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"
36 struct change_domains {
37 uint32_t invalidate_domains;
38 uint32_t flush_domains;
43 * Set the next domain for the specified object. This
44 * may not actually perform the necessary flushing/invaliding though,
45 * as that may want to be batched with other set_domain operations
47 * This is (we hope) the only really tricky part of gem. The goal
48 * is fairly simple -- track which caches hold bits of the object
49 * and make sure they remain coherent. A few concrete examples may
50 * help to explain how it works. For shorthand, we use the notation
51 * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
52 * a pair of read and write domain masks.
54 * Case 1: the batch buffer
60 * 5. Unmapped from GTT
63 * Let's take these a step at a time
66 * Pages allocated from the kernel may still have
67 * cache contents, so we set them to (CPU, CPU) always.
68 * 2. Written by CPU (using pwrite)
69 * The pwrite function calls set_domain (CPU, CPU) and
70 * this function does nothing (as nothing changes)
72 * This function asserts that the object is not
73 * currently in any GPU-based read or write domains
75 * i915_gem_execbuffer calls set_domain (COMMAND, 0).
76 * As write_domain is zero, this function adds in the
77 * current read domains (CPU+COMMAND, 0).
78 * flush_domains is set to CPU.
79 * invalidate_domains is set to COMMAND
80 * clflush is run to get data out of the CPU caches
81 * then i915_dev_set_domain calls i915_gem_flush to
82 * emit an MI_FLUSH and drm_agp_chipset_flush
83 * 5. Unmapped from GTT
84 * i915_gem_object_unbind calls set_domain (CPU, CPU)
85 * flush_domains and invalidate_domains end up both zero
86 * so no flushing/invalidating happens
90 * Case 2: The shared render buffer
94 * 3. Read/written by GPU
95 * 4. set_domain to (CPU,CPU)
96 * 5. Read/written by CPU
97 * 6. Read/written by GPU
100 * Same as last example, (CPU, CPU)
102 * Nothing changes (assertions find that it is not in the GPU)
103 * 3. Read/written by GPU
104 * execbuffer calls set_domain (RENDER, RENDER)
105 * flush_domains gets CPU
106 * invalidate_domains gets GPU
108 * MI_FLUSH and drm_agp_chipset_flush
109 * 4. set_domain (CPU, CPU)
110 * flush_domains gets GPU
111 * invalidate_domains gets CPU
112 * wait_rendering (obj) to make sure all drawing is complete.
113 * This will include an MI_FLUSH to get the data from GPU
115 * clflush (obj) to invalidate the CPU cache
116 * Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
117 * 5. Read/written by CPU
118 * cache lines are loaded and dirtied
119 * 6. Read written by GPU
120 * Same as last GPU access
122 * Case 3: The constant buffer
127 * 4. Updated (written) by CPU again
136 * flush_domains = CPU
137 * invalidate_domains = RENDER
140 * drm_agp_chipset_flush
141 * 4. Updated (written) by CPU again
143 * flush_domains = 0 (no previous write domain)
144 * invalidate_domains = 0 (no new read domains)
147 * flush_domains = CPU
148 * invalidate_domains = RENDER
151 * drm_agp_chipset_flush
154 i915_gem_object_set_to_gpu_domain(struct drm_i915_gem_object *obj,
155 struct intel_ring_buffer *ring,
156 struct change_domains *cd)
158 uint32_t invalidate_domains = 0, flush_domains = 0;
161 * If the object isn't moving to a new write domain,
162 * let the object stay in multiple read domains
164 if (obj->base.pending_write_domain == 0)
165 obj->base.pending_read_domains |= obj->base.read_domains;
168 * Flush the current write domain if
169 * the new read domains don't match. Invalidate
170 * any read domains which differ from the old
173 if (obj->base.write_domain &&
174 (((obj->base.write_domain != obj->base.pending_read_domains ||
175 obj->ring != ring)) ||
176 (obj->fenced_gpu_access && !obj->pending_fenced_gpu_access))) {
177 flush_domains |= obj->base.write_domain;
178 invalidate_domains |=
179 obj->base.pending_read_domains & ~obj->base.write_domain;
182 * Invalidate any read caches which may have
183 * stale data. That is, any new read domains.
