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>
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
32 #include <dev/drm2/drmP.h>
33 #include <dev/drm2/drm.h>
34 #include <dev/drm2/i915/i915_drm.h>
35 #include <dev/drm2/i915/i915_drv.h>
36 #include <dev/drm2/i915/intel_drv.h>
37 #include <sys/limits.h>
38 #include <sys/sf_buf.h>
40 struct change_domains {
41 uint32_t invalidate_domains;
42 uint32_t flush_domains;
48 * Set the next domain for the specified object. This
49 * may not actually perform the necessary flushing/invaliding though,
50 * as that may want to be batched with other set_domain operations
52 * This is (we hope) the only really tricky part of gem. The goal
53 * is fairly simple -- track which caches hold bits of the object
54 * and make sure they remain coherent. A few concrete examples may
55 * help to explain how it works. For shorthand, we use the notation
56 * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
57 * a pair of read and write domain masks.
59 * Case 1: the batch buffer
65 * 5. Unmapped from GTT
68 * Let's take these a step at a time
71 * Pages allocated from the kernel may still have
72 * cache contents, so we set them to (CPU, CPU) always.
73 * 2. Written by CPU (using pwrite)
74 * The pwrite function calls set_domain (CPU, CPU) and
75 * this function does nothing (as nothing changes)
77 * This function asserts that the object is not
78 * currently in any GPU-based read or write domains
80 * i915_gem_execbuffer calls set_domain (COMMAND, 0).
81 * As write_domain is zero, this function adds in the
82 * current read domains (CPU+COMMAND, 0).
83 * flush_domains is set to CPU.
84 * invalidate_domains is set to COMMAND
85 * clflush is run to get data out of the CPU caches
86 * then i915_dev_set_domain calls i915_gem_flush to
87 * emit an MI_FLUSH and drm_agp_chipset_flush
88 * 5. Unmapped from GTT
89 * i915_gem_object_unbind calls set_domain (CPU, CPU)
90 * flush_domains and invalidate_domains end up both zero
91 * so no flushing/invalidating happens
95 * Case 2: The shared render buffer
99 * 3. Read/written by GPU
100 * 4. set_domain to (CPU,CPU)
101 * 5. Read/written by CPU
102 * 6. Read/written by GPU
105 * Same as last example, (CPU, CPU)
107 * Nothing changes (assertions find that it is not in the GPU)
108 * 3. Read/written by GPU
109 * execbuffer calls set_domain (RENDER, RENDER)
110 * flush_domains gets CPU
111 * invalidate_domains gets GPU
113 * MI_FLUSH and drm_agp_chipset_flush
114 * 4. set_domain (CPU, CPU)
115 * flush_domains gets GPU
116 * invalidate_domains gets CPU
117 * wait_rendering (obj) to make sure all drawing is complete.
118 * This will include an MI_FLUSH to get the data from GPU
120 * clflush (obj) to invalidate the CPU cache
121 * Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
122 * 5. Read/written by CPU
123 * cache lines are loaded and dirtied
124 * 6. Read written by GPU
125 * Same as last GPU access
127 * Case 3: The constant buffer
132 * 4. Updated (written) by CPU again
141 * flush_domains = CPU
142 * invalidate_domains = RENDER
145 * drm_agp_chipset_flush
146 * 4. Updated (written) by CPU again
148 * flush_domains = 0 (no previous write domain)
149 * invalidate_domains = 0 (no new read domains)
152 * flush_domains = CPU
153 * invalidate_domains = RENDER
156 * drm_agp_chipset_flush
159 i915_gem_object_set_to_gpu_domain(struct drm_i915_gem_object *obj,
160 struct intel_ring_buffer *ring,
161 struct change_domains *cd)
163 uint32_t invalidate_domains = 0, flush_domains = 0;
166 * If the object isn't moving to a new write domain,
167 * let the object stay in multiple read domains
169 if (obj->base.pending_write_domain == 0)
170 obj->base.pending_read_domains |= obj->base.read_domains;
173 * Flush the current write domain if
174 * the new read domains don't match. Invalidate
175 * any read domains which differ from the old
178 if (obj->base.write_domain &&
179 (((obj->base.write_domain != obj->base.pending_read_domains ||
180 obj->ring != ring)) ||
181 (obj->fenced_gpu_access && !obj->pending_fenced_gpu_access))) {
182 flush_domains |= obj->base.write_domain;
183 invalidate_domains |=
184 obj->base.pending_read_domains & ~obj->base.write_domain;
187 * Invalidate any read caches which may have
188 * stale data. That is, any new read domains.
190 invalidate_domains |= obj->base.pending_read_domains & ~obj->base.read_domains;
191 if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU)
192 i915_gem_clflush_object(obj);
194 if (obj->base.pending_write_domain)
195 cd->flips |= atomic_load_acq_int(&obj->pending_flip);
197 /* The actual obj->write_domain will be updated with
198 * pending_write_domain after we emit the accumulated flush for all
199 * of our domain changes in execbuffers (which clears objects'
200 * write_domains). So if we have a current write domain that we
201 * aren't changing, set pending_write_domain to that.
