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;
211 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
212 struct drm_file *file_priv,
213 struct drm_i915_gem_exec_object2 *entry,
214 struct drm_i915_gem_relocation_entry *reloc)
216 struct drm_device *dev = obj->base.dev;
217 struct drm_gem_object *target_obj;
218 uint32_t target_offset;
221 target_obj = drm_gem_object_lookup(dev, file_priv,
222 reloc->target_handle);
223 if (target_obj == NULL)
226 target_offset = to_intel_bo(target_obj)->gtt_offset;
229 DRM_INFO("%s: obj %p offset %08x target %d "
230 "read %08x write %08x gtt %08x "
231 "presumed %08x delta %08x\n",
235 (int) reloc->target_handle,
236 (int) reloc->read_domains,
237 (int) reloc->write_domain,
239 (int) reloc->presumed_offset,
243 /* The target buffer should have appeared before us in the
244 * exec_object list, so it should have a GTT space bound by now.
246 if (target_offset == 0) {
247 DRM_ERROR("No GTT space found for object %d\n",
248 reloc->target_handle);
252 /* Validate that the target is in a valid r/w GPU domain */
253 if (reloc->write_domain & (reloc->write_domain - 1)) {
254 DRM_ERROR("reloc with multiple write domains: "
255 "obj %p target %d offset %d "
256 "read %08x write %08x",
257 obj, reloc->target_handle,
260 reloc->write_domain);
263 if (reloc->write_domain & I915_GEM_DOMAIN_CPU ||
264 reloc->read_domains & I915_GEM_DOMAIN_CPU) {
265 DRM_ERROR("reloc with read/write CPU domains: "
266 "obj %p target %d offset %d "
267 "read %08x write %08x",
268 obj, reloc->target_handle,
271 reloc->write_domain);
274 if (reloc->write_domain && target_obj->pending_write_domain &&
275 reloc->write_domain != target_obj->pending_write_domain) {
276 DRM_ERROR("Write domain conflict: "
277 "obj %p target %d offset %d "
278 "new %08x old %08x\n",
279 obj, reloc->target_handle,
282 target_obj->pending_write_domain);
286 target_obj->pending_read_domains |= reloc->read_domains;
287 target_obj->pending_write_domain |= reloc->write_domain;
289 /* If the relocation already has the right value in it, no
290 * more work needs to be done.
292 if (target_offset == reloc->presumed_offset)
295 /* Check that the relocation address is valid... */
296 if (reloc->offset > obj->base.size - 4) {
297 DRM_ERROR("Relocation beyond object bounds: "
298 "obj %p target %d offset %d size %d.\n",
299 obj, reloc->target_handle,
301 (int) obj->base.size);
304 if (reloc->offset & 3) {
305 DRM_ERROR("Relocation not 4-byte aligned: "
306 "obj %p target %d offset %d.\n",
307 obj, reloc->target_handle,
308 (int) reloc->offset);
312 /* and points to somewhere within the target object. */
313 if (reloc->delta >= target_obj->size) {
314 DRM_ERROR("Relocation beyond target object bounds: "
315 "obj %p target %d delta %d size %d.\n",
316 obj, reloc->target_handle,
318 (int) target_obj->size);
322 reloc->delta += target_offset;
323 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) {
324 uint32_t page_offset = reloc->offset & ~PAGE_MASK;
327 vaddr = kmap_atomic(obj->pages[reloc->offset >> PAGE_SHIFT]);
328 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
329 kunmap_atomic(vaddr);
331 struct drm_i915_private *dev_priv = dev->dev_private;
332 uint32_t __iomem *reloc_entry;
333 void __iomem *reloc_page;
335 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
339 /* Map the page containing the relocation we're going to perform. */
340 reloc->offset += obj->gtt_offset;
341 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
342 reloc->offset & PAGE_MASK);
343 reloc_entry = (uint32_t __iomem *)
344 (reloc_page + (reloc->offset & ~PAGE_MASK));
345 iowrite32(reloc->delta, reloc_entry);
346 io_mapping_unmap_atomic(reloc_page);
