2 * Copyright © 2008,2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
25 * Chris Wilson <chris@chris-wilson.co.uk>
30 #include <drm/i915_drm.h>
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/dma_remapping.h>
36 #define __EXEC_OBJECT_HAS_PIN (1<<31)
37 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
40 struct list_head vmas;
43 struct i915_vma *lut[0];
44 struct hlist_head buckets[0];
48 static struct eb_vmas *
49 eb_create(struct drm_i915_gem_execbuffer2 *args, struct i915_address_space *vm)
51 struct eb_vmas *eb = NULL;
53 if (args->flags & I915_EXEC_HANDLE_LUT) {
54 unsigned size = args->buffer_count;
55 size *= sizeof(struct i915_vma *);
56 size += sizeof(struct eb_vmas);
57 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
61 unsigned size = args->buffer_count;
62 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
63 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
64 while (count > 2*size)
66 eb = kzalloc(count*sizeof(struct hlist_head) +
67 sizeof(struct eb_vmas),
74 eb->and = -args->buffer_count;
76 INIT_LIST_HEAD(&eb->vmas);
81 eb_reset(struct eb_vmas *eb)
84 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
88 eb_lookup_vmas(struct eb_vmas *eb,
89 struct drm_i915_gem_exec_object2 *exec,
90 const struct drm_i915_gem_execbuffer2 *args,
91 struct i915_address_space *vm,
92 struct drm_file *file)
94 struct drm_i915_gem_object *obj;
95 struct list_head objects;
98 INIT_LIST_HEAD(&objects);
99 spin_lock(&file->table_lock);
100 /* Grab a reference to the object and release the lock so we can lookup
101 * or create the VMA without using GFP_ATOMIC */
102 for (i = 0; i < args->buffer_count; i++) {
103 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
105 spin_unlock(&file->table_lock);
106 DRM_DEBUG("Invalid object handle %d at index %d\n",
112 if (!list_empty(&obj->obj_exec_link)) {
113 spin_unlock(&file->table_lock);
114 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
115 obj, exec[i].handle, i);
120 drm_gem_object_reference(&obj->base);
121 list_add_tail(&obj->obj_exec_link, &objects);
123 spin_unlock(&file->table_lock);
126 while (!list_empty(&objects)) {
127 struct i915_vma *vma;
129 obj = list_first_entry(&objects,
130 struct drm_i915_gem_object,
134 * NOTE: We can leak any vmas created here when something fails
135 * later on. But that's no issue since vma_unbind can deal with
136 * vmas which are not actually bound. And since only
137 * lookup_or_create exists as an interface to get at the vma
138 * from the (obj, vm) we don't run the risk of creating
139 * duplicated vmas for the same vm.
141 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
143 DRM_DEBUG("Failed to lookup VMA\n");
148 /* Transfer ownership from the objects list to the vmas list. */
149 list_add_tail(&vma->exec_list, &eb->vmas);
150 list_del_init(&obj->obj_exec_link);
152 vma->exec_entry = &exec[i];
156 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
157 vma->exec_handle = handle;
158 hlist_add_head(&vma->exec_node,
159 &eb->buckets[handle & eb->and]);
168 while (!list_empty(&objects)) {
169 obj = list_first_entry(&objects,
170 struct drm_i915_gem_object,
172 list_del_init(&obj->obj_exec_link);
173 drm_gem_object_unreference(&obj->base);
176 * Objects already transfered to the vmas list will be unreferenced by
183 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
186 if (handle >= -eb->and)
188 return eb->lut[handle];
190 struct hlist_head *head;
191 struct hlist_node *node;
193 head = &eb->buckets[handle & eb->and];
194 hlist_for_each(node, head) {
195 struct i915_vma *vma;
197 vma = hlist_entry(node, struct i915_vma, exec_node);
198 if (vma->exec_handle == handle)
206 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
208 struct drm_i915_gem_exec_object2 *entry;
209 struct drm_i915_gem_object *obj = vma->obj;
211 if (!drm_mm_node_allocated(&vma->node))
214 entry = vma->exec_entry;
216 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
217 i915_gem_object_unpin_fence(obj);
219 if (entry->flags & __EXEC_OBJECT_HAS_PIN)
220 i915_gem_object_unpin(obj);
222 entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
