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>
36 struct change_domains {
37 uint32_t invalidate_domains;
38 uint32_t flush_domains;
44 * Set the next domain for the specified object. This
45 * may not actually perform the necessary flushing/invaliding though,
46 * as that may want to be batched with other set_domain operations
48 * This is (we hope) the only really tricky part of gem. The goal
49 * is fairly simple -- track which caches hold bits of the object
50 * and make sure they remain coherent. A few concrete examples may
51 * help to explain how it works. For shorthand, we use the notation
52 * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
53 * a pair of read and write domain masks.
55 * Case 1: the batch buffer
61 * 5. Unmapped from GTT
64 * Let's take these a step at a time
67 * Pages allocated from the kernel may still have
68 * cache contents, so we set them to (CPU, CPU) always.
69 * 2. Written by CPU (using pwrite)
70 * The pwrite function calls set_domain (CPU, CPU) and
71 * this function does nothing (as nothing changes)
73 * This function asserts that the object is not
74 * currently in any GPU-based read or write domains
76 * i915_gem_execbuffer calls set_domain (COMMAND, 0).
77 * As write_domain is zero, this function adds in the
78 * current read domains (CPU+COMMAND, 0).
79 * flush_domains is set to CPU.
80 * invalidate_domains is set to COMMAND
81 * clflush is run to get data out of the CPU caches
82 * then i915_dev_set_domain calls i915_gem_flush to
83 * emit an MI_FLUSH and drm_agp_chipset_flush
84 * 5. Unmapped from GTT
85 * i915_gem_object_unbind calls set_domain (CPU, CPU)
86 * flush_domains and invalidate_domains end up both zero
87 * so no flushing/invalidating happens
91 * Case 2: The shared render buffer
95 * 3. Read/written by GPU
96 * 4. set_domain to (CPU,CPU)
97 * 5. Read/written by CPU
98 * 6. Read/written by GPU
101 * Same as last example, (CPU, CPU)
103 * Nothing changes (assertions find that it is not in the GPU)
104 * 3. Read/written by GPU
105 * execbuffer calls set_domain (RENDER, RENDER)
106 * flush_domains gets CPU
107 * invalidate_domains gets GPU
109 * MI_FLUSH and drm_agp_chipset_flush
110 * 4. set_domain (CPU, CPU)
111 * flush_domains gets GPU
112 * invalidate_domains gets CPU
113 * wait_rendering (obj) to make sure all drawing is complete.
114 * This will include an MI_FLUSH to get the data from GPU
116 * clflush (obj) to invalidate the CPU cache
117 * Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
118 * 5. Read/written by CPU
119 * cache lines are loaded and dirtied
120 * 6. Read written by GPU
121 * Same as last GPU access
123 * Case 3: The constant buffer
128 * 4. Updated (written) by CPU again
137 * flush_domains = CPU
138 * invalidate_domains = RENDER
141 * drm_agp_chipset_flush
142 * 4. Updated (written) by CPU again
144 * flush_domains = 0 (no previous write domain)
145 * invalidate_domains = 0 (no new read domains)
148 * flush_domains = CPU
149 * invalidate_domains = RENDER
152 * drm_agp_chipset_flush
155 i915_gem_object_set_to_gpu_domain(struct drm_i915_gem_object *obj,
156 struct intel_ring_buffer *ring,
157 struct change_domains *cd)
159 uint32_t invalidate_domains = 0, flush_domains = 0;
162 * If the object isn't moving to a new write domain,
163 * let the object stay in multiple read domains
165 if (obj->base.pending_write_domain == 0)
166 obj->base.pending_read_domains |= obj->base.read_domains;
169 * Flush the current write domain if
170 * the new read domains don't match. Invalidate
171 * any read domains which differ from the old
174 if (obj->base.write_domain &&
175 (((obj->base.write_domain != obj->base.pending_read_domains ||
176 obj->ring != ring)) ||
177 (obj->fenced_gpu_access && !obj->pending_fenced_gpu_access))) {
178 flush_domains |= obj->base.write_domain;
179 invalidate_domains |=
180 obj->base.pending_read_domains & ~obj->base.write_domain;
183 * Invalidate any read caches which may have
184 * stale data. That is, any new read domains.