349 /* and update the user's relocation entry */
350 reloc->presumed_offset = target_offset;
355 drm_gem_object_unreference(target_obj);
360 i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
361 struct drm_file *file_priv,
362 struct drm_i915_gem_exec_object2 *entry)
364 struct drm_i915_gem_relocation_entry __user *user_relocs;
367 user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr;
368 for (i = 0; i < entry->relocation_count; i++) {
369 struct drm_i915_gem_relocation_entry reloc;
371 if (__copy_from_user_inatomic(&reloc,
376 ret = i915_gem_execbuffer_relocate_entry(obj, file_priv, entry, &reloc);
380 if (__copy_to_user_inatomic(&user_relocs[i].presumed_offset,
381 &reloc.presumed_offset,
382 sizeof(reloc.presumed_offset)))
390 i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
391 struct drm_file *file_priv,
392 struct drm_i915_gem_exec_object2 *entry,
393 struct drm_i915_gem_relocation_entry *relocs)
397 for (i = 0; i < entry->relocation_count; i++) {
398 ret = i915_gem_execbuffer_relocate_entry(obj, file_priv, entry, &relocs[i]);
407 i915_gem_execbuffer_relocate(struct drm_device *dev,
408 struct drm_file *file,
409 struct list_head *objects,
410 struct drm_i915_gem_exec_object2 *exec)
412 struct drm_i915_gem_object *obj;
415 list_for_each_entry(obj, objects, exec_list) {
416 obj->base.pending_read_domains = 0;
417 obj->base.pending_write_domain = 0;
418 ret = i915_gem_execbuffer_relocate_object(obj, file, exec++);
427 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
428 struct drm_file *file,
429 struct list_head *objects,
430 struct drm_i915_gem_exec_object2 *exec)
432 struct drm_i915_gem_object *obj;
433 struct drm_i915_gem_exec_object2 *entry;
435 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
437 /* Attempt to pin all of the buffers into the GTT.
438 * This is done in 3 phases:
440 * 1a. Unbind all objects that do not match the GTT constraints for
441 * the execbuffer (fenceable, mappable, alignment etc).
442 * 1b. Increment pin count for already bound objects.
443 * 2. Bind new objects.
444 * 3. Decrement pin count.
446 * This avoid unnecessary unbinding of later objects in order to makr
447 * room for the earlier objects *unless* we need to defragment.
453 /* Unbind any ill-fitting objects or pin. */
455 list_for_each_entry(obj, objects, exec_list) {
456 bool need_fence, need_mappable;
458 if (!obj->gtt_space) {
464 has_fenced_gpu_access &&
465 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
466 obj->tiling_mode != I915_TILING_NONE;
468 entry->relocation_count ? true : need_fence;
470 if ((entry->alignment && obj->gtt_offset & (entry->alignment - 1)) ||
471 (need_mappable && !obj->map_and_fenceable))
472 ret = i915_gem_object_unbind(obj);
474 ret = i915_gem_object_pin(obj,
483 /* Bind fresh objects */
485 list_for_each_entry(obj, objects, exec_list) {
489 has_fenced_gpu_access &&
490 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
491 obj->tiling_mode != I915_TILING_NONE;
493 if (!obj->gtt_space) {
495 entry->relocation_count ? true : need_fence;
497 ret = i915_gem_object_pin(obj,
504 if (has_fenced_gpu_access) {
506 ret = i915_gem_object_get_fence(obj, ring, 1);
509 } else if (entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
510 obj->tiling_mode == I915_TILING_NONE) {
512 ret = i915_gem_object_put_fence(obj);
516 obj->pending_fenced_gpu_access = need_fence;
519 entry->offset = obj->gtt_offset;
523 /* Decrement pin count for bound objects */
524 list_for_each_entry(obj, objects, exec_list) {
526 i915_gem_object_unpin(obj);
529 if (ret != -ENOSPC || retry > 1)
532 /* First attempt, just clear anything that is purgeable.
533 * Second attempt, clear the entire GTT.