185 invalidate_domains |= obj->base.pending_read_domains & ~obj->base.read_domains;
186 if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU)
187 i915_gem_clflush_object(obj);
189 /* blow away mappings if mapped through GTT */
190 if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_GTT)
191 i915_gem_release_mmap(obj);
193 /* The actual obj->write_domain will be updated with
194 * pending_write_domain after we emit the accumulated flush for all
195 * of our domain changes in execbuffers (which clears objects'
196 * write_domains). So if we have a current write domain that we
197 * aren't changing, set pending_write_domain to that.
199 if (flush_domains == 0 && obj->base.pending_write_domain == 0)
200 obj->base.pending_write_domain = obj->base.write_domain;
202 cd->invalidate_domains |= invalidate_domains;
203 cd->flush_domains |= flush_domains;
204 if (flush_domains & I915_GEM_GPU_DOMAINS)
205 cd->flush_rings |= obj->ring->id;
206 if (invalidate_domains & I915_GEM_GPU_DOMAINS)
207 cd->flush_rings |= ring->id;
212 struct hlist_head buckets[0];
215 static struct eb_objects *
218 struct eb_objects *eb;
219 int count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
222 eb = kzalloc(count*sizeof(struct hlist_head) +
223 sizeof(struct eb_objects),
233 eb_reset(struct eb_objects *eb)
235 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
239 eb_add_object(struct eb_objects *eb, struct drm_i915_gem_object *obj)
241 hlist_add_head(&obj->exec_node,
242 &eb->buckets[obj->exec_handle & eb->and]);
245 static struct drm_i915_gem_object *
246 eb_get_object(struct eb_objects *eb, unsigned long handle)
248 struct hlist_head *head;
249 struct hlist_node *node;
250 struct drm_i915_gem_object *obj;
252 head = &eb->buckets[handle & eb->and];
253 hlist_for_each(node, head) {
254 obj = hlist_entry(node, struct drm_i915_gem_object, exec_node);
255 if (obj->exec_handle == handle)
263 eb_destroy(struct eb_objects *eb)
269 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
270 struct eb_objects *eb,
271 struct drm_i915_gem_relocation_entry *reloc)
273 struct drm_device *dev = obj->base.dev;
274 struct drm_gem_object *target_obj;
275 uint32_t target_offset;
278 /* we've already hold a reference to all valid objects */
279 target_obj = &eb_get_object(eb, reloc->target_handle)->base;
280 if (unlikely(target_obj == NULL))
283 target_offset = to_intel_bo(target_obj)->gtt_offset;
285 /* The target buffer should have appeared before us in the
286 * exec_object list, so it should have a GTT space bound by now.
288 if (unlikely(target_offset == 0)) {
289 DRM_ERROR("No GTT space found for object %d\n",
290 reloc->target_handle);
294 /* Validate that the target is in a valid r/w GPU domain */
295 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
296 DRM_ERROR("reloc with multiple write domains: "
297 "obj %p target %d offset %d "
298 "read %08x write %08x",
299 obj, reloc->target_handle,
302 reloc->write_domain);
305 if (unlikely((reloc->write_domain | reloc->read_domains) & I915_GEM_DOMAIN_CPU)) {
306 DRM_ERROR("reloc with read/write CPU domains: "
307 "obj %p target %d offset %d "
308 "read %08x write %08x",
309 obj, reloc->target_handle,
312 reloc->write_domain);
315 if (unlikely(reloc->write_domain && target_obj->pending_write_domain &&
316 reloc->write_domain != target_obj->pending_write_domain)) {
317 DRM_ERROR("Write domain conflict: "
318 "obj %p target %d offset %d "
319 "new %08x old %08x\n",
320 obj, reloc->target_handle,
323 target_obj->pending_write_domain);
327 target_obj->pending_read_domains |= reloc->read_domains;
328 target_obj->pending_write_domain |= reloc->write_domain;
330 /* If the relocation already has the right value in it, no
331 * more work needs to be done.
