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>
37 struct list_head vmas;
40 struct i915_vma *lut[0];
41 struct hlist_head buckets[0];
45 static struct eb_vmas *
46 eb_create(struct drm_i915_gem_execbuffer2 *args, struct i915_address_space *vm)
48 struct eb_vmas *eb = NULL;
50 if (args->flags & I915_EXEC_HANDLE_LUT) {
51 unsigned size = args->buffer_count;
52 size *= sizeof(struct i915_vma *);
53 size += sizeof(struct eb_vmas);
54 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
58 unsigned size = args->buffer_count;
59 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
60 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
61 while (count > 2*size)
63 eb = kzalloc(count*sizeof(struct hlist_head) +
64 sizeof(struct eb_vmas),
71 eb->and = -args->buffer_count;
73 INIT_LIST_HEAD(&eb->vmas);
78 eb_reset(struct eb_vmas *eb)
81 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
85 eb_lookup_vmas(struct eb_vmas *eb,
86 struct drm_i915_gem_exec_object2 *exec,
87 const struct drm_i915_gem_execbuffer2 *args,
88 struct i915_address_space *vm,
89 struct drm_file *file)
91 struct drm_i915_gem_object *obj;
92 struct list_head objects;
95 INIT_LIST_HEAD(&objects);
96 spin_lock(&file->table_lock);
97 /* Grab a reference to the object and release the lock so we can lookup
98 * or create the VMA without using GFP_ATOMIC */
99 for (i = 0; i < args->buffer_count; i++) {
100 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
102 spin_unlock(&file->table_lock);
103 DRM_DEBUG("Invalid object handle %d at index %d\n",
109 if (!list_empty(&obj->obj_exec_link)) {
110 spin_unlock(&file->table_lock);
111 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
112 obj, exec[i].handle, i);
117 drm_gem_object_reference(&obj->base);
118 list_add_tail(&obj->obj_exec_link, &objects);
120 spin_unlock(&file->table_lock);
123 list_for_each_entry(obj, &objects, obj_exec_link) {
124 struct i915_vma *vma;
127 * NOTE: We can leak any vmas created here when something fails
128 * later on. But that's no issue since vma_unbind can deal with
129 * vmas which are not actually bound. And since only
130 * lookup_or_create exists as an interface to get at the vma
131 * from the (obj, vm) we don't run the risk of creating
132 * duplicated vmas for the same vm.
134 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
136 DRM_DEBUG("Failed to lookup VMA\n");
141 list_add_tail(&vma->exec_list, &eb->vmas);
143 vma->exec_entry = &exec[i];
147 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
148 vma->exec_handle = handle;
149 hlist_add_head(&vma->exec_node,
150 &eb->buckets[handle & eb->and]);
157 while (!list_empty(&objects)) {
158 obj = list_first_entry(&objects,
159 struct drm_i915_gem_object,
161 list_del_init(&obj->obj_exec_link);
163 drm_gem_object_unreference(&obj->base);
168 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
171 if (handle >= -eb->and)
173 return eb->lut[handle];
175 struct hlist_head *head;
176 struct hlist_node *node;
178 head = &eb->buckets[handle & eb->and];
179 hlist_for_each(node, head) {
180 struct i915_vma *vma;
182 vma = hlist_entry(node, struct i915_vma, exec_node);
183 if (vma->exec_handle == handle)
190 static void eb_destroy(struct eb_vmas *eb) {
191 while (!list_empty(&eb->vmas)) {
192 struct i915_vma *vma;
194 vma = list_first_entry(&eb->vmas,