225 static void eb_destroy(struct eb_vmas *eb)
227 while (!list_empty(&eb->vmas)) {
228 struct i915_vma *vma;
230 vma = list_first_entry(&eb->vmas,
233 list_del_init(&vma->exec_list);
234 i915_gem_execbuffer_unreserve_vma(vma);
235 drm_gem_object_unreference(&vma->obj->base);
240 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
242 return (HAS_LLC(obj->base.dev) ||
243 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
244 !obj->map_and_fenceable ||
245 obj->cache_level != I915_CACHE_NONE);
249 relocate_entry_cpu(struct drm_i915_gem_object *obj,
250 struct drm_i915_gem_relocation_entry *reloc)
252 struct drm_device *dev = obj->base.dev;
253 uint32_t page_offset = offset_in_page(reloc->offset);
257 ret = i915_gem_object_set_to_cpu_domain(obj, true);
261 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
262 reloc->offset >> PAGE_SHIFT));
263 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
265 if (INTEL_INFO(dev)->gen >= 8) {
266 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
268 if (page_offset == 0) {
269 kunmap_atomic(vaddr);
270 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
271 (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
274 *(uint32_t *)(vaddr + page_offset) = 0;
277 kunmap_atomic(vaddr);
283 relocate_entry_gtt(struct drm_i915_gem_object *obj,
284 struct drm_i915_gem_relocation_entry *reloc)
286 struct drm_device *dev = obj->base.dev;
287 struct drm_i915_private *dev_priv = dev->dev_private;
288 uint32_t __iomem *reloc_entry;
289 void __iomem *reloc_page;
292 ret = i915_gem_object_set_to_gtt_domain(obj, true);
296 ret = i915_gem_object_put_fence(obj);
300 /* Map the page containing the relocation we're going to perform. */
301 reloc->offset += i915_gem_obj_ggtt_offset(obj);
302 reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
303 reloc->offset & PAGE_MASK);
304 reloc_entry = (uint32_t __iomem *)
305 (reloc_page + offset_in_page(reloc->offset));
306 iowrite32(reloc->delta, reloc_entry);
308 if (INTEL_INFO(dev)->gen >= 8) {
311 if (offset_in_page(reloc->offset + sizeof(uint32_t)) == 0) {
312 io_mapping_unmap_atomic(reloc_page);
313 reloc_page = io_mapping_map_atomic_wc(
314 dev_priv->gtt.mappable,
315 reloc->offset + sizeof(uint32_t));
316 reloc_entry = reloc_page;
319 iowrite32(0, reloc_entry);
322 io_mapping_unmap_atomic(reloc_page);
328 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
330 struct drm_i915_gem_relocation_entry *reloc,
331 struct i915_address_space *vm)
333 struct drm_device *dev = obj->base.dev;
334 struct drm_gem_object *target_obj;
335 struct drm_i915_gem_object *target_i915_obj;
336 struct i915_vma *target_vma;
337 uint32_t target_offset;
340 /* we've already hold a reference to all valid objects */
341 target_vma = eb_get_vma(eb, reloc->target_handle);
342 if (unlikely(target_vma == NULL))
344 target_i915_obj = target_vma->obj;
345 target_obj = &target_vma->obj->base;
347 target_offset = i915_gem_obj_ggtt_offset(target_i915_obj);
349 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
350 * pipe_control writes because the gpu doesn't properly redirect them
351 * through the ppgtt for non_secure batchbuffers. */
352 if (unlikely(IS_GEN6(dev) &&
353 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
354 !target_i915_obj->has_global_gtt_mapping)) {
355 i915_gem_gtt_bind_object(target_i915_obj,
356 target_i915_obj->cache_level);
359 /* Validate that the target is in a valid r/w GPU domain */
360 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
361 DRM_DEBUG("reloc with multiple write domains: "
362 "obj %p target %d offset %d "
363 "read %08x write %08x",
364 obj, reloc->target_handle,
367 reloc->write_domain);
370 if (unlikely((reloc->write_domain | reloc->read_domains)
371 & ~I915_GEM_GPU_DOMAINS)) {
372 DRM_DEBUG("reloc with read/write non-GPU domains: "
373 "obj %p target %d offset %d "
374 "read %08x write %08x",
375 obj, reloc->target_handle,
378 reloc->write_domain);
382 target_obj->pending_read_domains |= reloc->read_domains;
383 target_obj->pending_write_domain |= reloc->write_domain;
385 /* If the relocation already has the right value in it, no
386 * more work needs to be done.