333 if (target_offset == reloc->presumed_offset)
336 /* Check that the relocation address is valid... */
337 if (unlikely(reloc->offset > obj->base.size - 4)) {
338 DRM_ERROR("Relocation beyond object bounds: "
339 "obj %p target %d offset %d size %d.\n",
340 obj, reloc->target_handle,
342 (int) obj->base.size);
345 if (unlikely(reloc->offset & 3)) {
346 DRM_ERROR("Relocation not 4-byte aligned: "
347 "obj %p target %d offset %d.\n",
348 obj, reloc->target_handle,
349 (int) reloc->offset);
353 /* and points to somewhere within the target object. */
354 if (unlikely(reloc->delta >= target_obj->size)) {
355 DRM_ERROR("Relocation beyond target object bounds: "
356 "obj %p target %d delta %d size %d.\n",
357 obj, reloc->target_handle,
359 (int) target_obj->size);
363 reloc->delta += target_offset;
364 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) {
365 uint32_t page_offset = reloc->offset & ~PAGE_MASK;
368 vaddr = kmap_atomic(obj->pages[reloc->offset >> PAGE_SHIFT]);
369 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
370 kunmap_atomic(vaddr);
372 struct drm_i915_private *dev_priv = dev->dev_private;
373 uint32_t __iomem *reloc_entry;
374 void __iomem *reloc_page;
376 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
380 /* Map the page containing the relocation we're going to perform. */
381 reloc->offset += obj->gtt_offset;
382 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
383 reloc->offset & PAGE_MASK);
384 reloc_entry = (uint32_t __iomem *)
385 (reloc_page + (reloc->offset & ~PAGE_MASK));
386 iowrite32(reloc->delta, reloc_entry);
387 io_mapping_unmap_atomic(reloc_page);
390 /* and update the user's relocation entry */
391 reloc->presumed_offset = target_offset;
397 i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
398 struct eb_objects *eb)
400 struct drm_i915_gem_relocation_entry __user *user_relocs;
401 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
404 user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr;
405 for (i = 0; i < entry->relocation_count; i++) {
406 struct drm_i915_gem_relocation_entry reloc;
408 if (__copy_from_user_inatomic(&reloc,
413 ret = i915_gem_execbuffer_relocate_entry(obj, eb, &reloc);
417 if (__copy_to_user_inatomic(&user_relocs[i].presumed_offset,
418 &reloc.presumed_offset,
419 sizeof(reloc.presumed_offset)))
427 i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
428 struct eb_objects *eb,
429 struct drm_i915_gem_relocation_entry *relocs)
431 const struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
434 for (i = 0; i < entry->relocation_count; i++) {
435 ret = i915_gem_execbuffer_relocate_entry(obj, eb, &relocs[i]);
444 i915_gem_execbuffer_relocate(struct drm_device *dev,
445 struct eb_objects *eb,
446 struct list_head *objects)
448 struct drm_i915_gem_object *obj;
451 list_for_each_entry(obj, objects, exec_list) {
452 ret = i915_gem_execbuffer_relocate_object(obj, eb);
461 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
462 struct drm_file *file,
463 struct list_head *objects)
465 struct drm_i915_gem_object *obj;
467 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
468 struct list_head ordered_objects;
470 INIT_LIST_HEAD(&ordered_objects);
471 while (!list_empty(objects)) {
472 struct drm_i915_gem_exec_object2 *entry;
473 bool need_fence, need_mappable;
475 obj = list_first_entry(objects,
476 struct drm_i915_gem_object,
478 entry = obj->exec_entry;
481 has_fenced_gpu_access &&
482 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
483 obj->tiling_mode != I915_TILING_NONE;
485 entry->relocation_count ? true : need_fence;
488 list_move(&obj->exec_list, &ordered_objects);
490 list_move_tail(&obj->exec_list, &ordered_objects);
492 obj->base.pending_read_domains = 0;
493 obj->base.pending_write_domain = 0;
495 list_splice(&ordered_objects, objects);
497 /* Attempt to pin all of the buffers into the GTT.
498 * This is done in 3 phases:
500 * 1a. Unbind all objects that do not match the GTT constraints for
501 * the execbuffer (fenceable, mappable, alignment etc).
502 * 1b. Increment pin count for already bound objects.
503 * 2. Bind new objects.