197 list_del_init(&vma->exec_list);
198 drm_gem_object_unreference(&vma->obj->base);
203 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
205 return (HAS_LLC(obj->base.dev) ||
206 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
207 !obj->map_and_fenceable ||
208 obj->cache_level != I915_CACHE_NONE);
212 relocate_entry_cpu(struct drm_i915_gem_object *obj,
213 struct drm_i915_gem_relocation_entry *reloc)
215 uint32_t page_offset = offset_in_page(reloc->offset);
219 ret = i915_gem_object_set_to_cpu_domain(obj, true);
223 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
224 reloc->offset >> PAGE_SHIFT));
225 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
226 kunmap_atomic(vaddr);
232 relocate_entry_gtt(struct drm_i915_gem_object *obj,
233 struct drm_i915_gem_relocation_entry *reloc)
235 struct drm_device *dev = obj->base.dev;
236 struct drm_i915_private *dev_priv = dev->dev_private;
237 uint32_t __iomem *reloc_entry;
238 void __iomem *reloc_page;
241 ret = i915_gem_object_set_to_gtt_domain(obj, true);
245 ret = i915_gem_object_put_fence(obj);
249 /* Map the page containing the relocation we're going to perform. */
250 reloc->offset += i915_gem_obj_ggtt_offset(obj);
251 reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
252 reloc->offset & PAGE_MASK);
253 reloc_entry = (uint32_t __iomem *)
254 (reloc_page + offset_in_page(reloc->offset));
255 iowrite32(reloc->delta, reloc_entry);
256 io_mapping_unmap_atomic(reloc_page);
262 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
264 struct drm_i915_gem_relocation_entry *reloc,
265 struct i915_address_space *vm)
267 struct drm_device *dev = obj->base.dev;
268 struct drm_gem_object *target_obj;
269 struct drm_i915_gem_object *target_i915_obj;
270 struct i915_vma *target_vma;
271 uint32_t target_offset;
274 /* we've already hold a reference to all valid objects */
275 target_vma = eb_get_vma(eb, reloc->target_handle);
276 if (unlikely(target_vma == NULL))
278 target_i915_obj = target_vma->obj;
279 target_obj = &target_vma->obj->base;
281 target_offset = i915_gem_obj_ggtt_offset(target_i915_obj);
283 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
284 * pipe_control writes because the gpu doesn't properly redirect them
285 * through the ppgtt for non_secure batchbuffers. */
286 if (unlikely(IS_GEN6(dev) &&
287 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
288 !target_i915_obj->has_global_gtt_mapping)) {
289 i915_gem_gtt_bind_object(target_i915_obj,
290 target_i915_obj->cache_level);
293 /* Validate that the target is in a valid r/w GPU domain */
294 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
295 DRM_DEBUG("reloc with multiple write domains: "
296 "obj %p target %d offset %d "
297 "read %08x write %08x",
298 obj, reloc->target_handle,
301 reloc->write_domain);
304 if (unlikely((reloc->write_domain | reloc->read_domains)
305 & ~I915_GEM_GPU_DOMAINS)) {
306 DRM_DEBUG("reloc with read/write non-GPU domains: "
307 "obj %p target %d offset %d "
308 "read %08x write %08x",
309 obj, reloc->target_handle,
312 reloc->write_domain);
316 target_obj->pending_read_domains |= reloc->read_domains;
317 target_obj->pending_write_domain |= reloc->write_domain;
319 /* If the relocation already has the right value in it, no
320 * more work needs to be done.