535 ret = i915_gem_evict_everything(ring->dev, retry == 0);
543 obj = list_entry(obj->exec_list.prev,
544 struct drm_i915_gem_object,
546 while (objects != &obj->exec_list) {
548 i915_gem_object_unpin(obj);
550 obj = list_entry(obj->exec_list.prev,
551 struct drm_i915_gem_object,
559 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
560 struct drm_file *file,
561 struct intel_ring_buffer *ring,
562 struct list_head *objects,
563 struct drm_i915_gem_exec_object2 *exec,
566 struct drm_i915_gem_relocation_entry *reloc;
567 struct drm_i915_gem_object *obj;
570 mutex_unlock(&dev->struct_mutex);
573 for (i = 0; i < count; i++)
574 total += exec[i].relocation_count;
576 reloc = drm_malloc_ab(total, sizeof(*reloc));
578 mutex_lock(&dev->struct_mutex);
583 for (i = 0; i < count; i++) {
584 struct drm_i915_gem_relocation_entry __user *user_relocs;
586 user_relocs = (void __user *)(uintptr_t)exec[i].relocs_ptr;
588 if (copy_from_user(reloc+total, user_relocs,
589 exec[i].relocation_count * sizeof(*reloc))) {
591 mutex_lock(&dev->struct_mutex);
595 total += exec[i].relocation_count;
598 ret = i915_mutex_lock_interruptible(dev);
600 mutex_lock(&dev->struct_mutex);
604 ret = i915_gem_execbuffer_reserve(ring, file, objects, exec);
609 list_for_each_entry(obj, objects, exec_list) {
610 obj->base.pending_read_domains = 0;
611 obj->base.pending_write_domain = 0;
612 ret = i915_gem_execbuffer_relocate_object_slow(obj, file,
618 total += exec->relocation_count;
622 /* Leave the user relocations as are, this is the painfully slow path,
623 * and we want to avoid the complication of dropping the lock whilst
624 * having buffers reserved in the aperture and so causing spurious
625 * ENOSPC for random operations.
629 drm_free_large(reloc);
634 i915_gem_execbuffer_flush(struct drm_device *dev,
635 uint32_t invalidate_domains,
636 uint32_t flush_domains,
637 uint32_t flush_rings)
639 drm_i915_private_t *dev_priv = dev->dev_private;
642 if (flush_domains & I915_GEM_DOMAIN_CPU)
643 intel_gtt_chipset_flush();
645 if ((flush_domains | invalidate_domains) & I915_GEM_GPU_DOMAINS) {
646 for (i = 0; i < I915_NUM_RINGS; i++)
647 if (flush_rings & (1 << i))
648 i915_gem_flush_ring(dev, &dev_priv->ring[i],
655 i915_gem_execbuffer_sync_rings(struct drm_i915_gem_object *obj,
656 struct intel_ring_buffer *to)
658 struct intel_ring_buffer *from = obj->ring;
662 if (from == NULL || to == from)
665 if (INTEL_INFO(obj->base.dev)->gen < 6)
666 return i915_gem_object_wait_rendering(obj, true);
668 idx = intel_ring_sync_index(from, to);
670 seqno = obj->last_rendering_seqno;
671 if (seqno <= from->sync_seqno[idx])
674 if (seqno == from->outstanding_lazy_request) {
675 struct drm_i915_gem_request *request;
677 request = kzalloc(sizeof(*request), GFP_KERNEL);
681 ret = i915_add_request(obj->base.dev, NULL, request, from);
687 seqno = request->seqno;
690 from->sync_seqno[idx] = seqno;
691 return intel_ring_sync(to, from, seqno - 1);
695 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
696 struct list_head *objects)
698 struct drm_i915_gem_object *obj;
699 struct change_domains cd;
702 cd.invalidate_domains = 0;