504 * 3. Decrement pin count.
506 * This avoid unnecessary unbinding of later objects in order to makr
507 * room for the earlier objects *unless* we need to defragment.
513 /* Unbind any ill-fitting objects or pin. */
514 list_for_each_entry(obj, objects, exec_list) {
515 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
516 bool need_fence, need_mappable;
521 has_fenced_gpu_access &&
522 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
523 obj->tiling_mode != I915_TILING_NONE;
525 entry->relocation_count ? true : need_fence;
527 if ((entry->alignment && obj->gtt_offset & (entry->alignment - 1)) ||
528 (need_mappable && !obj->map_and_fenceable))
529 ret = i915_gem_object_unbind(obj);
531 ret = i915_gem_object_pin(obj,
540 /* Bind fresh objects */
541 list_for_each_entry(obj, objects, exec_list) {
542 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
546 has_fenced_gpu_access &&
547 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
548 obj->tiling_mode != I915_TILING_NONE;
550 if (!obj->gtt_space) {
552 entry->relocation_count ? true : need_fence;
554 ret = i915_gem_object_pin(obj,
561 if (has_fenced_gpu_access) {
563 ret = i915_gem_object_get_fence(obj, ring, 1);
566 } else if (entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
567 obj->tiling_mode == I915_TILING_NONE) {
569 ret = i915_gem_object_put_fence(obj);
573 obj->pending_fenced_gpu_access = need_fence;
576 entry->offset = obj->gtt_offset;
579 /* Decrement pin count for bound objects */
580 list_for_each_entry(obj, objects, exec_list) {
582 i915_gem_object_unpin(obj);
585 if (ret != -ENOSPC || retry > 1)
588 /* First attempt, just clear anything that is purgeable.
589 * Second attempt, clear the entire GTT.
591 ret = i915_gem_evict_everything(ring->dev, retry == 0);
599 obj = list_entry(obj->exec_list.prev,
600 struct drm_i915_gem_object,
602 while (objects != &obj->exec_list) {
604 i915_gem_object_unpin(obj);
606 obj = list_entry(obj->exec_list.prev,
607 struct drm_i915_gem_object,
615 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
616 struct drm_file *file,
617 struct intel_ring_buffer *ring,
618 struct list_head *objects,
619 struct eb_objects *eb,
620 struct drm_i915_gem_exec_object2 *exec,
623 struct drm_i915_gem_relocation_entry *reloc;
624 struct drm_i915_gem_object *obj;
628 /* We may process another execbuffer during the unlock... */
629 while (!list_empty(objects)) {
630 obj = list_first_entry(objects,
631 struct drm_i915_gem_object,
633 list_del_init(&obj->exec_list);
634 drm_gem_object_unreference(&obj->base);
637 mutex_unlock(&dev->struct_mutex);
640 for (i = 0; i < count; i++)
641 total += exec[i].relocation_count;
643 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
644 reloc = drm_malloc_ab(total, sizeof(*reloc));
645 if (reloc == NULL || reloc_offset == NULL) {
646 drm_free_large(reloc);
647 drm_free_large(reloc_offset);
648 mutex_lock(&dev->struct_mutex);
653 for (i = 0; i < count; i++) {
654 struct drm_i915_gem_relocation_entry __user *user_relocs;
656 user_relocs = (void __user *)(uintptr_t)exec[i].relocs_ptr;
658 if (copy_from_user(reloc+total, user_relocs,
659 exec[i].relocation_count * sizeof(*reloc))) {
661 mutex_lock(&dev->struct_mutex);
665 reloc_offset[i] = total;
666 total += exec[i].relocation_count;
669 ret = i915_mutex_lock_interruptible(dev);
671 mutex_lock(&dev->struct_mutex);
675 /* reacquire the objects */
677 for (i = 0; i < count; i++) {
678 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
680 if (&obj->base == NULL) {
681 DRM_ERROR("Invalid object handle %d at index %d\n",
687 list_add_tail(&obj->exec_list, objects);
688 obj->exec_handle = exec[i].handle;
689 obj->exec_entry = &exec[i];
690 eb_add_object(eb, obj);
693 ret = i915_gem_execbuffer_reserve(ring, file, objects);
697 list_for_each_entry(obj, objects, exec_list) {
698 int offset = obj->exec_entry - exec;
699 ret = i915_gem_execbuffer_relocate_object_slow(obj, eb,
700 reloc + reloc_offset[offset]);
705 /* Leave the user relocations as are, this is the painfully slow path,
706 * and we want to avoid the complication of dropping the lock whilst
707 * having buffers reserved in the aperture and so causing spurious
708 * ENOSPC for random operations.