322 if (target_offset == reloc->presumed_offset)
325 /* Check that the relocation address is valid... */
326 if (unlikely(reloc->offset > obj->base.size - 4)) {
327 DRM_DEBUG("Relocation beyond object bounds: "
328 "obj %p target %d offset %d size %d.\n",
329 obj, reloc->target_handle,
331 (int) obj->base.size);
334 if (unlikely(reloc->offset & 3)) {
335 DRM_DEBUG("Relocation not 4-byte aligned: "
336 "obj %p target %d offset %d.\n",
337 obj, reloc->target_handle,
338 (int) reloc->offset);
342 /* We can't wait for rendering with pagefaults disabled */
343 if (obj->active && in_atomic())
346 reloc->delta += target_offset;
347 if (use_cpu_reloc(obj))
348 ret = relocate_entry_cpu(obj, reloc);
350 ret = relocate_entry_gtt(obj, reloc);
355 /* and update the user's relocation entry */
356 reloc->presumed_offset = target_offset;
362 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
365 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
366 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
367 struct drm_i915_gem_relocation_entry __user *user_relocs;
368 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
371 user_relocs = to_user_ptr(entry->relocs_ptr);
373 remain = entry->relocation_count;
375 struct drm_i915_gem_relocation_entry *r = stack_reloc;
377 if (count > ARRAY_SIZE(stack_reloc))
378 count = ARRAY_SIZE(stack_reloc);
381 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
385 u64 offset = r->presumed_offset;
387 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r,
392 if (r->presumed_offset != offset &&
393 __copy_to_user_inatomic(&user_relocs->presumed_offset,
395 sizeof(r->presumed_offset))) {
409 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
411 struct drm_i915_gem_relocation_entry *relocs)
413 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
416 for (i = 0; i < entry->relocation_count; i++) {
417 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i],
427 i915_gem_execbuffer_relocate(struct eb_vmas *eb,
428 struct i915_address_space *vm)
430 struct i915_vma *vma;
433 /* This is the fast path and we cannot handle a pagefault whilst
434 * holding the struct mutex lest the user pass in the relocations
435 * contained within a mmaped bo. For in such a case we, the page
436 * fault handler would call i915_gem_fault() and we would try to
437 * acquire the struct mutex again. Obviously this is bad and so
438 * lockdep complains vehemently.
441 list_for_each_entry(vma, &eb->vmas, exec_list) {
442 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
451 #define __EXEC_OBJECT_HAS_PIN (1<<31)
452 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
455 need_reloc_mappable(struct i915_vma *vma)
457 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
458 return entry->relocation_count && !use_cpu_reloc(vma->obj) &&
459 i915_is_ggtt(vma->vm);
463 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
464 struct intel_ring_buffer *ring,
467 struct drm_i915_private *dev_priv = ring->dev->dev_private;
468 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
469 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
470 bool need_fence, need_mappable;
471 struct drm_i915_gem_object *obj = vma->obj;
475 has_fenced_gpu_access &&
476 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
477 obj->tiling_mode != I915_TILING_NONE;
478 need_mappable = need_fence || need_reloc_mappable(vma);
480 ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, need_mappable,
485 entry->flags |= __EXEC_OBJECT_HAS_PIN;
487 if (has_fenced_gpu_access) {
488 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
489 ret = i915_gem_object_get_fence(obj);
493 if (i915_gem_object_pin_fence(obj))
494 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
496 obj->pending_fenced_gpu_access = true;
500 /* Ensure ppgtt mapping exists if needed */
501 if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) {
502 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
503 obj, obj->cache_level);
505 obj->has_aliasing_ppgtt_mapping = 1;
508 if (entry->offset != vma->node.start) {
509 entry->offset = vma->node.start;
513 if (entry->flags & EXEC_OBJECT_WRITE) {
514 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
515 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
518 if (entry->flags & EXEC_OBJECT_NEEDS_GTT &&
519 !obj->has_global_gtt_mapping)
520 i915_gem_gtt_bind_object(obj, obj->cache_level);
526 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
528 struct drm_i915_gem_exec_object2 *entry;
529 struct drm_i915_gem_object *obj = vma->obj;
531 if (!drm_mm_node_allocated(&vma->node))
534 entry = vma->exec_entry;
536 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
537 i915_gem_object_unpin_fence(obj);
539 if (entry->flags & __EXEC_OBJECT_HAS_PIN)
540 i915_gem_object_unpin(obj);
542 entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
546 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
547 struct list_head *vmas,
550 struct drm_i915_gem_object *obj;
551 struct i915_vma *vma;
552 struct i915_address_space *vm;
553 struct list_head ordered_vmas;
554 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
557 if (list_empty(vmas))
560 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
562 INIT_LIST_HEAD(&ordered_vmas);
563 while (!list_empty(vmas)) {
564 struct drm_i915_gem_exec_object2 *entry;
565 bool need_fence, need_mappable;
567 vma = list_first_entry(vmas, struct i915_vma, exec_list);
569 entry = vma->exec_entry;
572 has_fenced_gpu_access &&
573 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
574 obj->tiling_mode != I915_TILING_NONE;
575 need_mappable = need_fence || need_reloc_mappable(vma);
578 list_move(&vma->exec_list, &ordered_vmas);
580 list_move_tail(&vma->exec_list, &ordered_vmas);
582 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
583 obj->base.pending_write_domain = 0;
584 obj->pending_fenced_gpu_access = false;
586 list_splice(&ordered_vmas, vmas);
588 /* Attempt to pin all of the buffers into the GTT.