186 invalidate_domains |= obj->base.pending_read_domains & ~obj->base.read_domains;
187 if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU)
188 i915_gem_clflush_object(obj);
190 if (obj->base.pending_write_domain)
191 cd->flips |= atomic_read(&obj->pending_flip);
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 |= intel_ring_flag(obj->ring);
206 if (invalidate_domains & I915_GEM_GPU_DOMAINS)
207 cd->flush_rings |= intel_ring_flag(ring);
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)
268 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
270 return (obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
271 obj->cache_level != I915_CACHE_NONE);
275 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
276 struct eb_objects *eb,
277 struct drm_i915_gem_relocation_entry *reloc)
279 struct drm_device *dev = obj->base.dev;
280 struct drm_gem_object *target_obj;
281 struct drm_i915_gem_object *target_i915_obj;
282 uint32_t target_offset;
285 /* we've already hold a reference to all valid objects */
286 target_obj = &eb_get_object(eb, reloc->target_handle)->base;
287 if (unlikely(target_obj == NULL))
290 target_i915_obj = to_intel_bo(target_obj);
291 target_offset = target_i915_obj->gtt_offset;
293 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
294 * pipe_control writes because the gpu doesn't properly redirect them
295 * through the ppgtt for non_secure batchbuffers. */
296 if (unlikely(IS_GEN6(dev) &&
297 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
298 !target_i915_obj->has_global_gtt_mapping)) {
299 i915_gem_gtt_bind_object(target_i915_obj,
300 target_i915_obj->cache_level);
303 /* The target buffer should have appeared before us in the
304 * exec_object list, so it should have a GTT space bound by now.
306 if (unlikely(target_offset == 0)) {
307 DRM_DEBUG("No GTT space found for object %d\n",
308 reloc->target_handle);
312 /* Validate that the target is in a valid r/w GPU domain */
313 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
314 DRM_DEBUG("reloc with multiple write domains: "
315 "obj %p target %d offset %d "
316 "read %08x write %08x",
317 obj, reloc->target_handle,
320 reloc->write_domain);
323 if (unlikely((reloc->write_domain | reloc->read_domains)
324 & ~I915_GEM_GPU_DOMAINS)) {
325 DRM_DEBUG("reloc with read/write non-GPU domains: "
326 "obj %p target %d offset %d "
327 "read %08x write %08x",
328 obj, reloc->target_handle,
331 reloc->write_domain);
334 if (unlikely(reloc->write_domain && target_obj->pending_write_domain &&
335 reloc->write_domain != target_obj->pending_write_domain)) {
336 DRM_DEBUG("Write domain conflict: "
337 "obj %p target %d offset %d "
338 "new %08x old %08x\n",
339 obj, reloc->target_handle,
342 target_obj->pending_write_domain);
346 target_obj->pending_read_domains |= reloc->read_domains;
347 target_obj->pending_write_domain |= reloc->write_domain;
349 /* If the relocation already has the right value in it, no
350 * more work needs to be done.