712 drm_free_large(reloc);
713 drm_free_large(reloc_offset);
718 i915_gem_execbuffer_flush(struct drm_device *dev,
719 uint32_t invalidate_domains,
720 uint32_t flush_domains,
721 uint32_t flush_rings)
723 drm_i915_private_t *dev_priv = dev->dev_private;
726 if (flush_domains & I915_GEM_DOMAIN_CPU)
727 intel_gtt_chipset_flush();
729 if (flush_domains & I915_GEM_DOMAIN_GTT)
732 if ((flush_domains | invalidate_domains) & I915_GEM_GPU_DOMAINS) {
733 for (i = 0; i < I915_NUM_RINGS; i++)
734 if (flush_rings & (1 << i)) {
735 ret = i915_gem_flush_ring(&dev_priv->ring[i],
747 i915_gem_execbuffer_sync_rings(struct drm_i915_gem_object *obj,
748 struct intel_ring_buffer *to)
750 struct intel_ring_buffer *from = obj->ring;
754 if (from == NULL || to == from)
757 /* XXX gpu semaphores are currently causing hard hangs on SNB mobile */
758 if (INTEL_INFO(obj->base.dev)->gen < 6 || IS_MOBILE(obj->base.dev))
759 return i915_gem_object_wait_rendering(obj, true);
761 idx = intel_ring_sync_index(from, to);
763 seqno = obj->last_rendering_seqno;
764 if (seqno <= from->sync_seqno[idx])
767 if (seqno == from->outstanding_lazy_request) {
768 struct drm_i915_gem_request *request;
770 request = kzalloc(sizeof(*request), GFP_KERNEL);
774 ret = i915_add_request(from, NULL, request);
780 seqno = request->seqno;
783 from->sync_seqno[idx] = seqno;
784 return intel_ring_sync(to, from, seqno - 1);
788 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
789 struct list_head *objects)
791 struct drm_i915_gem_object *obj;
792 struct change_domains cd;
795 cd.invalidate_domains = 0;
796 cd.flush_domains = 0;
798 list_for_each_entry(obj, objects, exec_list)
799 i915_gem_object_set_to_gpu_domain(obj, ring, &cd);
801 if (cd.invalidate_domains | cd.flush_domains) {
802 ret = i915_gem_execbuffer_flush(ring->dev,
803 cd.invalidate_domains,
810 list_for_each_entry(obj, objects, exec_list) {
811 ret = i915_gem_execbuffer_sync_rings(obj, ring);
820 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
822 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
826 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
831 for (i = 0; i < count; i++) {
832 char __user *ptr = (char __user *)(uintptr_t)exec[i].relocs_ptr;
833 int length; /* limited by fault_in_pages_readable() */
835 /* First check for malicious input causing overflow */
836 if (exec[i].relocation_count >
837 INT_MAX / sizeof(struct drm_i915_gem_relocation_entry))
840 length = exec[i].relocation_count *
841 sizeof(struct drm_i915_gem_relocation_entry);
842 if (!access_ok(VERIFY_READ, ptr, length))
845 /* we may also need to update the presumed offsets */
846 if (!access_ok(VERIFY_WRITE, ptr, length))
849 if (fault_in_pages_readable(ptr, length))
857 i915_gem_execbuffer_wait_for_flips(struct intel_ring_buffer *ring,
858 struct list_head *objects)
860 struct drm_i915_gem_object *obj;
863 /* Check for any pending flips. As we only maintain a flip queue depth
864 * of 1, we can simply insert a WAIT for the next display flip prior
865 * to executing the batch and avoid stalling the CPU.