589 * This is done in 3 phases:
591 * 1a. Unbind all objects that do not match the GTT constraints for
592 * the execbuffer (fenceable, mappable, alignment etc).
593 * 1b. Increment pin count for already bound objects.
594 * 2. Bind new objects.
595 * 3. Decrement pin count.
597 * This avoid unnecessary unbinding of later objects in order to make
598 * room for the earlier objects *unless* we need to defragment.
604 /* Unbind any ill-fitting objects or pin. */
605 list_for_each_entry(vma, vmas, exec_list) {
606 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
607 bool need_fence, need_mappable;
611 if (!drm_mm_node_allocated(&vma->node))
615 has_fenced_gpu_access &&
616 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
617 obj->tiling_mode != I915_TILING_NONE;
618 need_mappable = need_fence || need_reloc_mappable(vma);
620 WARN_ON((need_mappable || need_fence) &&
621 !i915_is_ggtt(vma->vm));
623 if ((entry->alignment &&
624 vma->node.start & (entry->alignment - 1)) ||
625 (need_mappable && !obj->map_and_fenceable))
626 ret = i915_vma_unbind(vma);
628 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
633 /* Bind fresh objects */
634 list_for_each_entry(vma, vmas, exec_list) {
635 if (drm_mm_node_allocated(&vma->node))
638 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
643 err: /* Decrement pin count for bound objects */
644 list_for_each_entry(vma, vmas, exec_list)
645 i915_gem_execbuffer_unreserve_vma(vma);
647 if (ret != -ENOSPC || retry++)
650 ret = i915_gem_evict_vm(vm, true);
657 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
658 struct drm_i915_gem_execbuffer2 *args,
659 struct drm_file *file,
660 struct intel_ring_buffer *ring,
662 struct drm_i915_gem_exec_object2 *exec)
664 struct drm_i915_gem_relocation_entry *reloc;
665 struct i915_address_space *vm;
666 struct i915_vma *vma;
670 unsigned count = args->buffer_count;
672 if (WARN_ON(list_empty(&eb->vmas)))
675 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
677 /* We may process another execbuffer during the unlock... */
678 while (!list_empty(&eb->vmas)) {
679 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
680 list_del_init(&vma->exec_list);
681 drm_gem_object_unreference(&vma->obj->base);
684 mutex_unlock(&dev->struct_mutex);
687 for (i = 0; i < count; i++)
688 total += exec[i].relocation_count;
690 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
691 reloc = drm_malloc_ab(total, sizeof(*reloc));
692 if (reloc == NULL || reloc_offset == NULL) {
693 drm_free_large(reloc);
694 drm_free_large(reloc_offset);
695 mutex_lock(&dev->struct_mutex);
700 for (i = 0; i < count; i++) {
701 struct drm_i915_gem_relocation_entry __user *user_relocs;
702 u64 invalid_offset = (u64)-1;
705 user_relocs = to_user_ptr(exec[i].relocs_ptr);
707 if (copy_from_user(reloc+total, user_relocs,
708 exec[i].relocation_count * sizeof(*reloc))) {
710 mutex_lock(&dev->struct_mutex);
714 /* As we do not update the known relocation offsets after
715 * relocating (due to the complexities in lock handling),
716 * we need to mark them as invalid now so that we force the
717 * relocation processing next time. Just in case the target
718 * object is evicted and then rebound into its old
719 * presumed_offset before the next execbuffer - if that
720 * happened we would make the mistake of assuming that the
721 * relocations were valid.