203 if (flush_domains == 0 && obj->base.pending_write_domain == 0)
204 obj->base.pending_write_domain = obj->base.write_domain;
206 cd->invalidate_domains |= invalidate_domains;
207 cd->flush_domains |= flush_domains;
208 if (flush_domains & I915_GEM_GPU_DOMAINS)
209 cd->flush_rings |= intel_ring_flag(obj->ring);
210 if (invalidate_domains & I915_GEM_GPU_DOMAINS)
211 cd->flush_rings |= intel_ring_flag(ring);
216 LIST_HEAD(, drm_i915_gem_object) *buckets;
219 static struct eb_objects *
222 struct eb_objects *eb;
224 eb = malloc(sizeof(*eb), DRM_I915_GEM, M_WAITOK | M_ZERO);
225 eb->buckets = hashinit(size, DRM_I915_GEM, &eb->hashmask);
230 eb_reset(struct eb_objects *eb)
234 for (i = 0; i <= eb->hashmask; i++)
235 LIST_INIT(&eb->buckets[i]);
239 eb_add_object(struct eb_objects *eb, struct drm_i915_gem_object *obj)
242 LIST_INSERT_HEAD(&eb->buckets[obj->exec_handle & eb->hashmask],
246 static struct drm_i915_gem_object *
247 eb_get_object(struct eb_objects *eb, unsigned long handle)
249 struct drm_i915_gem_object *obj;
251 LIST_FOREACH(obj, &eb->buckets[handle & eb->hashmask], exec_node) {
252 if (obj->exec_handle == handle)
259 eb_destroy(struct eb_objects *eb)
262 free(eb->buckets, DRM_I915_GEM);
263 free(eb, DRM_I915_GEM);
266 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
268 return (obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
269 obj->cache_level != I915_CACHE_NONE);
273 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
274 struct eb_objects *eb,
275 struct drm_i915_gem_relocation_entry *reloc)
277 struct drm_device *dev = obj->base.dev;
278 struct drm_gem_object *target_obj;
279 struct drm_i915_gem_object *target_i915_obj;
280 uint32_t target_offset;
283 /* we've already hold a reference to all valid objects */
284 target_obj = &eb_get_object(eb, reloc->target_handle)->base;
285 if (unlikely(target_obj == NULL))
288 target_i915_obj = to_intel_bo(target_obj);
289 target_offset = target_i915_obj->gtt_offset;
292 DRM_INFO("%s: obj %p offset %08x target %d "
293 "read %08x write %08x gtt %08x "
294 "presumed %08x delta %08x\n",
298 (int) reloc->target_handle,
299 (int) reloc->read_domains,
300 (int) reloc->write_domain,
302 (int) reloc->presumed_offset,
306 /* The target buffer should have appeared before us in the
307 * exec_object list, so it should have a GTT space bound by now.
309 if (unlikely(target_offset == 0)) {
310 DRM_DEBUG("No GTT space found for object %d\n",
311 reloc->target_handle);
315 /* Validate that the target is in a valid r/w GPU domain */
316 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
317 DRM_DEBUG("reloc with multiple write domains: "
318 "obj %p target %d offset %d "
319 "read %08x write %08x",
320 obj, reloc->target_handle,
323 reloc->write_domain);
326 if (unlikely((reloc->write_domain | reloc->read_domains)
327 & ~I915_GEM_GPU_DOMAINS)) {
328 DRM_DEBUG("reloc with read/write non-GPU domains: "
329 "obj %p target %d offset %d "
330 "read %08x write %08x",
331 obj, reloc->target_handle,
334 reloc->write_domain);
337 if (unlikely(reloc->write_domain && target_obj->pending_write_domain &&
338 reloc->write_domain != target_obj->pending_write_domain)) {
339 DRM_DEBUG("Write domain conflict: "
340 "obj %p target %d offset %d "
341 "new %08x old %08x\n",
342 obj, reloc->target_handle,
345 target_obj->pending_write_domain);
349 target_obj->pending_read_domains |= reloc->read_domains;
350 target_obj->pending_write_domain |= reloc->write_domain;
352 /* If the relocation already has the right value in it, no
353 * more work needs to be done.