352 if (target_offset == reloc->presumed_offset)
355 /* Check that the relocation address is valid... */
356 if (unlikely(reloc->offset > obj->base.size - 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 uint32_t page_offset = reloc->offset & ~PAGE_MASK;
381 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
385 vaddr = kmap_atomic(obj->pages[reloc->offset >> PAGE_SHIFT]);
386 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
387 kunmap_atomic(vaddr);
389 struct drm_i915_private *dev_priv = dev->dev_private;
390 uint32_t __iomem *reloc_entry;
391 void __iomem *reloc_page;
393 ret = i915_gem_object_set_to_gtt_domain(obj, true);
397 ret = i915_gem_object_put_fence(obj);
401 /* Map the page containing the relocation we're going to perform. */
402 reloc->offset += obj->gtt_offset;
403 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
404 reloc->offset & PAGE_MASK);
405 reloc_entry = (uint32_t __iomem *)
406 (reloc_page + (reloc->offset & ~PAGE_MASK));
407 iowrite32(reloc->delta, reloc_entry);
408 io_mapping_unmap_atomic(reloc_page);
411 /* and update the user's relocation entry */
412 reloc->presumed_offset = target_offset;
418 i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
419 struct eb_objects *eb)
421 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
422 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
423 struct drm_i915_gem_relocation_entry __user *user_relocs;
424 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
427 user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr;
429 remain = entry->relocation_count;
431 struct drm_i915_gem_relocation_entry *r = stack_reloc;
433 if (count > ARRAY_SIZE(stack_reloc))
434 count = ARRAY_SIZE(stack_reloc);
437 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
441 u64 offset = r->presumed_offset;
443 ret = i915_gem_execbuffer_relocate_entry(obj, eb, r);
447 if (r->presumed_offset != offset &&
448 __copy_to_user_inatomic(&user_relocs->presumed_offset,
450 sizeof(r->presumed_offset))) {
464 i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
465 struct eb_objects *eb,
466 struct drm_i915_gem_relocation_entry *relocs)
468 const struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
471 for (i = 0; i < entry->relocation_count; i++) {
472 ret = i915_gem_execbuffer_relocate_entry(obj, eb, &relocs[i]);
481 i915_gem_execbuffer_relocate(struct drm_device *dev,
482 struct eb_objects *eb,
483 struct list_head *objects)
485 struct drm_i915_gem_object *obj;
488 /* This is the fast path and we cannot handle a pagefault whilst
489 * holding the struct mutex lest the user pass in the relocations
490 * contained within a mmaped bo. For in such a case we, the page
491 * fault handler would call i915_gem_fault() and we would try to
492 * acquire the struct mutex again. Obviously this is bad and so
493 * lockdep complains vehemently.
496 list_for_each_entry(obj, objects, exec_list) {
497 ret = i915_gem_execbuffer_relocate_object(obj, eb);
506 #define __EXEC_OBJECT_HAS_FENCE (1<<31)
509 need_reloc_mappable(struct drm_i915_gem_object *obj)
511 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
512 return entry->relocation_count && !use_cpu_reloc(obj);
516 pin_and_fence_object(struct drm_i915_gem_object *obj,
517 struct intel_ring_buffer *ring)
519 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
520 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
521 bool need_fence, need_mappable;
525 has_fenced_gpu_access &&
526 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
527 obj->tiling_mode != I915_TILING_NONE;
528 need_mappable = need_fence || need_reloc_mappable(obj);
530 ret = i915_gem_object_pin(obj, entry->alignment, need_mappable);
534 if (has_fenced_gpu_access) {
535 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
536 ret = i915_gem_object_get_fence(obj);
540 if (i915_gem_object_pin_fence(obj))
541 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
543 obj->pending_fenced_gpu_access = true;
547 entry->offset = obj->gtt_offset;
551 i915_gem_object_unpin(obj);
556 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
557 struct drm_file *file,
558 struct list_head *objects)
560 drm_i915_private_t *dev_priv = ring->dev->dev_private;
561 struct drm_i915_gem_object *obj;
563 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
564 struct list_head ordered_objects;
566 INIT_LIST_HEAD(&ordered_objects);
567 while (!list_empty(objects)) {
568 struct drm_i915_gem_exec_object2 *entry;
569 bool need_fence, need_mappable;
571 obj = list_first_entry(objects,
572 struct drm_i915_gem_object,
574 entry = obj->exec_entry;
577 has_fenced_gpu_access &&
578 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
579 obj->tiling_mode != I915_TILING_NONE;
580 need_mappable = need_fence || need_reloc_mappable(obj);
583 list_move(&obj->exec_list, &ordered_objects);
585 list_move_tail(&obj->exec_list, &ordered_objects);
587 obj->base.pending_read_domains = 0;
588 obj->base.pending_write_domain = 0;
590 list_splice(&ordered_objects, objects);
592 /* Attempt to pin all of the buffers into the GTT.
593 * This is done in 3 phases:
595 * 1a. Unbind all objects that do not match the GTT constraints for
596 * the execbuffer (fenceable, mappable, alignment etc).