723 for (j = 0; j < exec[i].relocation_count; j++) {
724 if (copy_to_user(&user_relocs[j].presumed_offset,
726 sizeof(invalid_offset))) {
728 mutex_lock(&dev->struct_mutex);
733 reloc_offset[i] = total;
734 total += exec[i].relocation_count;
737 ret = i915_mutex_lock_interruptible(dev);
739 mutex_lock(&dev->struct_mutex);
743 /* reacquire the objects */
745 ret = eb_lookup_vmas(eb, exec, args, vm, file);
749 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
750 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
754 list_for_each_entry(vma, &eb->vmas, exec_list) {
755 int offset = vma->exec_entry - exec;
756 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
757 reloc + reloc_offset[offset]);
762 /* Leave the user relocations as are, this is the painfully slow path,
763 * and we want to avoid the complication of dropping the lock whilst
764 * having buffers reserved in the aperture and so causing spurious
765 * ENOSPC for random operations.
769 drm_free_large(reloc);
770 drm_free_large(reloc_offset);
775 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
776 struct list_head *vmas)
778 struct i915_vma *vma;
779 uint32_t flush_domains = 0;
780 bool flush_chipset = false;
783 list_for_each_entry(vma, vmas, exec_list) {
784 struct drm_i915_gem_object *obj = vma->obj;
785 ret = i915_gem_object_sync(obj, ring);
789 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
790 flush_chipset |= i915_gem_clflush_object(obj, false);
792 flush_domains |= obj->base.write_domain;
796 i915_gem_chipset_flush(ring->dev);
798 if (flush_domains & I915_GEM_DOMAIN_GTT)
801 /* Unconditionally invalidate gpu caches and ensure that we do flush
802 * any residual writes from the previous batch.
804 return intel_ring_invalidate_all_caches(ring);
808 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
810 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
813 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
817 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
821 unsigned relocs_total = 0;
822 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
824 for (i = 0; i < count; i++) {
825 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
826 int length; /* limited by fault_in_pages_readable() */
828 if (exec[i].flags & __EXEC_OBJECT_UNKNOWN_FLAGS)
831 /* First check for malicious input causing overflow in
832 * the worst case where we need to allocate the entire
833 * relocation tree as a single array.
835 if (exec[i].relocation_count > relocs_max - relocs_total)
837 relocs_total += exec[i].relocation_count;
839 length = exec[i].relocation_count *
840 sizeof(struct drm_i915_gem_relocation_entry);
842 * We must check that the entire relocation array is safe
843 * to read, but since we may need to update the presumed
844 * offsets during execution, check for full write access.
