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/pagemap.h>
36 #define __EXEC_OBJECT_HAS_PIN (1<<31)
37 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
38 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
39 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
41 #define BATCH_OFFSET_BIAS (256*1024)
44 struct list_head vmas;
47 struct i915_vma *lut[0];
48 struct hlist_head buckets[0];
52 static struct eb_vmas *
53 eb_create(struct drm_i915_gem_execbuffer2 *args)
55 struct eb_vmas *eb = NULL;
57 if (args->flags & I915_EXEC_HANDLE_LUT) {
58 unsigned size = args->buffer_count;
59 size *= sizeof(struct i915_vma *);
60 size += sizeof(struct eb_vmas);
61 eb = kmalloc(size, M_DRM, M_NOWAIT);
65 unsigned size = args->buffer_count;
66 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
67 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
68 while (count > 2*size)
70 eb = kzalloc(count*sizeof(struct hlist_head) +
71 sizeof(struct eb_vmas),
78 eb->and = -args->buffer_count;
80 INIT_LIST_HEAD(&eb->vmas);
85 eb_reset(struct eb_vmas *eb)
88 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
92 eb_lookup_vmas(struct eb_vmas *eb,
93 struct drm_i915_gem_exec_object2 *exec,
94 const struct drm_i915_gem_execbuffer2 *args,
95 struct i915_address_space *vm,
96 struct drm_file *file)
98 struct drm_i915_gem_object *obj;
99 struct list_head objects;
102 INIT_LIST_HEAD(&objects);
103 lockmgr(&file->table_lock, LK_EXCLUSIVE);
104 /* Grab a reference to the object and release the lock so we can lookup
105 * or create the VMA without using GFP_ATOMIC */
106 for (i = 0; i < args->buffer_count; i++) {
107 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
109 lockmgr(&file->table_lock, LK_RELEASE);
110 DRM_DEBUG("Invalid object handle %d at index %d\n",
116 if (!list_empty(&obj->obj_exec_link)) {
117 lockmgr(&file->table_lock, LK_RELEASE);
118 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
119 obj, exec[i].handle, i);
124 drm_gem_object_reference(&obj->base);
125 list_add_tail(&obj->obj_exec_link, &objects);
127 lockmgr(&file->table_lock, LK_RELEASE);
130 while (!list_empty(&objects)) {
131 struct i915_vma *vma;
133 obj = list_first_entry(&objects,
134 struct drm_i915_gem_object,
138 * NOTE: We can leak any vmas created here when something fails
139 * later on. But that's no issue since vma_unbind can deal with
140 * vmas which are not actually bound. And since only
141 * lookup_or_create exists as an interface to get at the vma
142 * from the (obj, vm) we don't run the risk of creating
143 * duplicated vmas for the same vm.
145 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
147 DRM_DEBUG("Failed to lookup VMA\n");
152 /* Transfer ownership from the objects list to the vmas list. */
153 list_add_tail(&vma->exec_list, &eb->vmas);
154 list_del_init(&obj->obj_exec_link);
156 vma->exec_entry = &exec[i];
160 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
161 vma->exec_handle = handle;
162 hlist_add_head(&vma->exec_node,
163 &eb->buckets[handle & eb->and]);
172 while (!list_empty(&objects)) {
173 obj = list_first_entry(&objects,
174 struct drm_i915_gem_object,
176 list_del_init(&obj->obj_exec_link);
177 drm_gem_object_unreference(&obj->base);
180 * Objects already transfered to the vmas list will be unreferenced by
187 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
190 if (handle >= -eb->and)
192 return eb->lut[handle];
194 struct hlist_head *head;
195 struct hlist_node *node;
197 head = &eb->buckets[handle & eb->and];
198 hlist_for_each(node, head) {
199 struct i915_vma *vma;
201 vma = hlist_entry(node, struct i915_vma, exec_node);
202 if (vma->exec_handle == handle)
210 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
212 struct drm_i915_gem_exec_object2 *entry;
213 struct drm_i915_gem_object *obj = vma->obj;
215 if (!drm_mm_node_allocated(&vma->node))
218 entry = vma->exec_entry;
220 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
221 i915_gem_object_unpin_fence(obj);
223 if (entry->flags & __EXEC_OBJECT_HAS_PIN)
226 entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
229 static void eb_destroy(struct eb_vmas *eb)
231 while (!list_empty(&eb->vmas)) {