388 if (target_offset == reloc->presumed_offset)
391 /* Check that the relocation address is valid... */
392 if (unlikely(reloc->offset >
393 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
394 DRM_DEBUG("Relocation beyond object bounds: "
395 "obj %p target %d offset %d size %d.\n",
396 obj, reloc->target_handle,
398 (int) obj->base.size);
401 if (unlikely(reloc->offset & 3)) {
402 DRM_DEBUG("Relocation not 4-byte aligned: "
403 "obj %p target %d offset %d.\n",
404 obj, reloc->target_handle,
405 (int) reloc->offset);
409 /* We can't wait for rendering with pagefaults disabled */
410 if (obj->active && in_atomic())
413 reloc->delta += target_offset;
414 if (use_cpu_reloc(obj))
415 ret = relocate_entry_cpu(obj, reloc);
417 ret = relocate_entry_gtt(obj, reloc);
422 /* and update the user's relocation entry */
423 reloc->presumed_offset = target_offset;
429 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
432 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
433 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
434 struct drm_i915_gem_relocation_entry __user *user_relocs;
435 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
438 user_relocs = to_user_ptr(entry->relocs_ptr);
440 remain = entry->relocation_count;
442 struct drm_i915_gem_relocation_entry *r = stack_reloc;
444 if (count > ARRAY_SIZE(stack_reloc))
445 count = ARRAY_SIZE(stack_reloc);
448 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
452 u64 offset = r->presumed_offset;
454 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r,
459 if (r->presumed_offset != offset &&
460 __copy_to_user_inatomic(&user_relocs->presumed_offset,
462 sizeof(r->presumed_offset))) {
476 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
478 struct drm_i915_gem_relocation_entry *relocs)
480 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
483 for (i = 0; i < entry->relocation_count; i++) {
484 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i],
494 i915_gem_execbuffer_relocate(struct eb_vmas *eb,
495 struct i915_address_space *vm)
497 struct i915_vma *vma;
500 /* This is the fast path and we cannot handle a pagefault whilst
501 * holding the struct mutex lest the user pass in the relocations
502 * contained within a mmaped bo. For in such a case we, the page
503 * fault handler would call i915_gem_fault() and we would try to
504 * acquire the struct mutex again. Obviously this is bad and so
505 * lockdep complains vehemently.