846 if (!access_ok(VERIFY_WRITE, ptr, length))
849 if (likely(!i915_prefault_disable)) {
850 if (fault_in_multipages_readable(ptr, length))
859 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
860 struct intel_ring_buffer *ring)
862 struct i915_vma *vma;
864 list_for_each_entry(vma, vmas, exec_list) {
865 struct drm_i915_gem_object *obj = vma->obj;
866 u32 old_read = obj->base.read_domains;
867 u32 old_write = obj->base.write_domain;
869 obj->base.write_domain = obj->base.pending_write_domain;
870 if (obj->base.write_domain == 0)
871 obj->base.pending_read_domains |= obj->base.read_domains;
872 obj->base.read_domains = obj->base.pending_read_domains;
873 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
875 i915_vma_move_to_active(vma, ring);
876 if (obj->base.write_domain) {
878 obj->last_write_seqno = intel_ring_get_seqno(ring);
879 if (obj->pin_count) /* check for potential scanout */
880 intel_mark_fb_busy(obj, ring);
883 trace_i915_gem_object_change_domain(obj, old_read, old_write);
888 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
889 struct drm_file *file,
890 struct intel_ring_buffer *ring,
891 struct drm_i915_gem_object *obj)
893 /* Unconditionally force add_request to emit a full flush. */
894 ring->gpu_caches_dirty = true;
896 /* Add a breadcrumb for the completion of the batch buffer */
897 (void)__i915_add_request(ring, file, obj, NULL);
901 i915_reset_gen7_sol_offsets(struct drm_device *dev,
902 struct intel_ring_buffer *ring)
904 drm_i915_private_t *dev_priv = dev->dev_private;
907 if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS])
910 ret = intel_ring_begin(ring, 4 * 3);
914 for (i = 0; i < 4; i++) {
915 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
916 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
917 intel_ring_emit(ring, 0);
920 intel_ring_advance(ring);
926 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
927 struct drm_file *file,
928 struct drm_i915_gem_execbuffer2 *args,
929 struct drm_i915_gem_exec_object2 *exec,
930 struct i915_address_space *vm)
932 drm_i915_private_t *dev_priv = dev->dev_private;
934 struct drm_i915_gem_object *batch_obj;
935 struct drm_clip_rect *cliprects = NULL;
936 struct intel_ring_buffer *ring;
937 struct i915_ctx_hang_stats *hs;
938 u32 ctx_id = i915_execbuffer2_get_context_id(*args);
939 u32 exec_start, exec_len;
944 if (!i915_gem_check_execbuffer(args))
947 ret = validate_exec_list(exec, args->buffer_count);
952 if (args->flags & I915_EXEC_SECURE) {
953 if (!file->is_master || !capable(CAP_SYS_ADMIN))
956 flags |= I915_DISPATCH_SECURE;
958 if (args->flags & I915_EXEC_IS_PINNED)
959 flags |= I915_DISPATCH_PINNED;
961 switch (args->flags & I915_EXEC_RING_MASK) {
962 case I915_EXEC_DEFAULT:
963 case I915_EXEC_RENDER:
964 ring = &dev_priv->ring[RCS];
967 ring = &dev_priv->ring[VCS];
968 if (ctx_id != DEFAULT_CONTEXT_ID) {
969 DRM_DEBUG("Ring %s doesn't support contexts\n",
975 ring = &dev_priv->ring[BCS];
976 if (ctx_id != DEFAULT_CONTEXT_ID) {
977 DRM_DEBUG("Ring %s doesn't support contexts\n",
982 case I915_EXEC_VEBOX:
983 ring = &dev_priv->ring[VECS];
984 if (ctx_id != DEFAULT_CONTEXT_ID) {
985 DRM_DEBUG("Ring %s doesn't support contexts\n",
992 DRM_DEBUG("execbuf with unknown ring: %d\n",
993 (int)(args->flags & I915_EXEC_RING_MASK));
996 if (!intel_ring_initialized(ring)) {
997 DRM_DEBUG("execbuf with invalid ring: %d\n",
998 (int)(args->flags & I915_EXEC_RING_MASK));
1002 mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1003 mask = I915_EXEC_CONSTANTS_MASK;
1005 case I915_EXEC_CONSTANTS_REL_GENERAL:
1006 case I915_EXEC_CONSTANTS_ABSOLUTE:
1007 case I915_EXEC_CONSTANTS_REL_SURFACE:
1008 if (ring == &dev_priv->ring[RCS] &&
1009 mode != dev_priv->relative_constants_mode) {
1010 if (INTEL_INFO(dev)->gen < 4)
1013 if (INTEL_INFO(dev)->gen > 5 &&