597 * 1b. Increment pin count for already bound objects.
598 * 2. Bind new objects.
599 * 3. Decrement pin count.
601 * This avoid unnecessary unbinding of later objects in order to makr
602 * room for the earlier objects *unless* we need to defragment.
608 /* Unbind any ill-fitting objects or pin. */
609 list_for_each_entry(obj, objects, exec_list) {
610 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
611 bool need_fence, need_mappable;
617 has_fenced_gpu_access &&
618 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
619 obj->tiling_mode != I915_TILING_NONE;
620 need_mappable = need_fence || need_reloc_mappable(obj);
622 if ((entry->alignment && obj->gtt_offset & (entry->alignment - 1)) ||
623 (need_mappable && !obj->map_and_fenceable))
624 ret = i915_gem_object_unbind(obj);
626 ret = pin_and_fence_object(obj, ring);
631 /* Bind fresh objects */
632 list_for_each_entry(obj, objects, exec_list) {
636 ret = pin_and_fence_object(obj, ring);
640 /* This can potentially raise a harmless
641 * -EINVAL if we failed to bind in the above
642 * call. It cannot raise -EINTR since we know
643 * that the bo is freshly bound and so will
644 * not need to be flushed or waited upon.
646 ret_ignore = i915_gem_object_unbind(obj);
648 WARN_ON(obj->gtt_space);
653 /* Decrement pin count for bound objects */
654 list_for_each_entry(obj, objects, exec_list) {
655 struct drm_i915_gem_exec_object2 *entry;
660 entry = obj->exec_entry;
661 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
662 i915_gem_object_unpin_fence(obj);
663 entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
666 i915_gem_object_unpin(obj);
668 /* ... and ensure ppgtt mapping exist if needed. */
669 if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) {
670 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
671 obj, obj->cache_level);
673 obj->has_aliasing_ppgtt_mapping = 1;
677 if (ret != -ENOSPC || retry > 1)
680 /* First attempt, just clear anything that is purgeable.
681 * Second attempt, clear the entire GTT.
683 ret = i915_gem_evict_everything(ring->dev, retry == 0);
691 list_for_each_entry_continue_reverse(obj, objects, exec_list) {
692 struct drm_i915_gem_exec_object2 *entry;
697 entry = obj->exec_entry;
698 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
699 i915_gem_object_unpin_fence(obj);
700 entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
703 i915_gem_object_unpin(obj);
710 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
711 struct drm_file *file,
712 struct intel_ring_buffer *ring,
713 struct list_head *objects,
714 struct eb_objects *eb,
715 struct drm_i915_gem_exec_object2 *exec,
718 struct drm_i915_gem_relocation_entry *reloc;
719 struct drm_i915_gem_object *obj;
723 /* We may process another execbuffer during the unlock... */
724 while (!list_empty(objects)) {
725 obj = list_first_entry(objects,
726 struct drm_i915_gem_object,
728 list_del_init(&obj->exec_list);
729 drm_gem_object_unreference(&obj->base);
732 mutex_unlock(&dev->struct_mutex);
735 for (i = 0; i < count; i++)
736 total += exec[i].relocation_count;
738 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
739 reloc = drm_malloc_ab(total, sizeof(*reloc));
740 if (reloc == NULL || reloc_offset == NULL) {
741 drm_free_large(reloc);
742 drm_free_large(reloc_offset);
743 mutex_lock(&dev->struct_mutex);
748 for (i = 0; i < count; i++) {
749 struct drm_i915_gem_relocation_entry __user *user_relocs;
751 user_relocs = (void __user *)(uintptr_t)exec[i].relocs_ptr;
753 if (copy_from_user(reloc+total, user_relocs,
754 exec[i].relocation_count * sizeof(*reloc))) {
756 mutex_lock(&dev->struct_mutex);
760 reloc_offset[i] = total;
761 total += exec[i].relocation_count;
764 ret = i915_mutex_lock_interruptible(dev);
766 mutex_lock(&dev->struct_mutex);
770 /* reacquire the objects */
772 for (i = 0; i < count; i++) {
773 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
775 if (&obj->base == NULL) {
776 DRM_DEBUG("Invalid object handle %d at index %d\n",
782 list_add_tail(&obj->exec_list, objects);
783 obj->exec_handle = exec[i].handle;
784 obj->exec_entry = &exec[i];
785 eb_add_object(eb, obj);
788 ret = i915_gem_execbuffer_reserve(ring, file, objects);
792 list_for_each_entry(obj, objects, exec_list) {
793 int offset = obj->exec_entry - exec;
794 ret = i915_gem_execbuffer_relocate_object_slow(obj, eb,
795 reloc + reloc_offset[offset]);
800 /* Leave the user relocations as are, this is the painfully slow path,
801 * and we want to avoid the complication of dropping the lock whilst
802 * having buffers reserved in the aperture and so causing spurious
803 * ENOSPC for random operations.