508 list_for_each_entry(vma, &eb->vmas, exec_list) {
509 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
519 need_reloc_mappable(struct i915_vma *vma)
521 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
522 return entry->relocation_count && !use_cpu_reloc(vma->obj) &&
523 i915_is_ggtt(vma->vm);
527 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
528 struct intel_ring_buffer *ring,
531 struct drm_i915_private *dev_priv = ring->dev->dev_private;
532 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
533 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
534 bool need_fence, need_mappable;
535 struct drm_i915_gem_object *obj = vma->obj;
539 has_fenced_gpu_access &&
540 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
541 obj->tiling_mode != I915_TILING_NONE;
542 need_mappable = need_fence || need_reloc_mappable(vma);
544 ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, need_mappable,
549 entry->flags |= __EXEC_OBJECT_HAS_PIN;
551 if (has_fenced_gpu_access) {
552 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
553 ret = i915_gem_object_get_fence(obj);
557 if (i915_gem_object_pin_fence(obj))
558 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
560 obj->pending_fenced_gpu_access = true;
564 /* Ensure ppgtt mapping exists if needed */
565 if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) {
566 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
567 obj, obj->cache_level);
569 obj->has_aliasing_ppgtt_mapping = 1;
572 if (entry->offset != vma->node.start) {
573 entry->offset = vma->node.start;
577 if (entry->flags & EXEC_OBJECT_WRITE) {
578 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
579 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
582 if (entry->flags & EXEC_OBJECT_NEEDS_GTT &&
583 !obj->has_global_gtt_mapping)
584 i915_gem_gtt_bind_object(obj, obj->cache_level);
590 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
591 struct list_head *vmas,
594 struct drm_i915_gem_object *obj;
595 struct i915_vma *vma;
596 struct i915_address_space *vm;
597 struct list_head ordered_vmas;
598 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
601 if (list_empty(vmas))
604 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
606 INIT_LIST_HEAD(&ordered_vmas);
607 while (!list_empty(vmas)) {
608 struct drm_i915_gem_exec_object2 *entry;
609 bool need_fence, need_mappable;
611 vma = list_first_entry(vmas, struct i915_vma, exec_list);
613 entry = vma->exec_entry;
616 has_fenced_gpu_access &&
617 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
618 obj->tiling_mode != I915_TILING_NONE;
619 need_mappable = need_fence || need_reloc_mappable(vma);
622 list_move(&vma->exec_list, &ordered_vmas);
624 list_move_tail(&vma->exec_list, &ordered_vmas);
626 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
627 obj->base.pending_write_domain = 0;
628 obj->pending_fenced_gpu_access = false;
630 list_splice(&ordered_vmas, vmas);
632 /* Attempt to pin all of the buffers into the GTT.
633 * This is done in 3 phases:
635 * 1a. Unbind all objects that do not match the GTT constraints for
636 * the execbuffer (fenceable, mappable, alignment etc).
637 * 1b. Increment pin count for already bound objects.
638 * 2. Bind new objects.
639 * 3. Decrement pin count.
641 * This avoid unnecessary unbinding of later objects in order to make
642 * room for the earlier objects *unless* we need to defragment.
648 /* Unbind any ill-fitting objects or pin. */
649 list_for_each_entry(vma, vmas, exec_list) {
650 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
651 bool need_fence, need_mappable;
655 if (!drm_mm_node_allocated(&vma->node))
659 has_fenced_gpu_access &&
660 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
661 obj->tiling_mode != I915_TILING_NONE;
662 need_mappable = need_fence || need_reloc_mappable(vma);
664 WARN_ON((need_mappable || need_fence) &&
665 !i915_is_ggtt(vma->vm));
667 if ((entry->alignment &&
668 vma->node.start & (entry->alignment - 1)) ||
669 (need_mappable && !obj->map_and_fenceable))
670 ret = i915_vma_unbind(vma);
672 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
677 /* Bind fresh objects */
678 list_for_each_entry(vma, vmas, exec_list) {
679 if (drm_mm_node_allocated(&vma->node))
682 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
688 if (ret != -ENOSPC || retry++)
691 /* Decrement pin count for bound objects */
692 list_for_each_entry(vma, vmas, exec_list)
693 i915_gem_execbuffer_unreserve_vma(vma);
695 ret = i915_gem_evict_vm(vm, true);
702 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
703 struct drm_i915_gem_execbuffer2 *args,
704 struct drm_file *file,
705 struct intel_ring_buffer *ring,
707 struct drm_i915_gem_exec_object2 *exec)
709 struct drm_i915_gem_relocation_entry *reloc;
710 struct i915_address_space *vm;
711 struct i915_vma *vma;
715 unsigned count = args->buffer_count;
717 if (WARN_ON(list_empty(&eb->vmas)))
720 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
722 /* We may process another execbuffer during the unlock... */
723 while (!list_empty(&eb->vmas)) {
724 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
725 list_del_init(&vma->exec_list);
726 i915_gem_execbuffer_unreserve_vma(vma);
727 drm_gem_object_unreference(&vma->obj->base);
730 mutex_unlock(&dev->struct_mutex);
733 for (i = 0; i < count; i++)
734 total += exec[i].relocation_count;
736 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
737 reloc = drm_malloc_ab(total, sizeof(*reloc));
738 if (reloc == NULL || reloc_offset == NULL) {
739 drm_free_large(reloc);
740 drm_free_large(reloc_offset);
741 mutex_lock(&dev->struct_mutex);
746 for (i = 0; i < count; i++) {
747 struct drm_i915_gem_relocation_entry __user *user_relocs;
748 u64 invalid_offset = (u64)-1;
751 user_relocs = to_user_ptr(exec[i].relocs_ptr);
753 if (copy_from_user(reloc+total, user_relocs,
754 exec[i].relocation_count * sizeof(*reloc))) {
756 mutex_lock(&dev->struct_mutex);
760 /* As we do not update the known relocation offsets after
761 * relocating (due to the complexities in lock handling),
762 * we need to mark them as invalid now so that we force the
763 * relocation processing next time. Just in case the target
764 * object is evicted and then rebound into its old
765 * presumed_offset before the next execbuffer - if that
766 * happened we would make the mistake of assuming that the
767 * relocations were valid.