1014 mode == I915_EXEC_CONSTANTS_REL_SURFACE)
1017 /* The HW changed the meaning on this bit on gen6 */
1018 if (INTEL_INFO(dev)->gen >= 6)
1019 mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1023 DRM_DEBUG("execbuf with unknown constants: %d\n", mode);
1027 if (args->buffer_count < 1) {
1028 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1032 if (args->num_cliprects != 0) {
1033 if (ring != &dev_priv->ring[RCS]) {
1034 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1038 if (INTEL_INFO(dev)->gen >= 5) {
1039 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1043 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1044 DRM_DEBUG("execbuf with %u cliprects\n",
1045 args->num_cliprects);
1049 cliprects = kcalloc(args->num_cliprects,
1052 if (cliprects == NULL) {
1057 if (copy_from_user(cliprects,
1058 to_user_ptr(args->cliprects_ptr),
1059 sizeof(*cliprects)*args->num_cliprects)) {
1065 ret = i915_mutex_lock_interruptible(dev);
1069 if (dev_priv->ums.mm_suspended) {
1070 mutex_unlock(&dev->struct_mutex);
1075 eb = eb_create(args, vm);
1077 mutex_unlock(&dev->struct_mutex);
1082 /* Look up object handles */
1083 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1087 /* take note of the batch buffer before we might reorder the lists */
1088 batch_obj = list_entry(eb->vmas.prev, struct i915_vma, exec_list)->obj;
1090 /* Move the objects en-masse into the GTT, evicting if necessary. */
1091 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1092 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
1096 /* The objects are in their final locations, apply the relocations. */
1098 ret = i915_gem_execbuffer_relocate(eb, vm);
1100 if (ret == -EFAULT) {
1101 ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1103 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1109 /* Set the pending read domains for the batch buffer to COMMAND */
1110 if (batch_obj->base.pending_write_domain) {
1111 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1115 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1117 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1118 * batch" bit. Hence we need to pin secure batches into the global gtt.
1119 * hsw should have this fixed, but let's be paranoid and do it
1120 * unconditionally for now. */
1121 if (flags & I915_DISPATCH_SECURE && !batch_obj->has_global_gtt_mapping)
1122 i915_gem_gtt_bind_object(batch_obj, batch_obj->cache_level);
1124 ret = i915_gem_execbuffer_move_to_gpu(ring, &eb->vmas);
1128 hs = i915_gem_context_get_hang_stats(dev, file, ctx_id);
1139 ret = i915_switch_context(ring, file, ctx_id);
1143 if (ring == &dev_priv->ring[RCS] &&
1144 mode != dev_priv->relative_constants_mode) {
1145 ret = intel_ring_begin(ring, 4);
1149 intel_ring_emit(ring, MI_NOOP);
1150 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1151 intel_ring_emit(ring, INSTPM);
1152 intel_ring_emit(ring, mask << 16 | mode);
1153 intel_ring_advance(ring);
1155 dev_priv->relative_constants_mode = mode;
1158 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1159 ret = i915_reset_gen7_sol_offsets(dev, ring);
1164 exec_start = i915_gem_obj_offset(batch_obj, vm) +
1165 args->batch_start_offset;
1166 exec_len = args->batch_len;
1168 for (i = 0; i < args->num_cliprects; i++) {
1169 ret = i915_emit_box(dev, &cliprects[i],
1170 args->DR1, args->DR4);
1174 ret = ring->dispatch_execbuffer(ring,
1175 exec_start, exec_len,
1181 ret = ring->dispatch_execbuffer(ring,
1182 exec_start, exec_len,
1188 trace_i915_gem_ring_dispatch(ring, intel_ring_get_seqno(ring), flags);
1190 i915_gem_execbuffer_move_to_active(&eb->vmas, ring);
1191 i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
1196 mutex_unlock(&dev->struct_mutex);
1204 * Legacy execbuffer just creates an exec2 list from the original exec object
1205 * list array and passes it to the real function.