232 struct i915_vma *vma;
234 vma = list_first_entry(&eb->vmas,
237 list_del_init(&vma->exec_list);
238 i915_gem_execbuffer_unreserve_vma(vma);
239 drm_gem_object_unreference(&vma->obj->base);
244 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
246 return (HAS_LLC(obj->base.dev) ||
247 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
248 !obj->map_and_fenceable ||
249 obj->cache_level != I915_CACHE_NONE);
253 relocate_entry_cpu(struct drm_i915_gem_object *obj,
254 struct drm_i915_gem_relocation_entry *reloc,
255 uint64_t target_offset)
257 struct drm_device *dev = obj->base.dev;
258 uint32_t page_offset = offset_in_page(reloc->offset);
259 uint64_t delta = reloc->delta + target_offset;
263 ret = i915_gem_object_set_to_cpu_domain(obj, true);
267 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
268 reloc->offset >> PAGE_SHIFT));
269 *(uint32_t *)(vaddr + page_offset) = lower_32_bits(delta);
271 if (INTEL_INFO(dev)->gen >= 8) {
272 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
274 if (page_offset == 0) {
275 kunmap_atomic(vaddr);
276 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
277 (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
280 *(uint32_t *)(vaddr + page_offset) = upper_32_bits(delta);
283 kunmap_atomic(vaddr);
289 relocate_entry_gtt(struct drm_i915_gem_object *obj,
290 struct drm_i915_gem_relocation_entry *reloc,
291 uint64_t target_offset)
293 struct drm_device *dev = obj->base.dev;
294 struct drm_i915_private *dev_priv = dev->dev_private;
295 uint64_t delta = reloc->delta + target_offset;
297 void __iomem *reloc_page;
300 ret = i915_gem_object_set_to_gtt_domain(obj, true);
304 ret = i915_gem_object_put_fence(obj);
308 /* Map the page containing the relocation we're going to perform. */
309 offset = i915_gem_obj_ggtt_offset(obj);
310 offset += reloc->offset;
311 reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
312 offset & ~PAGE_MASK);
313 iowrite32(lower_32_bits(delta), reloc_page + offset_in_page(offset));
315 if (INTEL_INFO(dev)->gen >= 8) {
316 offset += sizeof(uint32_t);
318 if (offset_in_page(offset) == 0) {
319 io_mapping_unmap_atomic(reloc_page);
321 io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
325 iowrite32(upper_32_bits(delta),
326 reloc_page + offset_in_page(offset));
329 io_mapping_unmap_atomic(reloc_page);
335 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
337 struct drm_i915_gem_relocation_entry *reloc)
339 struct drm_device *dev = obj->base.dev;
340 struct drm_gem_object *target_obj;
341 struct drm_i915_gem_object *target_i915_obj;
342 struct i915_vma *target_vma;
343 uint64_t target_offset;
346 /* we've already hold a reference to all valid objects */
347 target_vma = eb_get_vma(eb, reloc->target_handle);
348 if (unlikely(target_vma == NULL))
350 target_i915_obj = target_vma->obj;
351 target_obj = &target_vma->obj->base;
353 target_offset = target_vma->node.start;
355 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
356 * pipe_control writes because the gpu doesn't properly redirect them
357 * through the ppgtt for non_secure batchbuffers. */
358 if (unlikely(IS_GEN6(dev) &&
359 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
360 !target_i915_obj->has_global_gtt_mapping)) {
361 struct i915_vma *vma =
362 list_first_entry(&target_i915_obj->vma_list,
363 typeof(*vma), vma_link);
364 vma->bind_vma(vma, target_i915_obj->cache_level, GLOBAL_BIND);
367 /* Validate that the target is in a valid r/w GPU domain */
368 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
369 DRM_DEBUG("reloc with multiple write domains: "
370 "obj %p target %d offset %d "
371 "read %08x write %08x",
372 obj, reloc->target_handle,
375 reloc->write_domain);
378 if (unlikely((reloc->write_domain | reloc->read_domains)
379 & ~I915_GEM_GPU_DOMAINS)) {
380 DRM_DEBUG("reloc with read/write non-GPU domains: "
381 "obj %p target %d offset %d "
382 "read %08x write %08x",
383 obj, reloc->target_handle,
386 reloc->write_domain);
390 target_obj->pending_read_domains |= reloc->read_domains;
391 target_obj->pending_write_domain |= reloc->write_domain;
393 /* If the relocation already has the right value in it, no
394 * more work needs to be done.