769 for (j = 0; j < exec[i].relocation_count; j++) {
770 if (copy_to_user(&user_relocs[j].presumed_offset,
772 sizeof(invalid_offset))) {
774 mutex_lock(&dev->struct_mutex);
779 reloc_offset[i] = total;
780 total += exec[i].relocation_count;
783 ret = i915_mutex_lock_interruptible(dev);
785 mutex_lock(&dev->struct_mutex);
789 /* reacquire the objects */
791 ret = eb_lookup_vmas(eb, exec, args, vm, file);
795 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
796 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
800 list_for_each_entry(vma, &eb->vmas, exec_list) {
801 int offset = vma->exec_entry - exec;
802 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
803 reloc + reloc_offset[offset]);
808 /* Leave the user relocations as are, this is the painfully slow path,
809 * and we want to avoid the complication of dropping the lock whilst
810 * having buffers reserved in the aperture and so causing spurious
811 * ENOSPC for random operations.
815 drm_free_large(reloc);
816 drm_free_large(reloc_offset);
821 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
822 struct list_head *vmas)
824 struct i915_vma *vma;
825 uint32_t flush_domains = 0;
826 bool flush_chipset = false;
829 list_for_each_entry(vma, vmas, exec_list) {
830 struct drm_i915_gem_object *obj = vma->obj;
831 ret = i915_gem_object_sync(obj, ring);
835 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
836 flush_chipset |= i915_gem_clflush_object(obj, false);
838 flush_domains |= obj->base.write_domain;
842 i915_gem_chipset_flush(ring->dev);
844 if (flush_domains & I915_GEM_DOMAIN_GTT)
847 /* Unconditionally invalidate gpu caches and ensure that we do flush
848 * any residual writes from the previous batch.
850 return intel_ring_invalidate_all_caches(ring);
854 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
856 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
859 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
863 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
867 unsigned relocs_total = 0;
868 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
870 for (i = 0; i < count; i++) {
871 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
872 int length; /* limited by fault_in_pages_readable() */
874 if (exec[i].flags & __EXEC_OBJECT_UNKNOWN_FLAGS)
877 /* First check for malicious input causing overflow in
878 * the worst case where we need to allocate the entire
879 * relocation tree as a single array.
881 if (exec[i].relocation_count > relocs_max - relocs_total)
883 relocs_total += exec[i].relocation_count;
885 length = exec[i].relocation_count *
886 sizeof(struct drm_i915_gem_relocation_entry);
888 * We must check that the entire relocation array is safe
889 * to read, but since we may need to update the presumed
890 * offsets during execution, check for full write access.