703 cd.flush_domains = 0;
705 list_for_each_entry(obj, objects, exec_list)
706 i915_gem_object_set_to_gpu_domain(obj, ring, &cd);
708 if (cd.invalidate_domains | cd.flush_domains) {
710 DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
712 cd.invalidate_domains,
715 i915_gem_execbuffer_flush(ring->dev,
716 cd.invalidate_domains,
721 list_for_each_entry(obj, objects, exec_list) {
722 ret = i915_gem_execbuffer_sync_rings(obj, ring);
731 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
733 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
737 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
742 for (i = 0; i < count; i++) {
743 char __user *ptr = (char __user *)(uintptr_t)exec[i].relocs_ptr;
744 int length; /* limited by fault_in_pages_readable() */
746 /* First check for malicious input causing overflow */
747 if (exec[i].relocation_count >
748 INT_MAX / sizeof(struct drm_i915_gem_relocation_entry))
751 length = exec[i].relocation_count *
752 sizeof(struct drm_i915_gem_relocation_entry);
753 if (!access_ok(VERIFY_READ, ptr, length))
756 /* we may also need to update the presumed offsets */
757 if (!access_ok(VERIFY_WRITE, ptr, length))
760 if (fault_in_pages_readable(ptr, length))
768 i915_gem_execbuffer_wait_for_flips(struct intel_ring_buffer *ring,
769 struct list_head *objects)
771 struct drm_i915_gem_object *obj;
774 /* Check for any pending flips. As we only maintain a flip queue depth
775 * of 1, we can simply insert a WAIT for the next display flip prior
776 * to executing the batch and avoid stalling the CPU.
779 list_for_each_entry(obj, objects, exec_list) {
780 if (obj->base.write_domain)
781 flips |= atomic_read(&obj->pending_flip);
784 int plane, flip_mask, ret;
786 for (plane = 0; flips >> plane; plane++) {
787 if (((flips >> plane) & 1) == 0)
791 flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
793 flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
795 ret = intel_ring_begin(ring, 2);
799 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | flip_mask);
800 intel_ring_emit(ring, MI_NOOP);
801 intel_ring_advance(ring);
809 i915_gem_execbuffer_move_to_active(struct list_head *objects,
810 struct intel_ring_buffer *ring,
813 struct drm_i915_gem_object *obj;
815 list_for_each_entry(obj, objects, exec_list) {
816 obj->base.read_domains = obj->base.pending_read_domains;
817 obj->base.write_domain = obj->base.pending_write_domain;
818 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
820 i915_gem_object_move_to_active(obj, ring, seqno);
821 if (obj->base.write_domain) {
823 obj->pending_gpu_write = true;
824 list_move_tail(&obj->gpu_write_list,
825 &ring->gpu_write_list);
826 intel_mark_busy(ring->dev, obj);
829 trace_i915_gem_object_change_domain(obj,
830 obj->base.read_domains,
831 obj->base.write_domain);
836 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
837 struct drm_file *file,
838 struct intel_ring_buffer *ring)
840 struct drm_i915_gem_request *request;
844 * Ensure that the commands in the batch buffer are
845 * finished before the interrupt fires.
847 * The sampler always gets flushed on i965 (sigh).