396 if (target_offset == reloc->presumed_offset)
399 /* Check that the relocation address is valid... */
400 if (unlikely(reloc->offset >
401 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
402 DRM_DEBUG("Relocation beyond object bounds: "
403 "obj %p target %d offset %d size %d.\n",
404 obj, reloc->target_handle,
406 (int) obj->base.size);
409 if (unlikely(reloc->offset & 3)) {
410 DRM_DEBUG("Relocation not 4-byte aligned: "
411 "obj %p target %d offset %d.\n",
412 obj, reloc->target_handle,
413 (int) reloc->offset);
417 /* We can't wait for rendering with pagefaults disabled */
418 if (obj->active && (curthread->td_flags & TDF_NOFAULT))
421 if (use_cpu_reloc(obj))
422 ret = relocate_entry_cpu(obj, reloc, target_offset);
424 ret = relocate_entry_gtt(obj, reloc, target_offset);
429 /* and update the user's relocation entry */
430 reloc->presumed_offset = target_offset;
436 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
439 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
440 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
441 struct drm_i915_gem_relocation_entry __user *user_relocs;
442 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
445 user_relocs = to_user_ptr(entry->relocs_ptr);
447 remain = entry->relocation_count;
449 struct drm_i915_gem_relocation_entry *r = stack_reloc;
451 if (count > ARRAY_SIZE(stack_reloc))
452 count = ARRAY_SIZE(stack_reloc);
455 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
459 u64 offset = r->presumed_offset;
461 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r);
465 if (r->presumed_offset != offset &&
466 __copy_to_user_inatomic(&user_relocs->presumed_offset,
468 sizeof(r->presumed_offset))) {
482 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
484 struct drm_i915_gem_relocation_entry *relocs)
486 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
489 for (i = 0; i < entry->relocation_count; i++) {
490 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i]);
499 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
501 struct i915_vma *vma;
504 /* This is the fast path and we cannot handle a pagefault whilst
505 * holding the struct mutex lest the user pass in the relocations
506 * contained within a mmaped bo. For in such a case we, the page
507 * fault handler would call i915_gem_fault() and we would try to
508 * acquire the struct mutex again. Obviously this is bad and so
509 * lockdep complains vehemently.