807 drm_free_large(reloc);
808 drm_free_large(reloc_offset);
813 i915_gem_execbuffer_flush(struct drm_device *dev,
814 uint32_t invalidate_domains,
815 uint32_t flush_domains)
817 if (flush_domains & I915_GEM_DOMAIN_CPU)
818 intel_gtt_chipset_flush();
820 if (flush_domains & I915_GEM_DOMAIN_GTT)
825 i915_gem_execbuffer_wait_for_flips(struct intel_ring_buffer *ring, u32 flips)
827 u32 plane, flip_mask;
830 /* Check for any pending flips. As we only maintain a flip queue depth
831 * of 1, we can simply insert a WAIT for the next display flip prior
832 * to executing the batch and avoid stalling the CPU.
835 for (plane = 0; flips >> plane; plane++) {
836 if (((flips >> plane) & 1) == 0)
840 flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
842 flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
844 ret = intel_ring_begin(ring, 2);
848 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | flip_mask);
849 intel_ring_emit(ring, MI_NOOP);
850 intel_ring_advance(ring);
858 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
859 struct list_head *objects)
861 struct drm_i915_gem_object *obj;
862 struct change_domains cd;
865 memset(&cd, 0, sizeof(cd));
866 list_for_each_entry(obj, objects, exec_list)
867 i915_gem_object_set_to_gpu_domain(obj, ring, &cd);
869 if (cd.invalidate_domains | cd.flush_domains) {
870 i915_gem_execbuffer_flush(ring->dev,
871 cd.invalidate_domains,
876 ret = i915_gem_execbuffer_wait_for_flips(ring, cd.flips);
881 list_for_each_entry(obj, objects, exec_list) {
882 ret = i915_gem_object_sync(obj, ring);
887 /* Unconditionally invalidate gpu caches and ensure that we do flush
888 * any residual writes from the previous batch.
890 ret = i915_gem_flush_ring(ring,
891 I915_GEM_GPU_DOMAINS,
892 ring->gpu_caches_dirty ? I915_GEM_GPU_DOMAINS : 0);
896 ring->gpu_caches_dirty = false;
901 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
903 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
907 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
912 for (i = 0; i < count; i++) {
913 char __user *ptr = (char __user *)(uintptr_t)exec[i].relocs_ptr;
914 int length; /* limited by fault_in_pages_readable() */
916 /* First check for malicious input causing overflow */
917 if (exec[i].relocation_count >
918 INT_MAX / sizeof(struct drm_i915_gem_relocation_entry))
921 length = exec[i].relocation_count *
922 sizeof(struct drm_i915_gem_relocation_entry);
923 if (!access_ok(VERIFY_READ, ptr, length))
926 /* we may also need to update the presumed offsets */
927 if (!access_ok(VERIFY_WRITE, ptr, length))
930 if (fault_in_multipages_readable(ptr, length))
938 i915_gem_execbuffer_move_to_active(struct list_head *objects,
939 struct intel_ring_buffer *ring,
942 struct drm_i915_gem_object *obj;
944 list_for_each_entry(obj, objects, exec_list) {
945 u32 old_read = obj->base.read_domains;
946 u32 old_write = obj->base.write_domain;
949 obj->base.read_domains = obj->base.pending_read_domains;
950 obj->base.write_domain = obj->base.pending_write_domain;
951 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
953 i915_gem_object_move_to_active(obj, ring, seqno);
954 if (obj->base.write_domain) {
956 obj->pending_gpu_write = true;
957 list_move_tail(&obj->gpu_write_list,
958 &ring->gpu_write_list);
959 if (obj->pin_count) /* check for potential scanout */
960 intel_mark_busy(ring->dev, obj);
963 trace_i915_gem_object_change_domain(obj, old_read, old_write);
966 intel_mark_busy(ring->dev, NULL);
970 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