850 if (INTEL_INFO(dev)->gen >= 4)
851 flush_domains |= I915_GEM_DOMAIN_SAMPLER;
853 ring->flush(ring, I915_GEM_DOMAIN_COMMAND, flush_domains);
855 /* Add a breadcrumb for the completion of the batch buffer */
856 request = kzalloc(sizeof(*request), GFP_KERNEL);
857 if (request == NULL || i915_add_request(dev, file, request, ring)) {
858 i915_gem_next_request_seqno(dev, ring);
864 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
865 struct drm_file *file,
866 struct drm_i915_gem_execbuffer2 *args,
867 struct drm_i915_gem_exec_object2 *exec)
869 drm_i915_private_t *dev_priv = dev->dev_private;
870 struct list_head objects;
871 struct drm_i915_gem_object *batch_obj;
872 struct drm_clip_rect *cliprects = NULL;
873 struct intel_ring_buffer *ring;
874 u32 exec_start, exec_len;
878 if (!i915_gem_check_execbuffer(args)) {
879 DRM_ERROR("execbuf with invalid offset/length\n");
883 ret = validate_exec_list(exec, args->buffer_count);
888 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
889 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
891 switch (args->flags & I915_EXEC_RING_MASK) {
892 case I915_EXEC_DEFAULT:
893 case I915_EXEC_RENDER:
894 ring = &dev_priv->ring[RCS];
898 DRM_ERROR("execbuf with invalid ring (BSD)\n");
901 ring = &dev_priv->ring[VCS];
905 DRM_ERROR("execbuf with invalid ring (BLT)\n");
908 ring = &dev_priv->ring[BCS];
911 DRM_ERROR("execbuf with unknown ring: %d\n",
912 (int)(args->flags & I915_EXEC_RING_MASK));
916 if (args->buffer_count < 1) {
917 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
921 if (args->num_cliprects != 0) {
922 if (ring != &dev_priv->ring[RCS]) {
923 DRM_ERROR("clip rectangles are only valid with the render ring\n");
927 cliprects = kmalloc(args->num_cliprects * sizeof(*cliprects),
929 if (cliprects == NULL) {
934 if (copy_from_user(cliprects,
935 (struct drm_clip_rect __user *)(uintptr_t)
937 sizeof(*cliprects)*args->num_cliprects)) {
943 ret = i915_mutex_lock_interruptible(dev);
947 if (dev_priv->mm.suspended) {
948 mutex_unlock(&dev->struct_mutex);
953 /* Look up object handles */
954 INIT_LIST_HEAD(&objects);
955 for (i = 0; i < args->buffer_count; i++) {
956 struct drm_i915_gem_object *obj;
958 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
961 DRM_ERROR("Invalid object handle %d at index %d\n",
963 /* prevent error path from reading uninitialized data */
968 if (!list_empty(&obj->exec_list)) {
969 DRM_ERROR("Object %p [handle %d, index %d] appears more than once in object list\n",
970 obj, exec[i].handle, i);
975 list_add_tail(&obj->exec_list, &objects);
978 /* Move the objects en-masse into the GTT, evicting if necessary. */
979 ret = i915_gem_execbuffer_reserve(ring, file, &objects, exec);
983 /* The objects are in their final locations, apply the relocations. */
984 ret = i915_gem_execbuffer_relocate(dev, file, &objects, exec);
986 if (ret == -EFAULT) {
987 ret = i915_gem_execbuffer_relocate_slow(dev, file, ring,
990 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
996 /* Set the pending read domains for the batch buffer to COMMAND */
997 batch_obj = list_entry(objects.prev,
998 struct drm_i915_gem_object,
1000 if (batch_obj->base.pending_write_domain) {
1001 DRM_ERROR("Attempting to use self-modifying batch buffer\n");
1005 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1007 ret = i915_gem_execbuffer_move_to_gpu(ring, &objects);
1011 ret = i915_gem_execbuffer_wait_for_flips(ring, &objects);
1015 seqno = i915_gem_next_request_seqno(dev, ring);
1016 for (i = 0; i < I915_NUM_RINGS-1; i++) {
1017 if (seqno < ring->sync_seqno[i]) {
1018 /* The GPU can not handle its semaphore value wrapping,
1019 * so every billion or so execbuffers, we need to stall
1020 * the GPU in order to reset the counters.
1022 ret = i915_gpu_idle(dev);
1026 BUG_ON(ring->sync_seqno[i]);
1030 exec_start = batch_obj->gtt_offset + args->batch_start_offset;
1031 exec_len = args->batch_len;
1033 for (i = 0; i < args->num_cliprects; i++) {
1034 ret = i915_emit_box(dev, &cliprects[i],
1035 args->DR1, args->DR4);
1039 ret = ring->dispatch_execbuffer(ring,
1040 exec_start, exec_len);
1045 ret = ring->dispatch_execbuffer(ring, exec_start, exec_len);
1050 i915_gem_execbuffer_move_to_active(&objects, ring, seqno);
1051 i915_gem_execbuffer_retire_commands(dev, file, ring);
1054 while (!list_empty(&objects)) {
1055 struct drm_i915_gem_object *obj;
1057 obj = list_first_entry(&objects,
1058 struct drm_i915_gem_object,
1060 list_del_init(&obj->exec_list);
1061 drm_gem_object_unreference(&obj->base);
1064 mutex_unlock(&dev->struct_mutex);
1072 * Legacy execbuffer just creates an exec2 list from the original exec object
1073 * list array and passes it to the real function.