892 if (!access_ok(VERIFY_WRITE, ptr, length))
895 if (likely(!i915_prefault_disable)) {
896 if (fault_in_multipages_readable(ptr, length))
905 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
906 struct intel_ring_buffer *ring)
908 struct i915_vma *vma;
910 list_for_each_entry(vma, vmas, exec_list) {
911 struct drm_i915_gem_object *obj = vma->obj;
912 u32 old_read = obj->base.read_domains;
913 u32 old_write = obj->base.write_domain;
915 obj->base.write_domain = obj->base.pending_write_domain;
916 if (obj->base.write_domain == 0)
917 obj->base.pending_read_domains |= obj->base.read_domains;
918 obj->base.read_domains = obj->base.pending_read_domains;
919 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
921 i915_vma_move_to_active(vma, ring);
922 if (obj->base.write_domain) {
924 obj->last_write_seqno = intel_ring_get_seqno(ring);
925 if (obj->pin_count) /* check for potential scanout */
926 intel_mark_fb_busy(obj, ring);
929 trace_i915_gem_object_change_domain(obj, old_read, old_write);
934 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
935 struct drm_file *file,
936 struct intel_ring_buffer *ring,
937 struct drm_i915_gem_object *obj)
939 /* Unconditionally force add_request to emit a full flush. */
940 ring->gpu_caches_dirty = true;
942 /* Add a breadcrumb for the completion of the batch buffer */
943 (void)__i915_add_request(ring, file, obj, NULL);
947 i915_reset_gen7_sol_offsets(struct drm_device *dev,
948 struct intel_ring_buffer *ring)
950 drm_i915_private_t *dev_priv = dev->dev_private;
953 if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS])
956 ret = intel_ring_begin(ring, 4 * 3);
960 for (i = 0; i < 4; i++) {
961 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
962 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
963 intel_ring_emit(ring, 0);
966 intel_ring_advance(ring);
972 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
973 struct drm_file *file,
974 struct drm_i915_gem_execbuffer2 *args,
975 struct drm_i915_gem_exec_object2 *exec,
976 struct i915_address_space *vm)
978 drm_i915_private_t *dev_priv = dev->dev_private;
980 struct drm_i915_gem_object *batch_obj;
981 struct drm_clip_rect *cliprects = NULL;
982 struct intel_ring_buffer *ring;
983 struct i915_ctx_hang_stats *hs;
984 u32 ctx_id = i915_execbuffer2_get_context_id(*args);
985 u32 exec_start, exec_len;
990 if (!i915_gem_check_execbuffer(args))
993 ret = validate_exec_list(exec, args->buffer_count);
998 if (args->flags & I915_EXEC_SECURE) {
999 if (!file->is_master || !capable(CAP_SYS_ADMIN))
1002 flags |= I915_DISPATCH_SECURE;
1004 if (args->flags & I915_EXEC_IS_PINNED)
1005 flags |= I915_DISPATCH_PINNED;
1007 switch (args->flags & I915_EXEC_RING_MASK) {
1008 case I915_EXEC_DEFAULT:
1009 case I915_EXEC_RENDER:
1010 ring = &dev_priv->ring[RCS];
1013 ring = &dev_priv->ring[VCS];
1014 if (ctx_id != DEFAULT_CONTEXT_ID) {
1015 DRM_DEBUG("Ring %s doesn't support contexts\n",
1021 ring = &dev_priv->ring[BCS];
1022 if (ctx_id != DEFAULT_CONTEXT_ID) {
1023 DRM_DEBUG("Ring %s doesn't support contexts\n",
1028 case I915_EXEC_VEBOX:
1029 ring = &dev_priv->ring[VECS];
1030 if (ctx_id != DEFAULT_CONTEXT_ID) {
1031 DRM_DEBUG("Ring %s doesn't support contexts\n",
1038 DRM_DEBUG("execbuf with unknown ring: %d\n",
1039 (int)(args->flags & I915_EXEC_RING_MASK));
1042 if (!intel_ring_initialized(ring)) {
1043 DRM_DEBUG("execbuf with invalid ring: %d\n",
1044 (int)(args->flags & I915_EXEC_RING_MASK));
1048 mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1049 mask = I915_EXEC_CONSTANTS_MASK;
1051 case I915_EXEC_CONSTANTS_REL_GENERAL:
1052 case I915_EXEC_CONSTANTS_ABSOLUTE:
1053 case I915_EXEC_CONSTANTS_REL_SURFACE:
1054 if (ring == &dev_priv->ring[RCS] &&
1055 mode != dev_priv->relative_constants_mode) {
1056 if (INTEL_INFO(dev)->gen < 4)
1059 if (INTEL_INFO(dev)->gen > 5 &&