868 list_for_each_entry(obj, objects, exec_list) {
869 if (obj->base.write_domain)
870 flips |= atomic_read(&obj->pending_flip);
873 int plane, flip_mask, ret;
875 for (plane = 0; flips >> plane; plane++) {
876 if (((flips >> plane) & 1) == 0)
880 flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
882 flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
884 ret = intel_ring_begin(ring, 2);
888 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | flip_mask);
889 intel_ring_emit(ring, MI_NOOP);
890 intel_ring_advance(ring);
898 i915_gem_execbuffer_move_to_active(struct list_head *objects,
899 struct intel_ring_buffer *ring,
902 struct drm_i915_gem_object *obj;
904 list_for_each_entry(obj, objects, exec_list) {
905 u32 old_read = obj->base.read_domains;
906 u32 old_write = obj->base.write_domain;
909 obj->base.read_domains = obj->base.pending_read_domains;
910 obj->base.write_domain = obj->base.pending_write_domain;
911 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
913 i915_gem_object_move_to_active(obj, ring, seqno);
914 if (obj->base.write_domain) {
916 obj->pending_gpu_write = true;
917 list_move_tail(&obj->gpu_write_list,
918 &ring->gpu_write_list);
919 intel_mark_busy(ring->dev, obj);
922 trace_i915_gem_object_change_domain(obj, old_read, old_write);
927 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
928 struct drm_file *file,
929 struct intel_ring_buffer *ring)
931 struct drm_i915_gem_request *request;
935 * Ensure that the commands in the batch buffer are
936 * finished before the interrupt fires.
938 * The sampler always gets flushed on i965 (sigh).
940 invalidate = I915_GEM_DOMAIN_COMMAND;
941 if (INTEL_INFO(dev)->gen >= 4)
942 invalidate |= I915_GEM_DOMAIN_SAMPLER;
943 if (ring->flush(ring, invalidate, 0)) {
944 i915_gem_next_request_seqno(ring);
948 /* Add a breadcrumb for the completion of the batch buffer */
949 request = kzalloc(sizeof(*request), GFP_KERNEL);
950 if (request == NULL || i915_add_request(ring, file, request)) {
951 i915_gem_next_request_seqno(ring);
957 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
958 struct drm_file *file,
959 struct drm_i915_gem_execbuffer2 *args,
960 struct drm_i915_gem_exec_object2 *exec)
962 drm_i915_private_t *dev_priv = dev->dev_private;
963 struct list_head objects;
964 struct eb_objects *eb;
965 struct drm_i915_gem_object *batch_obj;
966 struct drm_clip_rect *cliprects = NULL;
967 struct intel_ring_buffer *ring;
968 u32 exec_start, exec_len;
972 if (!i915_gem_check_execbuffer(args)) {
973 DRM_ERROR("execbuf with invalid offset/length\n");
977 ret = validate_exec_list(exec, args->buffer_count);
981 switch (args->flags & I915_EXEC_RING_MASK) {
982 case I915_EXEC_DEFAULT:
983 case I915_EXEC_RENDER:
984 ring = &dev_priv->ring[RCS];
988 DRM_ERROR("execbuf with invalid ring (BSD)\n");
991 ring = &dev_priv->ring[VCS];
995 DRM_ERROR("execbuf with invalid ring (BLT)\n");
998 ring = &dev_priv->ring[BCS];
1001 DRM_ERROR("execbuf with unknown ring: %d\n",
1002 (int)(args->flags & I915_EXEC_RING_MASK));
1006 mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1008 case I915_EXEC_CONSTANTS_REL_GENERAL:
1009 case I915_EXEC_CONSTANTS_ABSOLUTE:
1010 case I915_EXEC_CONSTANTS_REL_SURFACE:
1011 if (ring == &dev_priv->ring[RCS] &&
1012 mode != dev_priv->relative_constants_mode) {
1013 if (INTEL_INFO(dev)->gen < 4)
1016 if (INTEL_INFO(dev)->gen > 5 &&
1017 mode == I915_EXEC_CONSTANTS_REL_SURFACE)
1020 ret = intel_ring_begin(ring, 4);
1024 intel_ring_emit(ring, MI_NOOP);
1025 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1026 intel_ring_emit(ring, INSTPM);
1027 intel_ring_emit(ring,
1028 I915_EXEC_CONSTANTS_MASK << 16 | mode);
1029 intel_ring_advance(ring);
1031 dev_priv->relative_constants_mode = mode;