512 list_for_each_entry(vma, &eb->vmas, exec_list) {
513 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
523 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
524 struct intel_engine_cs *ring,
527 struct drm_i915_gem_object *obj = vma->obj;
528 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
533 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
534 flags |= PIN_MAPPABLE;
535 if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
537 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
538 flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
540 ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
544 entry->flags |= __EXEC_OBJECT_HAS_PIN;
546 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
547 ret = i915_gem_object_get_fence(obj);
551 if (i915_gem_object_pin_fence(obj))
552 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
555 if (entry->offset != vma->node.start) {
556 entry->offset = vma->node.start;
560 if (entry->flags & EXEC_OBJECT_WRITE) {
561 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
562 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
569 need_reloc_mappable(struct i915_vma *vma)
571 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
573 if (entry->relocation_count == 0)
576 if (!i915_is_ggtt(vma->vm))
579 /* See also use_cpu_reloc() */
580 if (HAS_LLC(vma->obj->base.dev))
583 if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
590 eb_vma_misplaced(struct i915_vma *vma)
592 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
593 struct drm_i915_gem_object *obj = vma->obj;
595 WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
596 !i915_is_ggtt(vma->vm));
598 if (entry->alignment &&
599 vma->node.start & (entry->alignment - 1))
602 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
605 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
606 vma->node.start < BATCH_OFFSET_BIAS)
613 i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
614 struct list_head *vmas,
617 struct drm_i915_gem_object *obj;
618 struct i915_vma *vma;
619 struct i915_address_space *vm;
620 struct list_head ordered_vmas;
621 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
624 i915_gem_retire_requests_ring(ring);
626 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
628 INIT_LIST_HEAD(&ordered_vmas);
629 while (!list_empty(vmas)) {
630 struct drm_i915_gem_exec_object2 *entry;
631 bool need_fence, need_mappable;
633 vma = list_first_entry(vmas, struct i915_vma, exec_list);
635 entry = vma->exec_entry;
637 if (!has_fenced_gpu_access)
638 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
640 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
641 obj->tiling_mode != I915_TILING_NONE;
642 need_mappable = need_fence || need_reloc_mappable(vma);
645 entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
646 list_move(&vma->exec_list, &ordered_vmas);
648 list_move_tail(&vma->exec_list, &ordered_vmas);
650 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
651 obj->base.pending_write_domain = 0;
653 list_splice(&ordered_vmas, vmas);
655 /* Attempt to pin all of the buffers into the GTT.
656 * This is done in 3 phases:
658 * 1a. Unbind all objects that do not match the GTT constraints for
659 * the execbuffer (fenceable, mappable, alignment etc).
660 * 1b. Increment pin count for already bound objects.
661 * 2. Bind new objects.
662 * 3. Decrement pin count.
664 * This avoid unnecessary unbinding of later objects in order to make
665 * room for the earlier objects *unless* we need to defragment.
671 /* Unbind any ill-fitting objects or pin. */
672 list_for_each_entry(vma, vmas, exec_list) {
673 if (!drm_mm_node_allocated(&vma->node))
676 if (eb_vma_misplaced(vma))
677 ret = i915_vma_unbind(vma);
679 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
684 /* Bind fresh objects */
685 list_for_each_entry(vma, vmas, exec_list) {
686 if (drm_mm_node_allocated(&vma->node))
689 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
695 if (ret != -ENOSPC || retry++)
698 /* Decrement pin count for bound objects */
699 list_for_each_entry(vma, vmas, exec_list)
700 i915_gem_execbuffer_unreserve_vma(vma);
702 ret = i915_gem_evict_vm(vm, true);
709 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
710 struct drm_i915_gem_execbuffer2 *args,
711 struct drm_file *file,
712 struct intel_engine_cs *ring,
714 struct drm_i915_gem_exec_object2 *exec)
716 struct drm_i915_gem_relocation_entry *reloc;
717 struct i915_address_space *vm;
718 struct i915_vma *vma;
722 unsigned count = args->buffer_count;
724 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
726 /* We may process another execbuffer during the unlock... */
727 while (!list_empty(&eb->vmas)) {
728 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
729 list_del_init(&vma->exec_list);
730 i915_gem_execbuffer_unreserve_vma(vma);
731 drm_gem_object_unreference(&vma->obj->base);
734 mutex_unlock(&dev->struct_mutex);
737 for (i = 0; i < count; i++)
738 total += exec[i].relocation_count;
740 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
741 reloc = drm_malloc_ab(total, sizeof(*reloc));
742 if (reloc == NULL || reloc_offset == NULL) {
743 drm_free_large(reloc);
744 drm_free_large(reloc_offset);
745 mutex_lock(&dev->struct_mutex);
750 for (i = 0; i < count; i++) {
751 struct drm_i915_gem_relocation_entry __user *user_relocs;
752 u64 invalid_offset = (u64)-1;
755 user_relocs = to_user_ptr(exec[i].relocs_ptr);
757 if (copy_from_user(reloc+total, user_relocs,
758 exec[i].relocation_count * sizeof(*reloc))) {
760 mutex_lock(&dev->struct_mutex);
764 /* As we do not update the known relocation offsets after
765 * relocating (due to the complexities in lock handling),
766 * we need to mark them as invalid now so that we force the
767 * relocation processing next time. Just in case the target
768 * object is evicted and then rebound into its old
769 * presumed_offset before the next execbuffer - if that
770 * happened we would make the mistake of assuming that the
771 * relocations were valid.