1060 mode == I915_EXEC_CONSTANTS_REL_SURFACE)
1063 /* The HW changed the meaning on this bit on gen6 */
1064 if (INTEL_INFO(dev)->gen >= 6)
1065 mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1069 DRM_DEBUG("execbuf with unknown constants: %d\n", mode);
1073 if (args->buffer_count < 1) {
1074 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1078 if (args->num_cliprects != 0) {
1079 if (ring != &dev_priv->ring[RCS]) {
1080 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1084 if (INTEL_INFO(dev)->gen >= 5) {
1085 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1089 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1090 DRM_DEBUG("execbuf with %u cliprects\n",
1091 args->num_cliprects);
1095 cliprects = kcalloc(args->num_cliprects,
1098 if (cliprects == NULL) {
1103 if (copy_from_user(cliprects,
1104 to_user_ptr(args->cliprects_ptr),
1105 sizeof(*cliprects)*args->num_cliprects)) {
1111 ret = i915_mutex_lock_interruptible(dev);
1115 if (dev_priv->ums.mm_suspended) {
1116 mutex_unlock(&dev->struct_mutex);
1121 eb = eb_create(args, vm);
1123 mutex_unlock(&dev->struct_mutex);
1128 /* Look up object handles */
1129 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1133 /* take note of the batch buffer before we might reorder the lists */
1134 batch_obj = list_entry(eb->vmas.prev, struct i915_vma, exec_list)->obj;
1136 /* Move the objects en-masse into the GTT, evicting if necessary. */
1137 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1138 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
1142 /* The objects are in their final locations, apply the relocations. */
1144 ret = i915_gem_execbuffer_relocate(eb, vm);
1146 if (ret == -EFAULT) {
1147 ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1149 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1155 /* Set the pending read domains for the batch buffer to COMMAND */
1156 if (batch_obj->base.pending_write_domain) {
1157 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1161 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1163 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1164 * batch" bit. Hence we need to pin secure batches into the global gtt.
1165 * hsw should have this fixed, but bdw mucks it up again. */
1166 if (flags & I915_DISPATCH_SECURE && !batch_obj->has_global_gtt_mapping)
1167 i915_gem_gtt_bind_object(batch_obj, batch_obj->cache_level);
1169 ret = i915_gem_execbuffer_move_to_gpu(ring, &eb->vmas);
1173 hs = i915_gem_context_get_hang_stats(dev, file, ctx_id);
1184 ret = i915_switch_context(ring, file, ctx_id);
1188 if (ring == &dev_priv->ring[RCS] &&
1189 mode != dev_priv->relative_constants_mode) {
1190 ret = intel_ring_begin(ring, 4);
1194 intel_ring_emit(ring, MI_NOOP);
1195 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1196 intel_ring_emit(ring, INSTPM);
1197 intel_ring_emit(ring, mask << 16 | mode);
1198 intel_ring_advance(ring);
1200 dev_priv->relative_constants_mode = mode;
1203 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1204 ret = i915_reset_gen7_sol_offsets(dev, ring);
1209 exec_start = i915_gem_obj_offset(batch_obj, vm) +
1210 args->batch_start_offset;
1211 exec_len = args->batch_len;
1213 for (i = 0; i < args->num_cliprects; i++) {
1214 ret = i915_emit_box(dev, &cliprects[i],
1215 args->DR1, args->DR4);
1219 ret = ring->dispatch_execbuffer(ring,
1220 exec_start, exec_len,
1226 ret = ring->dispatch_execbuffer(ring,
1227 exec_start, exec_len,
1233 trace_i915_gem_ring_dispatch(ring, intel_ring_get_seqno(ring), flags);
1235 i915_gem_execbuffer_move_to_active(&eb->vmas, ring);
1236 i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
1241 mutex_unlock(&dev->struct_mutex);
1249 * Legacy execbuffer just creates an exec2 list from the original exec object
1250 * list array and passes it to the real function.