1035 DRM_ERROR("execbuf with unknown constants: %d\n", mode);
1039 if (args->buffer_count < 1) {
1040 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
1044 if (args->num_cliprects != 0) {
1045 if (ring != &dev_priv->ring[RCS]) {
1046 DRM_ERROR("clip rectangles are only valid with the render ring\n");
1050 cliprects = kmalloc(args->num_cliprects * sizeof(*cliprects),
1052 if (cliprects == NULL) {
1057 if (copy_from_user(cliprects,
1058 (struct drm_clip_rect __user *)(uintptr_t)
1059 args->cliprects_ptr,
1060 sizeof(*cliprects)*args->num_cliprects)) {
1066 ret = i915_mutex_lock_interruptible(dev);
1070 if (dev_priv->mm.suspended) {
1071 mutex_unlock(&dev->struct_mutex);
1076 eb = eb_create(args->buffer_count);
1078 mutex_unlock(&dev->struct_mutex);
1083 /* Look up object handles */
1084 INIT_LIST_HEAD(&objects);
1085 for (i = 0; i < args->buffer_count; i++) {
1086 struct drm_i915_gem_object *obj;
1088 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
1090 if (&obj->base == NULL) {
1091 DRM_ERROR("Invalid object handle %d at index %d\n",
1093 /* prevent error path from reading uninitialized data */
1098 if (!list_empty(&obj->exec_list)) {
1099 DRM_ERROR("Object %p [handle %d, index %d] appears more than once in object list\n",
1100 obj, exec[i].handle, i);
1105 list_add_tail(&obj->exec_list, &objects);
1106 obj->exec_handle = exec[i].handle;
1107 obj->exec_entry = &exec[i];
1108 eb_add_object(eb, obj);
1111 /* take note of the batch buffer before we might reorder the lists */
1112 batch_obj = list_entry(objects.prev,
1113 struct drm_i915_gem_object,
1116 /* Move the objects en-masse into the GTT, evicting if necessary. */
1117 ret = i915_gem_execbuffer_reserve(ring, file, &objects);
1121 /* The objects are in their final locations, apply the relocations. */
1122 ret = i915_gem_execbuffer_relocate(dev, eb, &objects);
1124 if (ret == -EFAULT) {
1125 ret = i915_gem_execbuffer_relocate_slow(dev, file, ring,
1128 args->buffer_count);
1129 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1135 /* Set the pending read domains for the batch buffer to COMMAND */
1136 if (batch_obj->base.pending_write_domain) {
1137 DRM_ERROR("Attempting to use self-modifying batch buffer\n");
1141 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1143 ret = i915_gem_execbuffer_move_to_gpu(ring, &objects);
1147 ret = i915_gem_execbuffer_wait_for_flips(ring, &objects);
1151 seqno = i915_gem_next_request_seqno(ring);
1152 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++) {
1153 if (seqno < ring->sync_seqno[i]) {
1154 /* The GPU can not handle its semaphore value wrapping,
1155 * so every billion or so execbuffers, we need to stall
1156 * the GPU in order to reset the counters.
1158 ret = i915_gpu_idle(dev);
1162 BUG_ON(ring->sync_seqno[i]);
1166 trace_i915_gem_ring_dispatch(ring, seqno);
1168 exec_start = batch_obj->gtt_offset + args->batch_start_offset;
1169 exec_len = args->batch_len;
1171 for (i = 0; i < args->num_cliprects; i++) {
1172 ret = i915_emit_box(dev, &cliprects[i],
1173 args->DR1, args->DR4);
1177 ret = ring->dispatch_execbuffer(ring,
1178 exec_start, exec_len);
1183 ret = ring->dispatch_execbuffer(ring, exec_start, exec_len);
1188 i915_gem_execbuffer_move_to_active(&objects, ring, seqno);
1189 i915_gem_execbuffer_retire_commands(dev, file, ring);
1193 while (!list_empty(&objects)) {
1194 struct drm_i915_gem_object *obj;
1196 obj = list_first_entry(&objects,
1197 struct drm_i915_gem_object,
1199 list_del_init(&obj->exec_list);
1200 drm_gem_object_unreference(&obj->base);
1203 mutex_unlock(&dev->struct_mutex);
1211 * Legacy execbuffer just creates an exec2 list from the original exec object
1212 * list array and passes it to the real function.