971 struct drm_file *file,
972 struct intel_ring_buffer *ring)
974 struct drm_i915_gem_request *request;
976 /* Unconditionally force add_request to emit a full flush. */
977 ring->gpu_caches_dirty = true;
979 /* Add a breadcrumb for the completion of the batch buffer */
980 request = kzalloc(sizeof(*request), GFP_KERNEL);
981 if (request == NULL || i915_add_request(ring, file, request)) {
987 i915_reset_gen7_sol_offsets(struct drm_device *dev,
988 struct intel_ring_buffer *ring)
990 drm_i915_private_t *dev_priv = dev->dev_private;
993 if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS])
996 ret = intel_ring_begin(ring, 4 * 3);
1000 for (i = 0; i < 4; i++) {
1001 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1002 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1003 intel_ring_emit(ring, 0);
1006 intel_ring_advance(ring);
1012 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1013 struct drm_file *file,
1014 struct drm_i915_gem_execbuffer2 *args,
1015 struct drm_i915_gem_exec_object2 *exec)
1017 drm_i915_private_t *dev_priv = dev->dev_private;
1018 struct list_head objects;
1019 struct eb_objects *eb;
1020 struct drm_i915_gem_object *batch_obj;
1021 struct drm_clip_rect *cliprects = NULL;
1022 struct intel_ring_buffer *ring;
1023 u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1024 u32 exec_start, exec_len;
1029 if (!i915_gem_check_execbuffer(args)) {
1030 DRM_DEBUG("execbuf with invalid offset/length\n");
1034 ret = validate_exec_list(exec, args->buffer_count);
1038 switch (args->flags & I915_EXEC_RING_MASK) {
1039 case I915_EXEC_DEFAULT:
1040 case I915_EXEC_RENDER:
1041 ring = &dev_priv->ring[RCS];
1044 ring = &dev_priv->ring[VCS];
1046 DRM_DEBUG("Ring %s doesn't support contexts\n",
1052 ring = &dev_priv->ring[BCS];
1054 DRM_DEBUG("Ring %s doesn't support contexts\n",
1060 DRM_DEBUG("execbuf with unknown ring: %d\n",
1061 (int)(args->flags & I915_EXEC_RING_MASK));
1064 if (!intel_ring_initialized(ring)) {
1065 DRM_DEBUG("execbuf with invalid ring: %d\n",
1066 (int)(args->flags & I915_EXEC_RING_MASK));
1070 mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1071 mask = I915_EXEC_CONSTANTS_MASK;
1073 case I915_EXEC_CONSTANTS_REL_GENERAL:
1074 case I915_EXEC_CONSTANTS_ABSOLUTE:
1075 case I915_EXEC_CONSTANTS_REL_SURFACE:
1076 if (ring == &dev_priv->ring[RCS] &&
1077 mode != dev_priv->relative_constants_mode) {
1078 if (INTEL_INFO(dev)->gen < 4)
1081 if (INTEL_INFO(dev)->gen > 5 &&
1082 mode == I915_EXEC_CONSTANTS_REL_SURFACE)
1085 /* The HW changed the meaning on this bit on gen6 */
1086 if (INTEL_INFO(dev)->gen >= 6)
1087 mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1091 DRM_DEBUG("execbuf with unknown constants: %d\n", mode);
1095 if (args->buffer_count < 1) {
1096 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1100 if (args->num_cliprects != 0) {
1101 if (ring != &dev_priv->ring[RCS]) {
1102 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1106 if (INTEL_INFO(dev)->gen >= 5) {
1107 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1111 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1112 DRM_DEBUG("execbuf with %u cliprects\n",
1113 args->num_cliprects);
1117 cliprects = kmalloc(args->num_cliprects * sizeof(*cliprects),
1119 if (cliprects == NULL) {
1124 if (copy_from_user(cliprects,
1125 (struct drm_clip_rect __user *)(uintptr_t)
1126 args->cliprects_ptr,
1127 sizeof(*cliprects)*args->num_cliprects)) {
1133 ret = i915_mutex_lock_interruptible(dev);
1137 if (dev_priv->mm.suspended) {
1138 mutex_unlock(&dev->struct_mutex);