1253 i915_gem_execbuffer(struct drm_device *dev, void *data,
1254 struct drm_file *file)
1256 struct drm_i915_private *dev_priv = dev->dev_private;
1257 struct drm_i915_gem_execbuffer *args = data;
1258 struct drm_i915_gem_execbuffer2 exec2;
1259 struct drm_i915_gem_exec_object *exec_list = NULL;
1260 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1263 if (args->buffer_count < 1) {
1264 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1268 /* Copy in the exec list from userland */
1269 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1270 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1271 if (exec_list == NULL || exec2_list == NULL) {
1272 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1273 args->buffer_count);
1274 drm_free_large(exec_list);
1275 drm_free_large(exec2_list);
1278 ret = copy_from_user(exec_list,
1279 to_user_ptr(args->buffers_ptr),
1280 sizeof(*exec_list) * args->buffer_count);
1282 DRM_DEBUG("copy %d exec entries failed %d\n",
1283 args->buffer_count, ret);
1284 drm_free_large(exec_list);
1285 drm_free_large(exec2_list);
1289 for (i = 0; i < args->buffer_count; i++) {
1290 exec2_list[i].handle = exec_list[i].handle;
1291 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1292 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1293 exec2_list[i].alignment = exec_list[i].alignment;
1294 exec2_list[i].offset = exec_list[i].offset;
1295 if (INTEL_INFO(dev)->gen < 4)
1296 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1298 exec2_list[i].flags = 0;
1301 exec2.buffers_ptr = args->buffers_ptr;
1302 exec2.buffer_count = args->buffer_count;
1303 exec2.batch_start_offset = args->batch_start_offset;
1304 exec2.batch_len = args->batch_len;
1305 exec2.DR1 = args->DR1;
1306 exec2.DR4 = args->DR4;
1307 exec2.num_cliprects = args->num_cliprects;
1308 exec2.cliprects_ptr = args->cliprects_ptr;
1309 exec2.flags = I915_EXEC_RENDER;
1310 i915_execbuffer2_set_context_id(exec2, 0);
1312 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list,
1313 &dev_priv->gtt.base);
1315 /* Copy the new buffer offsets back to the user's exec list. */
1316 for (i = 0; i < args->buffer_count; i++)
1317 exec_list[i].offset = exec2_list[i].offset;
1318 /* ... and back out to userspace */
1319 ret = copy_to_user(to_user_ptr(args->buffers_ptr),
1321 sizeof(*exec_list) * args->buffer_count);
1324 DRM_DEBUG("failed to copy %d exec entries "
1325 "back to user (%d)\n",
1326 args->buffer_count, ret);
1330 drm_free_large(exec_list);
1331 drm_free_large(exec2_list);
1336 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1337 struct drm_file *file)
1339 struct drm_i915_private *dev_priv = dev->dev_private;
1340 struct drm_i915_gem_execbuffer2 *args = data;
1341 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1344 if (args->buffer_count < 1 ||
1345 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1346 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1350 exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1351 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
1352 if (exec2_list == NULL)
1353 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1354 args->buffer_count);
1355 if (exec2_list == NULL) {
1356 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1357 args->buffer_count);
1360 ret = copy_from_user(exec2_list,
1361 to_user_ptr(args->buffers_ptr),
1362 sizeof(*exec2_list) * args->buffer_count);
1364 DRM_DEBUG("copy %d exec entries failed %d\n",
1365 args->buffer_count, ret);
1366 drm_free_large(exec2_list);
1370 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list,
1371 &dev_priv->gtt.base);
1373 /* Copy the new buffer offsets back to the user's exec list. */
1374 ret = copy_to_user(to_user_ptr(args->buffers_ptr),
1376 sizeof(*exec2_list) * args->buffer_count);
1379 DRM_DEBUG("failed to copy %d exec entries "
1380 "back to user (%d)\n",
1381 args->buffer_count, ret);
1385 drm_free_large(exec2_list);