1215 i915_gem_execbuffer(struct drm_device *dev, void *data,
1216 struct drm_file *file)
1218 struct drm_i915_gem_execbuffer *args = data;
1219 struct drm_i915_gem_execbuffer2 exec2;
1220 struct drm_i915_gem_exec_object *exec_list = NULL;
1221 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1224 if (args->buffer_count < 1) {
1225 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
1229 /* Copy in the exec list from userland */
1230 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1231 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1232 if (exec_list == NULL || exec2_list == NULL) {
1233 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
1234 args->buffer_count);
1235 drm_free_large(exec_list);
1236 drm_free_large(exec2_list);
1239 ret = copy_from_user(exec_list,
1240 (struct drm_i915_relocation_entry __user *)
1241 (uintptr_t) args->buffers_ptr,
1242 sizeof(*exec_list) * args->buffer_count);
1244 DRM_ERROR("copy %d exec entries failed %d\n",
1245 args->buffer_count, ret);
1246 drm_free_large(exec_list);
1247 drm_free_large(exec2_list);
1251 for (i = 0; i < args->buffer_count; i++) {
1252 exec2_list[i].handle = exec_list[i].handle;
1253 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1254 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1255 exec2_list[i].alignment = exec_list[i].alignment;
1256 exec2_list[i].offset = exec_list[i].offset;
1257 if (INTEL_INFO(dev)->gen < 4)
1258 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1260 exec2_list[i].flags = 0;
1263 exec2.buffers_ptr = args->buffers_ptr;
1264 exec2.buffer_count = args->buffer_count;
1265 exec2.batch_start_offset = args->batch_start_offset;
1266 exec2.batch_len = args->batch_len;
1267 exec2.DR1 = args->DR1;
1268 exec2.DR4 = args->DR4;
1269 exec2.num_cliprects = args->num_cliprects;
1270 exec2.cliprects_ptr = args->cliprects_ptr;
1271 exec2.flags = I915_EXEC_RENDER;
1273 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1275 /* Copy the new buffer offsets back to the user's exec list. */
1276 for (i = 0; i < args->buffer_count; i++)
1277 exec_list[i].offset = exec2_list[i].offset;
1278 /* ... and back out to userspace */
1279 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1280 (uintptr_t) args->buffers_ptr,
1282 sizeof(*exec_list) * args->buffer_count);
1285 DRM_ERROR("failed to copy %d exec entries "
1286 "back to user (%d)\n",
1287 args->buffer_count, ret);
1291 drm_free_large(exec_list);
1292 drm_free_large(exec2_list);
1297 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1298 struct drm_file *file)
1300 struct drm_i915_gem_execbuffer2 *args = data;
1301 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1304 if (args->buffer_count < 1) {
1305 DRM_ERROR("execbuf2 with %d buffers\n", args->buffer_count);
1309 exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1310 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
1311 if (exec2_list == NULL)
1312 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1313 args->buffer_count);
1314 if (exec2_list == NULL) {
1315 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
1316 args->buffer_count);
1319 ret = copy_from_user(exec2_list,
1320 (struct drm_i915_relocation_entry __user *)
1321 (uintptr_t) args->buffers_ptr,
1322 sizeof(*exec2_list) * args->buffer_count);
1324 DRM_ERROR("copy %d exec entries failed %d\n",
1325 args->buffer_count, ret);
1326 drm_free_large(exec2_list);
1330 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1332 /* Copy the new buffer offsets back to the user's exec list. */
1333 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1334 (uintptr_t) args->buffers_ptr,
1336 sizeof(*exec2_list) * args->buffer_count);
1339 DRM_ERROR("failed to copy %d exec entries "
1340 "back to user (%d)\n",
1341 args->buffer_count, ret);
1345 drm_free_large(exec2_list);