355 if (target_offset == reloc->presumed_offset)
358 /* Check that the relocation address is valid... */
359 if (unlikely(reloc->offset > obj->base.size - 4)) {
360 DRM_DEBUG("Relocation beyond object bounds: "
361 "obj %p target %d offset %d size %d.\n",
362 obj, reloc->target_handle,
364 (int) obj->base.size);
367 if (unlikely(reloc->offset & 3)) {
368 DRM_DEBUG("Relocation not 4-byte aligned: "
369 "obj %p target %d offset %d.\n",
370 obj, reloc->target_handle,
371 (int) reloc->offset);
375 /* We can't wait for rendering with pagefaults disabled */
376 if (obj->active && (curthread->td_pflags & TDP_NOFAULTING) != 0)
379 reloc->delta += target_offset;
380 if (use_cpu_reloc(obj)) {
381 uint32_t page_offset = reloc->offset & PAGE_MASK;
385 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
389 sf = sf_buf_alloc(obj->pages[OFF_TO_IDX(reloc->offset)],
393 vaddr = (void *)sf_buf_kva(sf);
394 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
397 uint32_t *reloc_entry;
400 ret = i915_gem_object_set_to_gtt_domain(obj, true);
404 ret = i915_gem_object_put_fence(obj);
409 * Map the page containing the relocation we're going
412 reloc->offset += obj->gtt_offset;
413 reloc_page = pmap_mapdev_attr(dev->agp->base + (reloc->offset &
414 ~PAGE_MASK), PAGE_SIZE, PAT_WRITE_COMBINING);
415 reloc_entry = (uint32_t *)(reloc_page + (reloc->offset &
417 *(volatile uint32_t *)reloc_entry = reloc->delta;
418 pmap_unmapdev((vm_offset_t)reloc_page, PAGE_SIZE);
421 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
422 * pipe_control writes because the gpu doesn't properly redirect them
423 * through the ppgtt for non_secure batchbuffers. */
424 if (unlikely(IS_GEN6(dev) &&
425 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
426 !target_i915_obj->has_global_gtt_mapping)) {
427 i915_gem_gtt_bind_object(target_i915_obj,
428 target_i915_obj->cache_level);
431 /* and update the user's relocation entry */
432 reloc->presumed_offset = target_offset;
438 i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
439 struct eb_objects *eb)
441 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
442 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
443 struct drm_i915_gem_relocation_entry *user_relocs;
444 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
447 user_relocs = (void *)(uintptr_t)entry->relocs_ptr;
448 remain = entry->relocation_count;
450 struct drm_i915_gem_relocation_entry *r = stack_reloc;
452 if (count > DRM_ARRAY_SIZE(stack_reloc))
453 count = DRM_ARRAY_SIZE(stack_reloc);
456 ret = -copyin_nofault(user_relocs, r, count*sizeof(r[0]));
461 u64 offset = r->presumed_offset;
463 ret = i915_gem_execbuffer_relocate_entry(obj, eb, r);
467 if (r->presumed_offset != offset &&
468 copyout_nofault(&r->presumed_offset,
469 &user_relocs->presumed_offset,
470 sizeof(r->presumed_offset))) {
483 i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
484 struct eb_objects *eb, struct drm_i915_gem_relocation_entry *relocs)
486 const struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
489 for (i = 0; i < entry->relocation_count; i++) {
490 ret = i915_gem_execbuffer_relocate_entry(obj, eb, &relocs[i]);
499 i915_gem_execbuffer_relocate(struct drm_device *dev,
500 struct eb_objects *eb,
501 struct list_head *objects)
503 struct drm_i915_gem_object *obj;
506 /* Try to move as many of the relocation targets off the active list
507 * to avoid unnecessary fallbacks to the slow path, as we cannot wait
508 * for the retirement with pagefaults disabled.
510 i915_gem_retire_requests(dev);
513 pflags = vm_fault_disable_pagefaults();
514 /* This is the fast path and we cannot handle a pagefault whilst
515 * holding the device lock lest the user pass in the relocations
516 * contained within a mmaped bo. For in such a case we, the page
517 * fault handler would call i915_gem_fault() and we would try to
518 * acquire the device lock again. Obviously this is bad.
521 list_for_each_entry(obj, objects, exec_list) {
522 ret = i915_gem_execbuffer_relocate_object(obj, eb);
526 vm_fault_enable_pagefaults(pflags);
530 #define __EXEC_OBJECT_HAS_FENCE (1<<31)
533 need_reloc_mappable(struct drm_i915_gem_object *obj)
535 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
536 return entry->relocation_count && !use_cpu_reloc(obj);
540 pin_and_fence_object(struct drm_i915_gem_object *obj,
541 struct intel_ring_buffer *ring)
543 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
544 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
545 bool need_fence, need_mappable;
549 has_fenced_gpu_access &&
550 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
551 obj->tiling_mode != I915_TILING_NONE;
552 need_mappable = need_fence || need_reloc_mappable(obj);
554 ret = i915_gem_object_pin(obj, entry->alignment, need_mappable);
558 if (has_fenced_gpu_access) {
559 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
560 ret = i915_gem_object_get_fence(obj);
564 if (i915_gem_object_pin_fence(obj))
565 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
567 obj->pending_fenced_gpu_access = true;
571 entry->offset = obj->gtt_offset;
575 i915_gem_object_unpin(obj);
580 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
581 struct drm_file *file,
582 struct list_head *objects)
584 drm_i915_private_t *dev_priv;
585 struct drm_i915_gem_object *obj;
587 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
588 struct list_head ordered_objects;
590 dev_priv = ring->dev->dev_private;
591 INIT_LIST_HEAD(&ordered_objects);
592 while (!list_empty(objects)) {
593 struct drm_i915_gem_exec_object2 *entry;
594 bool need_fence, need_mappable;
596 obj = list_first_entry(objects,
597 struct drm_i915_gem_object,
599 entry = obj->exec_entry;
602 has_fenced_gpu_access &&
603 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
604 obj->tiling_mode != I915_TILING_NONE;
605 need_mappable = need_fence || need_reloc_mappable(obj);
608 list_move(&obj->exec_list, &ordered_objects);
610 list_move_tail(&obj->exec_list, &ordered_objects);
612 obj->base.pending_read_domains = 0;
613 obj->base.pending_write_domain = 0;
615 list_splice(&ordered_objects, objects);
617 /* Attempt to pin all of the buffers into the GTT.