1143 eb = eb_create(args->buffer_count);
1145 mutex_unlock(&dev->struct_mutex);
1150 /* Look up object handles */
1151 INIT_LIST_HEAD(&objects);
1152 for (i = 0; i < args->buffer_count; i++) {
1153 struct drm_i915_gem_object *obj;
1155 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
1157 if (&obj->base == NULL) {
1158 DRM_DEBUG("Invalid object handle %d at index %d\n",
1160 /* prevent error path from reading uninitialized data */
1165 if (!list_empty(&obj->exec_list)) {
1166 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
1167 obj, exec[i].handle, i);
1172 list_add_tail(&obj->exec_list, &objects);
1173 obj->exec_handle = exec[i].handle;
1174 obj->exec_entry = &exec[i];
1175 eb_add_object(eb, obj);
1178 /* take note of the batch buffer before we might reorder the lists */
1179 batch_obj = list_entry(objects.prev,
1180 struct drm_i915_gem_object,
1183 /* Move the objects en-masse into the GTT, evicting if necessary. */
1184 ret = i915_gem_execbuffer_reserve(ring, file, &objects);
1188 /* The objects are in their final locations, apply the relocations. */
1189 ret = i915_gem_execbuffer_relocate(dev, eb, &objects);
1191 if (ret == -EFAULT) {
1192 ret = i915_gem_execbuffer_relocate_slow(dev, file, ring,
1195 args->buffer_count);
1196 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1202 /* Set the pending read domains for the batch buffer to COMMAND */
1203 if (batch_obj->base.pending_write_domain) {
1204 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1208 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1210 ret = i915_gem_execbuffer_move_to_gpu(ring, &objects);
1214 seqno = i915_gem_next_request_seqno(ring);
1215 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++) {
1216 if (seqno < ring->sync_seqno[i]) {
1217 /* The GPU can not handle its semaphore value wrapping,
1218 * so every billion or so execbuffers, we need to stall
1219 * the GPU in order to reset the counters.
1221 ret = i915_gpu_idle(dev);
1224 i915_gem_retire_requests(dev);
1226 BUG_ON(ring->sync_seqno[i]);
1230 ret = i915_switch_context(ring, file, ctx_id);
1234 if (ring == &dev_priv->ring[RCS] &&
1235 mode != dev_priv->relative_constants_mode) {
1236 ret = intel_ring_begin(ring, 4);
1240 intel_ring_emit(ring, MI_NOOP);
1241 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1242 intel_ring_emit(ring, INSTPM);
1243 intel_ring_emit(ring, mask << 16 | mode);
1244 intel_ring_advance(ring);
1246 dev_priv->relative_constants_mode = mode;
1249 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1250 ret = i915_reset_gen7_sol_offsets(dev, ring);
1255 trace_i915_gem_ring_dispatch(ring, seqno);
1257 exec_start = batch_obj->gtt_offset + args->batch_start_offset;
1258 exec_len = args->batch_len;
1260 for (i = 0; i < args->num_cliprects; i++) {
1261 ret = i915_emit_box(dev, &cliprects[i],
1262 args->DR1, args->DR4);
1266 ret = ring->dispatch_execbuffer(ring,
1267 exec_start, exec_len);
1272 ret = ring->dispatch_execbuffer(ring, exec_start, exec_len);
1277 i915_gem_execbuffer_move_to_active(&objects, ring, seqno);
1278 i915_gem_execbuffer_retire_commands(dev, file, ring);
1282 while (!list_empty(&objects)) {
1283 struct drm_i915_gem_object *obj;
1285 obj = list_first_entry(&objects,
1286 struct drm_i915_gem_object,
1288 list_del_init(&obj->exec_list);
1289 drm_gem_object_unreference(&obj->base);
1292 mutex_unlock(&dev->struct_mutex);
1300 * Legacy execbuffer just creates an exec2 list from the original exec object
1301 * list array and passes it to the real function.