1076 i915_gem_execbuffer(struct drm_device *dev, void *data,
1077 struct drm_file *file)
1079 struct drm_i915_gem_execbuffer *args = data;
1080 struct drm_i915_gem_execbuffer2 exec2;
1081 struct drm_i915_gem_exec_object *exec_list = NULL;
1082 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1086 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
1087 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
1090 if (args->buffer_count < 1) {
1091 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
1095 /* Copy in the exec list from userland */
1096 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1097 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1098 if (exec_list == NULL || exec2_list == NULL) {
1099 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
1100 args->buffer_count);
1101 drm_free_large(exec_list);
1102 drm_free_large(exec2_list);
1105 ret = copy_from_user(exec_list,
1106 (struct drm_i915_relocation_entry __user *)
1107 (uintptr_t) args->buffers_ptr,
1108 sizeof(*exec_list) * args->buffer_count);
1110 DRM_ERROR("copy %d exec entries failed %d\n",
1111 args->buffer_count, ret);
1112 drm_free_large(exec_list);
1113 drm_free_large(exec2_list);
1117 for (i = 0; i < args->buffer_count; i++) {
1118 exec2_list[i].handle = exec_list[i].handle;
1119 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1120 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1121 exec2_list[i].alignment = exec_list[i].alignment;
1122 exec2_list[i].offset = exec_list[i].offset;
1123 if (INTEL_INFO(dev)->gen < 4)
1124 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1126 exec2_list[i].flags = 0;
1129 exec2.buffers_ptr = args->buffers_ptr;
1130 exec2.buffer_count = args->buffer_count;
1131 exec2.batch_start_offset = args->batch_start_offset;
1132 exec2.batch_len = args->batch_len;
1133 exec2.DR1 = args->DR1;
1134 exec2.DR4 = args->DR4;
1135 exec2.num_cliprects = args->num_cliprects;
1136 exec2.cliprects_ptr = args->cliprects_ptr;
1137 exec2.flags = I915_EXEC_RENDER;
1139 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1141 /* Copy the new buffer offsets back to the user's exec list. */
1142 for (i = 0; i < args->buffer_count; i++)
1143 exec_list[i].offset = exec2_list[i].offset;
1144 /* ... and back out to userspace */
1145 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1146 (uintptr_t) args->buffers_ptr,
1148 sizeof(*exec_list) * args->buffer_count);
1151 DRM_ERROR("failed to copy %d exec entries "
1152 "back to user (%d)\n",
1153 args->buffer_count, ret);
1157 drm_free_large(exec_list);
1158 drm_free_large(exec2_list);
1163 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1164 struct drm_file *file)
1166 struct drm_i915_gem_execbuffer2 *args = data;
1167 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1171 DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
1172 (int) args->buffers_ptr, args->buffer_count, args->batch_len);
1175 if (args->buffer_count < 1) {
1176 DRM_ERROR("execbuf2 with %d buffers\n", args->buffer_count);
1180 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1181 if (exec2_list == NULL) {
1182 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
1183 args->buffer_count);
1186 ret = copy_from_user(exec2_list,
1187 (struct drm_i915_relocation_entry __user *)
1188 (uintptr_t) args->buffers_ptr,
1189 sizeof(*exec2_list) * args->buffer_count);
1191 DRM_ERROR("copy %d exec entries failed %d\n",
1192 args->buffer_count, ret);
1193 drm_free_large(exec2_list);
1197 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1199 /* Copy the new buffer offsets back to the user's exec list. */
1200 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1201 (uintptr_t) args->buffers_ptr,
1203 sizeof(*exec2_list) * args->buffer_count);
1206 DRM_ERROR("failed to copy %d exec entries "
1207 "back to user (%d)\n",
1208 args->buffer_count, ret);
1212 drm_free_large(exec2_list);