618 * This is done in 3 phases:
620 * 1a. Unbind all objects that do not match the GTT constraints for
621 * the execbuffer (fenceable, mappable, alignment etc).
622 * 1b. Increment pin count for already bound objects and obtain
623 * a fence register if required.
624 * 2. Bind new objects.
625 * 3. Decrement pin count.
627 * This avoid unnecessary unbinding of later objects in order to makr
628 * room for the earlier objects *unless* we need to defragment.
634 /* Unbind any ill-fitting objects or pin. */
635 list_for_each_entry(obj, objects, exec_list) {
636 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
637 bool need_fence, need_mappable;
643 has_fenced_gpu_access &&
644 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
645 obj->tiling_mode != I915_TILING_NONE;
646 need_mappable = need_fence || need_reloc_mappable(obj);
648 if ((entry->alignment && obj->gtt_offset & (entry->alignment - 1)) ||
649 (need_mappable && !obj->map_and_fenceable))
650 ret = i915_gem_object_unbind(obj);
652 ret = pin_and_fence_object(obj, ring);
657 /* Bind fresh objects */
658 list_for_each_entry(obj, objects, exec_list) {
662 ret = pin_and_fence_object(obj, ring);
666 /* This can potentially raise a harmless
667 * -EINVAL if we failed to bind in the above
668 * call. It cannot raise -EINTR since we know
669 * that the bo is freshly bound and so will
670 * not need to be flushed or waited upon.
672 ret_ignore = i915_gem_object_unbind(obj);
674 if (obj->gtt_space != NULL)
675 printf("%s: gtt_space\n", __func__);
680 /* Decrement pin count for bound objects */
681 list_for_each_entry(obj, objects, exec_list) {
682 struct drm_i915_gem_exec_object2 *entry;
687 entry = obj->exec_entry;
688 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
689 i915_gem_object_unpin_fence(obj);
690 entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
693 i915_gem_object_unpin(obj);
695 /* ... and ensure ppgtt mapping exist if needed. */
696 if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) {
697 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
698 obj, obj->cache_level);
700 obj->has_aliasing_ppgtt_mapping = 1;
704 if (ret != -ENOSPC || retry > 1)
707 /* First attempt, just clear anything that is purgeable.
708 * Second attempt, clear the entire GTT.
710 ret = i915_gem_evict_everything(ring->dev, retry == 0);
718 list_for_each_entry_continue_reverse(obj, objects, exec_list) {
719 struct drm_i915_gem_exec_object2 *entry;
724 entry = obj->exec_entry;
725 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
726 i915_gem_object_unpin_fence(obj);
727 entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
730 i915_gem_object_unpin(obj);
737 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
738 struct drm_file *file, struct intel_ring_buffer *ring,
739 struct list_head *objects, struct eb_objects *eb,
740 struct drm_i915_gem_exec_object2 *exec, int count)
742 struct drm_i915_gem_relocation_entry *reloc;
743 struct drm_i915_gem_object *obj;
747 /* We may process another execbuffer during the unlock... */
748 while (!list_empty(objects)) {
749 obj = list_first_entry(objects,
750 struct drm_i915_gem_object,
752 list_del_init(&obj->exec_list);
753 drm_gem_object_unreference(&obj->base);
759 for (i = 0; i < count; i++)
760 total += exec[i].relocation_count;
762 reloc_offset = malloc(count * sizeof(*reloc_offset), DRM_I915_GEM,
764 reloc = malloc(total * sizeof(*reloc), DRM_I915_GEM, M_WAITOK | M_ZERO);
767 for (i = 0; i < count; i++) {
768 struct drm_i915_gem_relocation_entry *user_relocs;
770 user_relocs = (void *)(uintptr_t)exec[i].relocs_ptr;
771 ret = -copyin(user_relocs, reloc + total,
772 exec[i].relocation_count * sizeof(*reloc));
778 reloc_offset[i] = total;
779 total += exec[i].relocation_count;
782 ret = i915_mutex_lock_interruptible(dev);
788 /* reacquire the objects */
790 for (i = 0; i < count; i++) {
791 struct drm_i915_gem_object *obj;
793 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
795 if (&obj->base == NULL) {
796 DRM_DEBUG("Invalid object handle %d at index %d\n",
802 list_add_tail(&obj->exec_list, objects);
803 obj->exec_handle = exec[i].handle;
804 obj->exec_entry = &exec[i];
805 eb_add_object(eb, obj);
808 ret = i915_gem_execbuffer_reserve(ring, file, objects);
812 list_for_each_entry(obj, objects, exec_list) {
813 int offset = obj->exec_entry - exec;
814 ret = i915_gem_execbuffer_relocate_object_slow(obj, eb,
815 reloc + reloc_offset[offset]);
820 /* Leave the user relocations as are, this is the painfully slow path,
821 * and we want to avoid the complication of dropping the lock whilst
822 * having buffers reserved in the aperture and so causing spurious
823 * ENOSPC for random operations.