1208 i915_gem_execbuffer(struct drm_device *dev, void *data,
1209 struct drm_file *file)
1211 struct drm_i915_private *dev_priv = dev->dev_private;
1212 struct drm_i915_gem_execbuffer *args = data;
1213 struct drm_i915_gem_execbuffer2 exec2;
1214 struct drm_i915_gem_exec_object *exec_list = NULL;
1215 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1218 if (args->buffer_count < 1) {
1219 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1223 /* Copy in the exec list from userland */
1224 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1225 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1226 if (exec_list == NULL || exec2_list == NULL) {
1227 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1228 args->buffer_count);
1229 drm_free_large(exec_list);
1230 drm_free_large(exec2_list);
1233 ret = copy_from_user(exec_list,
1234 to_user_ptr(args->buffers_ptr),
1235 sizeof(*exec_list) * args->buffer_count);
1237 DRM_DEBUG("copy %d exec entries failed %d\n",
1238 args->buffer_count, ret);
1239 drm_free_large(exec_list);
1240 drm_free_large(exec2_list);
1244 for (i = 0; i < args->buffer_count; i++) {
1245 exec2_list[i].handle = exec_list[i].handle;
1246 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1247 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1248 exec2_list[i].alignment = exec_list[i].alignment;
1249 exec2_list[i].offset = exec_list[i].offset;
1250 if (INTEL_INFO(dev)->gen < 4)
1251 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1253 exec2_list[i].flags = 0;
1256 exec2.buffers_ptr = args->buffers_ptr;
1257 exec2.buffer_count = args->buffer_count;
1258 exec2.batch_start_offset = args->batch_start_offset;
1259 exec2.batch_len = args->batch_len;
1260 exec2.DR1 = args->DR1;
1261 exec2.DR4 = args->DR4;
1262 exec2.num_cliprects = args->num_cliprects;
1263 exec2.cliprects_ptr = args->cliprects_ptr;
1264 exec2.flags = I915_EXEC_RENDER;
1265 i915_execbuffer2_set_context_id(exec2, 0);
1267 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list,
1268 &dev_priv->gtt.base);
1270 /* Copy the new buffer offsets back to the user's exec list. */
1271 for (i = 0; i < args->buffer_count; i++)
1272 exec_list[i].offset = exec2_list[i].offset;
1273 /* ... and back out to userspace */
1274 ret = copy_to_user(to_user_ptr(args->buffers_ptr),
1276 sizeof(*exec_list) * args->buffer_count);
1279 DRM_DEBUG("failed to copy %d exec entries "
1280 "back to user (%d)\n",
1281 args->buffer_count, ret);
1285 drm_free_large(exec_list);
1286 drm_free_large(exec2_list);
1291 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1292 struct drm_file *file)
1294 struct drm_i915_private *dev_priv = dev->dev_private;
1295 struct drm_i915_gem_execbuffer2 *args = data;
1296 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1299 if (args->buffer_count < 1 ||
1300 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1301 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1305 exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1306 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
1307 if (exec2_list == NULL)
1308 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1309 args->buffer_count);
1310 if (exec2_list == NULL) {
1311 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1312 args->buffer_count);
1315 ret = copy_from_user(exec2_list,
1316 to_user_ptr(args->buffers_ptr),
1317 sizeof(*exec2_list) * args->buffer_count);
1319 DRM_DEBUG("copy %d exec entries failed %d\n",
1320 args->buffer_count, ret);
1321 drm_free_large(exec2_list);
1325 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list,
1326 &dev_priv->gtt.base);
1328 /* Copy the new buffer offsets back to the user's exec list. */
1329 ret = copy_to_user(to_user_ptr(args->buffers_ptr),
1331 sizeof(*exec2_list) * args->buffer_count);
1334 DRM_DEBUG("failed to copy %d exec entries "
1335 "back to user (%d)\n",
1336 args->buffer_count, ret);
1340 drm_free_large(exec2_list);