1304 i915_gem_execbuffer(struct drm_device *dev, void *data,
1305 struct drm_file *file)
1307 struct drm_i915_gem_execbuffer *args = data;
1308 struct drm_i915_gem_execbuffer2 exec2;
1309 struct drm_i915_gem_exec_object *exec_list = NULL;
1310 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1313 if (args->buffer_count < 1) {
1314 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1318 /* Copy in the exec list from userland */
1319 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1320 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1321 if (exec_list == NULL || exec2_list == NULL) {
1322 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1323 args->buffer_count);
1324 drm_free_large(exec_list);
1325 drm_free_large(exec2_list);
1328 ret = copy_from_user(exec_list,
1329 (struct drm_i915_relocation_entry __user *)
1330 (uintptr_t) args->buffers_ptr,
1331 sizeof(*exec_list) * args->buffer_count);
1333 DRM_DEBUG("copy %d exec entries failed %d\n",
1334 args->buffer_count, ret);
1335 drm_free_large(exec_list);
1336 drm_free_large(exec2_list);
1340 for (i = 0; i < args->buffer_count; i++) {
1341 exec2_list[i].handle = exec_list[i].handle;
1342 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1343 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1344 exec2_list[i].alignment = exec_list[i].alignment;
1345 exec2_list[i].offset = exec_list[i].offset;
1346 if (INTEL_INFO(dev)->gen < 4)
1347 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1349 exec2_list[i].flags = 0;
1352 exec2.buffers_ptr = args->buffers_ptr;
1353 exec2.buffer_count = args->buffer_count;
1354 exec2.batch_start_offset = args->batch_start_offset;
1355 exec2.batch_len = args->batch_len;
1356 exec2.DR1 = args->DR1;
1357 exec2.DR4 = args->DR4;
1358 exec2.num_cliprects = args->num_cliprects;
1359 exec2.cliprects_ptr = args->cliprects_ptr;
1360 exec2.flags = I915_EXEC_RENDER;
1361 i915_execbuffer2_set_context_id(exec2, 0);
1363 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1365 /* Copy the new buffer offsets back to the user's exec list. */
1366 for (i = 0; i < args->buffer_count; i++)
1367 exec_list[i].offset = exec2_list[i].offset;
1368 /* ... and back out to userspace */
1369 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1370 (uintptr_t) args->buffers_ptr,
1372 sizeof(*exec_list) * args->buffer_count);
1375 DRM_DEBUG("failed to copy %d exec entries "
1376 "back to user (%d)\n",
1377 args->buffer_count, ret);
1381 drm_free_large(exec_list);
1382 drm_free_large(exec2_list);
1387 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1388 struct drm_file *file)
1390 struct drm_i915_gem_execbuffer2 *args = data;
1391 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1394 if (args->buffer_count < 1 ||
1395 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1396 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1400 exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1401 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
1402 if (exec2_list == NULL)
1403 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1404 args->buffer_count);
1405 if (exec2_list == NULL) {
1406 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1407 args->buffer_count);
1410 ret = copy_from_user(exec2_list,
1411 (struct drm_i915_relocation_entry __user *)
1412 (uintptr_t) args->buffers_ptr,
1413 sizeof(*exec2_list) * args->buffer_count);
1415 DRM_DEBUG("copy %d exec entries failed %d\n",
1416 args->buffer_count, ret);
1417 drm_free_large(exec2_list);
1421 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1423 /* Copy the new buffer offsets back to the user's exec list. */
1424 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1425 (uintptr_t) args->buffers_ptr,
1427 sizeof(*exec2_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(exec2_list);