827 free(reloc, DRM_I915_GEM);
828 free(reloc_offset, DRM_I915_GEM);
833 i915_gem_execbuffer_flush(struct drm_device *dev,
834 uint32_t invalidate_domains,
835 uint32_t flush_domains,
836 uint32_t flush_rings)
838 drm_i915_private_t *dev_priv = dev->dev_private;
841 if (flush_domains & I915_GEM_DOMAIN_CPU)
842 intel_gtt_chipset_flush();
844 if (flush_domains & I915_GEM_DOMAIN_GTT)
847 if ((flush_domains | invalidate_domains) & I915_GEM_GPU_DOMAINS) {
848 for (i = 0; i < I915_NUM_RINGS; i++)
849 if (flush_rings & (1 << i)) {
850 ret = i915_gem_flush_ring(&dev_priv->rings[i],
851 invalidate_domains, flush_domains);
861 i915_gem_execbuffer_wait_for_flips(struct intel_ring_buffer *ring, u32 flips)
863 u32 plane, flip_mask;
866 /* Check for any pending flips. As we only maintain a flip queue depth
867 * of 1, we can simply insert a WAIT for the next display flip prior
868 * to executing the batch and avoid stalling the CPU.
871 for (plane = 0; flips >> plane; plane++) {
872 if (((flips >> plane) & 1) == 0)
876 flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
878 flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
880 ret = intel_ring_begin(ring, 2);
884 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | flip_mask);
885 intel_ring_emit(ring, MI_NOOP);
886 intel_ring_advance(ring);
893 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
894 struct list_head *objects)
896 struct drm_i915_gem_object *obj;
897 struct change_domains cd;
900 memset(&cd, 0, sizeof(cd));
901 list_for_each_entry(obj, objects, exec_list)
902 i915_gem_object_set_to_gpu_domain(obj, ring, &cd);
904 if (cd.invalidate_domains | cd.flush_domains) {
906 DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
908 cd.invalidate_domains,
911 ret = i915_gem_execbuffer_flush(ring->dev,
912 cd.invalidate_domains,
920 ret = i915_gem_execbuffer_wait_for_flips(ring, cd.flips);
925 list_for_each_entry(obj, objects, exec_list) {
926 ret = i915_gem_object_sync(obj, ring);
935 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
937 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
941 validate_exec_list(struct drm_i915_gem_exec_object2 *exec, int count,
945 int i, length, page_count;
947 /* XXXKIB various limits checking is missing there */
948 *map = malloc(count * sizeof(*ma), DRM_I915_GEM, M_WAITOK | M_ZERO);
949 for (i = 0; i < count; i++) {
950 /* First check for malicious input causing overflow */
951 if (exec[i].relocation_count >
952 INT_MAX / sizeof(struct drm_i915_gem_relocation_entry))
955 length = exec[i].relocation_count *
956 sizeof(struct drm_i915_gem_relocation_entry);
962 * Since both start and end of the relocation region
963 * may be not aligned on the page boundary, be
964 * conservative and request a page slot for each
965 * partial page. Thus +2.
967 page_count = howmany(length, PAGE_SIZE) + 2;
968 ma = (*map)[i] = malloc(page_count * sizeof(vm_page_t),
969 DRM_I915_GEM, M_WAITOK | M_ZERO);
970 if (vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
971 exec[i].relocs_ptr, length, VM_PROT_READ | VM_PROT_WRITE,
972 ma, page_count) == -1) {
973 free(ma, DRM_I915_GEM);
983 i915_gem_execbuffer_move_to_active(struct list_head *objects,
984 struct intel_ring_buffer *ring,
987 struct drm_i915_gem_object *obj;
988 uint32_t old_read, old_write;
990 list_for_each_entry(obj, objects, exec_list) {
991 old_read = obj->base.read_domains;
992 old_write = obj->base.write_domain;
994 obj->base.read_domains = obj->base.pending_read_domains;
995 obj->base.write_domain = obj->base.pending_write_domain;
996 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
998 i915_gem_object_move_to_active(obj, ring, seqno);
999 if (obj->base.write_domain) {
1001 obj->pending_gpu_write = true;
1002 list_move_tail(&obj->gpu_write_list,
1003 &ring->gpu_write_list);
1004 if (obj->pin_count) /* check for potential scanout */
1005 intel_mark_busy(ring->dev, obj);
1007 CTR3(KTR_DRM, "object_change_domain move_to_active %p %x %x",
1008 obj, old_read, old_write);
1011 intel_mark_busy(ring->dev, NULL);
1014 int i915_gem_sync_exec_requests;
1017 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
1018 struct drm_file *file,
1019 struct intel_ring_buffer *ring)
1021 struct drm_i915_gem_request *request;
1025 * Ensure that the commands in the batch buffer are
1026 * finished before the interrupt fires.
1028 * The sampler always gets flushed on i965 (sigh).
1030 invalidate = I915_GEM_DOMAIN_COMMAND;
1031 if (INTEL_INFO(dev)->gen >= 4)
1032 invalidate |= I915_GEM_DOMAIN_SAMPLER;
1033 if (ring->flush(ring, invalidate, 0)) {
1034 i915_gem_next_request_seqno(ring);
1038 /* Add a breadcrumb for the completion of the batch buffer */
1039 request = malloc(sizeof(*request), DRM_I915_GEM, M_WAITOK | M_ZERO);
1040 if (request == NULL || i915_add_request(ring, file, request)) {
1041 i915_gem_next_request_seqno(ring);
1042 free(request, DRM_I915_GEM);
1043 } else if (i915_gem_sync_exec_requests) {
1044 i915_wait_request(ring, request->seqno);
1045 i915_gem_retire_requests(dev);
1050 i915_gem_fix_mi_batchbuffer_end(struct drm_i915_gem_object *batch_obj,
1051 uint32_t batch_start_offset, uint32_t batch_len)
1054 uint64_t po_r, po_w;
1057 po_r = batch_obj->base.dev->agp->base + batch_obj->gtt_offset +
1058 batch_start_offset + batch_len;
1061 mkva = pmap_mapdev_attr(trunc_page(po_r), 2 * PAGE_SIZE,
1062 PAT_WRITE_COMBINING);
1064 cmd = *(uint32_t *)(mkva + po_r);
1066 if (cmd != MI_BATCH_BUFFER_END) {
1068 * batch_len != 0 due to the check at the start of
1069 * i915_gem_do_execbuffer
1071 if (batch_obj->base.size > batch_start_offset + batch_len) {
1073 /* DRM_DEBUG("batchbuffer does not end by MI_BATCH_BUFFER_END !\n"); */
1076 DRM_DEBUG("batchbuffer does not end by MI_BATCH_BUFFER_END, overwriting last bo cmd !\n");
1078 *(uint32_t *)(mkva + po_w) = MI_BATCH_BUFFER_END;
1081 pmap_unmapdev((vm_offset_t)mkva, 2 * PAGE_SIZE);
1084 int i915_fix_mi_batchbuffer_end = 0;
1087 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1088 struct intel_ring_buffer *ring)
1090 drm_i915_private_t *dev_priv = dev->dev_private;
1093 if (!IS_GEN7(dev) || ring != &dev_priv->rings[RCS])
1096 ret = intel_ring_begin(ring, 4 * 3);
1100 for (i = 0; i < 4; i++) {
1101 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1102 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1103 intel_ring_emit(ring, 0);
1106 intel_ring_advance(ring);
1112 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1113 struct drm_file *file,
1114 struct drm_i915_gem_execbuffer2 *args,
1115 struct drm_i915_gem_exec_object2 *exec)
1117 drm_i915_private_t *dev_priv = dev->dev_private;
1118 struct list_head objects;
1119 struct eb_objects *eb;
1120 struct drm_i915_gem_object *batch_obj;
1121 struct drm_clip_rect *cliprects = NULL;
1122 struct intel_ring_buffer *ring;
1123 vm_page_t **relocs_ma;
1124 u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1125 u32 exec_start, exec_len;
1130 if (!i915_gem_check_execbuffer(args)) {
1131 DRM_DEBUG("execbuf with invalid offset/length\n");
1135 if (args->batch_len == 0)
1138 ret = validate_exec_list(exec, args->buffer_count, &relocs_ma);
1140 goto pre_struct_lock_err;
1142 switch (args->flags & I915_EXEC_RING_MASK) {
1143 case I915_EXEC_DEFAULT:
1144 case I915_EXEC_RENDER:
1145 ring = &dev_priv->rings[RCS];
1148 ring = &dev_priv->rings[VCS];
1150 DRM_DEBUG("Ring %s doesn't support contexts\n",
1156 ring = &dev_priv->rings[BCS];
1158 DRM_DEBUG("Ring %s doesn't support contexts\n",
1164 DRM_DEBUG("execbuf with unknown ring: %d\n",
1165 (int)(args->flags & I915_EXEC_RING_MASK));
1167 goto pre_struct_lock_err;
1169 if (!intel_ring_initialized(ring)) {
1170 DRM_DEBUG("execbuf with invalid ring: %d\n",
1171 (int)(args->flags & I915_EXEC_RING_MASK));
1175 mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1176 mask = I915_EXEC_CONSTANTS_MASK;
1178 case I915_EXEC_CONSTANTS_REL_GENERAL:
1179 case I915_EXEC_CONSTANTS_ABSOLUTE:
1180 case I915_EXEC_CONSTANTS_REL_SURFACE:
1181 if (ring == &dev_priv->rings[RCS] &&
1182 mode != dev_priv->relative_constants_mode) {
1183 if (INTEL_INFO(dev)->gen < 4) {
1185 goto pre_struct_lock_err;
1188 if (INTEL_INFO(dev)->gen > 5 &&
1189 mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1191 goto pre_struct_lock_err;
1194 /* The HW changed the meaning on this bit on gen6 */
1195 if (INTEL_INFO(dev)->gen >= 6)
1196 mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1200 DRM_DEBUG("execbuf with unknown constants: %d\n", mode);
1202 goto pre_struct_lock_err;
1205 if (args->buffer_count < 1) {
1206 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1208 goto pre_struct_lock_err;
1211 if (args->num_cliprects != 0) {
1212 if (ring != &dev_priv->rings[RCS]) {
1213 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1215 goto pre_struct_lock_err;
1218 if (INTEL_INFO(dev)->gen >= 5) {
1219 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1221 goto pre_struct_lock_err;
1224 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1225 DRM_DEBUG("execbuf with %u cliprects\n",
1226 args->num_cliprects);
1228 goto pre_struct_lock_err;
1230 cliprects = malloc( sizeof(*cliprects) * args->num_cliprects,
1231 DRM_I915_GEM, M_WAITOK | M_ZERO);
1232 ret = -copyin((void *)(uintptr_t)args->cliprects_ptr, cliprects,
1233 sizeof(*cliprects) * args->num_cliprects);
1235 goto pre_struct_lock_err;
1238 ret = i915_mutex_lock_interruptible(dev);
1240 goto pre_struct_lock_err;
1242 if (dev_priv->mm.suspended) {
1245 goto pre_struct_lock_err;
1248 eb = eb_create(args->buffer_count);
1252 goto pre_struct_lock_err;
1255 /* Look up object handles */
1256 INIT_LIST_HEAD(&objects);
1257 for (i = 0; i < args->buffer_count; i++) {
1258 struct drm_i915_gem_object *obj;
1259 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
1261 if (&obj->base == NULL) {
1262 DRM_DEBUG("Invalid object handle %d at index %d\n",
1264 /* prevent error path from reading uninitialized data */
1269 if (!list_empty(&obj->exec_list)) {
1270 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
1271 obj, exec[i].handle, i);
1276 list_add_tail(&obj->exec_list, &objects);
1277 obj->exec_handle = exec[i].handle;
1278 obj->exec_entry = &exec[i];
1279 eb_add_object(eb, obj);
1282 /* take note of the batch buffer before we might reorder the lists */
1283 batch_obj = list_entry(objects.prev,
1284 struct drm_i915_gem_object,
1287 /* Move the objects en-masse into the GTT, evicting if necessary. */
1288 ret = i915_gem_execbuffer_reserve(ring, file, &objects);
1292 /* The objects are in their final locations, apply the relocations. */
1293 ret = i915_gem_execbuffer_relocate(dev, eb, &objects);
1295 if (ret == -EFAULT) {
1296 ret = i915_gem_execbuffer_relocate_slow(dev, file, ring,
1297 &objects, eb, exec, args->buffer_count);
1298 DRM_LOCK_ASSERT(dev);
1304 /* Set the pending read domains for the batch buffer to COMMAND */
1305 if (batch_obj->base.pending_write_domain) {
1306 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1310 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1312 ret = i915_gem_execbuffer_move_to_gpu(ring, &objects);
1316 ret = i915_switch_context(ring, file, ctx_id);
1320 seqno = i915_gem_next_request_seqno(ring);
1321 for (i = 0; i < I915_NUM_RINGS - 1; i++) {
1322 if (seqno < ring->sync_seqno[i]) {
1323 /* The GPU can not handle its semaphore value wrapping,
1324 * so every billion or so execbuffers, we need to stall
1325 * the GPU in order to reset the counters.
1327 ret = i915_gpu_idle(dev);
1330 i915_gem_retire_requests(dev);
1332 KASSERT(ring->sync_seqno[i] == 0, ("Non-zero sync_seqno"));
1336 if (ring == &dev_priv->rings[RCS] &&
1337 mode != dev_priv->relative_constants_mode) {
1338 ret = intel_ring_begin(ring, 4);
1342 intel_ring_emit(ring, MI_NOOP);
1343 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1344 intel_ring_emit(ring, INSTPM);
1345 intel_ring_emit(ring, mask << 16 | mode);
1346 intel_ring_advance(ring);
1348 dev_priv->relative_constants_mode = mode;
1351 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1352 ret = i915_reset_gen7_sol_offsets(dev, ring);
1357 exec_start = batch_obj->gtt_offset + args->batch_start_offset;
1358 exec_len = args->batch_len;
1360 if (i915_fix_mi_batchbuffer_end) {
1361 i915_gem_fix_mi_batchbuffer_end(batch_obj,
1362 args->batch_start_offset, args->batch_len);
1365 CTR4(KTR_DRM, "ring_dispatch %s %d exec %x %x", ring->name, seqno,
1366 exec_start, exec_len);
1369 for (i = 0; i < args->num_cliprects; i++) {
1370 ret = i915_emit_box_p(dev, &cliprects[i],
1371 args->DR1, args->DR4);
1375 ret = ring->dispatch_execbuffer(ring, exec_start,
1381 ret = ring->dispatch_execbuffer(ring, exec_start, exec_len);
1386 i915_gem_execbuffer_move_to_active(&objects, ring, seqno);
1387 i915_gem_execbuffer_retire_commands(dev, file, ring);
1391 while (!list_empty(&objects)) {
1392 struct drm_i915_gem_object *obj;
1394 obj = list_first_entry(&objects, struct drm_i915_gem_object,
1396 list_del_init(&obj->exec_list);
1397 drm_gem_object_unreference(&obj->base);
1401 pre_struct_lock_err:
1402 for (i = 0; i < args->buffer_count; i++) {
1403 if (relocs_ma[i] != NULL) {
1404 vm_page_unhold_pages(relocs_ma[i], howmany(
1405 exec[i].relocation_count *
1406 sizeof(struct drm_i915_gem_relocation_entry),
1408 free(relocs_ma[i], DRM_I915_GEM);
1411 free(relocs_ma, DRM_I915_GEM);
1412 free(cliprects, DRM_I915_GEM);
1417 * Legacy execbuffer just creates an exec2 list from the original exec object
1418 * list array and passes it to the real function.
1421 i915_gem_execbuffer(struct drm_device *dev, void *data,
1422 struct drm_file *file)
1424 struct drm_i915_gem_execbuffer *args = data;
1425 struct drm_i915_gem_execbuffer2 exec2;
1426 struct drm_i915_gem_exec_object *exec_list = NULL;
1427 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1430 DRM_DEBUG("buffers_ptr %d buffer_count %d len %08x\n",
1431 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
1433 if (args->buffer_count < 1) {
1434 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1438 /* Copy in the exec list from userland */
1439 /* XXXKIB user-controlled malloc size */
1440 exec_list = malloc(sizeof(*exec_list) * args->buffer_count,
1441 DRM_I915_GEM, M_WAITOK);
1442 exec2_list = malloc(sizeof(*exec2_list) * args->buffer_count,
1443 DRM_I915_GEM, M_WAITOK);
1444 ret = -copyin((void *)(uintptr_t)args->buffers_ptr, exec_list,
1445 sizeof(*exec_list) * args->buffer_count);
1447 DRM_DEBUG("copy %d exec entries failed %d\n",
1448 args->buffer_count, ret);
1449 free(exec_list, DRM_I915_GEM);
1450 free(exec2_list, DRM_I915_GEM);
1454 for (i = 0; i < args->buffer_count; i++) {
1455 exec2_list[i].handle = exec_list[i].handle;
1456 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1457 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1458 exec2_list[i].alignment = exec_list[i].alignment;
1459 exec2_list[i].offset = exec_list[i].offset;
1460 if (INTEL_INFO(dev)->gen < 4)
1461 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1463 exec2_list[i].flags = 0;
1466 exec2.buffers_ptr = args->buffers_ptr;
1467 exec2.buffer_count = args->buffer_count;
1468 exec2.batch_start_offset = args->batch_start_offset;
1469 exec2.batch_len = args->batch_len;
1470 exec2.DR1 = args->DR1;
1471 exec2.DR4 = args->DR4;
1472 exec2.num_cliprects = args->num_cliprects;
1473 exec2.cliprects_ptr = args->cliprects_ptr;
1474 exec2.flags = I915_EXEC_RENDER;
1475 i915_execbuffer2_set_context_id(exec2, 0);
1477 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1479 /* Copy the new buffer offsets back to the user's exec list. */
1480 for (i = 0; i < args->buffer_count; i++)
1481 exec_list[i].offset = exec2_list[i].offset;
1482 /* ... and back out to userspace */
1483 ret = -copyout(exec_list, (void *)(uintptr_t)args->buffers_ptr,
1484 sizeof(*exec_list) * args->buffer_count);
1486 DRM_DEBUG("failed to copy %d exec entries "
1487 "back to user (%d)\n",
1488 args->buffer_count, ret);
1492 free(exec_list, DRM_I915_GEM);
1493 free(exec2_list, DRM_I915_GEM);
1498 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1499 struct drm_file *file)
1501 struct drm_i915_gem_execbuffer2 *args = data;
1502 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1505 DRM_DEBUG("buffers_ptr %jx buffer_count %d len %08x\n",
1506 (uintmax_t)args->buffers_ptr, args->buffer_count, args->batch_len);
1508 if (args->buffer_count < 1 ||
1509 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1510 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1514 /* XXXKIB user-controllable malloc size */
1515 exec2_list = malloc(sizeof(*exec2_list) * args->buffer_count,
1516 DRM_I915_GEM, M_WAITOK);
1517 ret = -copyin((void *)(uintptr_t)args->buffers_ptr, exec2_list,
1518 sizeof(*exec2_list) * args->buffer_count);
1520 DRM_DEBUG("copy %d exec entries failed %d\n",
1521 args->buffer_count, ret);
1522 free(exec2_list, DRM_I915_GEM);
1526 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1528 /* Copy the new buffer offsets back to the user's exec list. */
1529 ret = -copyout(exec2_list, (void *)(uintptr_t)args->buffers_ptr,
1530 sizeof(*exec2_list) * args->buffer_count);
1532 DRM_DEBUG("failed to copy %d exec entries "
1533 "back to user (%d)\n",
1534 args->buffer_count, ret);
1538 free(exec2_list, DRM_I915_GEM);