773 for (j = 0; j < exec[i].relocation_count; j++) {
774 if (__copy_to_user(&user_relocs[j].presumed_offset,
776 sizeof(invalid_offset))) {
778 mutex_lock(&dev->struct_mutex);
783 reloc_offset[i] = total;
784 total += exec[i].relocation_count;
787 ret = i915_mutex_lock_interruptible(dev);
789 mutex_lock(&dev->struct_mutex);
793 /* reacquire the objects */
795 ret = eb_lookup_vmas(eb, exec, args, vm, file);
799 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
800 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
804 list_for_each_entry(vma, &eb->vmas, exec_list) {
805 int offset = vma->exec_entry - exec;
806 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
807 reloc + reloc_offset[offset]);
812 /* Leave the user relocations as are, this is the painfully slow path,
813 * and we want to avoid the complication of dropping the lock whilst
814 * having buffers reserved in the aperture and so causing spurious
815 * ENOSPC for random operations.
819 drm_free_large(reloc);
820 drm_free_large(reloc_offset);
825 i915_gem_execbuffer_move_to_gpu(struct intel_engine_cs *ring,
826 struct list_head *vmas)
828 struct i915_vma *vma;
829 uint32_t flush_domains = 0;
830 bool flush_chipset = false;
833 list_for_each_entry(vma, vmas, exec_list) {
834 struct drm_i915_gem_object *obj = vma->obj;
835 ret = i915_gem_object_sync(obj, ring);
839 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
840 flush_chipset |= i915_gem_clflush_object(obj, false);
842 flush_domains |= obj->base.write_domain;
846 i915_gem_chipset_flush(ring->dev);
848 if (flush_domains & I915_GEM_DOMAIN_GTT)
851 /* Unconditionally invalidate gpu caches and ensure that we do flush
852 * any residual writes from the previous batch.
854 return intel_ring_invalidate_all_caches(ring);
858 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
860 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
863 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
867 validate_exec_list(struct drm_device *dev,
868 struct drm_i915_gem_exec_object2 *exec,
871 unsigned relocs_total = 0;
872 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
873 unsigned invalid_flags;
876 invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
877 if (USES_FULL_PPGTT(dev))
878 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
880 for (i = 0; i < count; i++) {
881 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
882 int length; /* limited by fault_in_pages_readable() */
884 if (exec[i].flags & invalid_flags)
887 /* First check for malicious input causing overflow in
888 * the worst case where we need to allocate the entire
889 * relocation tree as a single array.
891 if (exec[i].relocation_count > relocs_max - relocs_total)
893 relocs_total += exec[i].relocation_count;
895 length = exec[i].relocation_count *
896 sizeof(struct drm_i915_gem_relocation_entry);
898 * We must check that the entire relocation array is safe
899 * to read, but since we may need to update the presumed
900 * offsets during execution, check for full write access.
903 if (!access_ok(VERIFY_WRITE, ptr, length))
907 if (likely(!i915.prefault_disable)) {
908 if (fault_in_multipages_readable(ptr, length))
916 static struct intel_context *
917 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
918 struct intel_engine_cs *ring, const u32 ctx_id)
920 struct intel_context *ctx = NULL;
921 struct i915_ctx_hang_stats *hs;
923 if (ring->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
924 return ERR_PTR(-EINVAL);
926 ctx = i915_gem_context_get(file->driver_priv, ctx_id);
930 hs = &ctx->hang_stats;
932 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
933 return ERR_PTR(-EIO);
936 if (i915.enable_execlists && !ctx->engine[ring->id].state) {
937 int ret = intel_lr_context_deferred_create(ctx, ring);
939 DRM_DEBUG("Could not create LRC %u: %d\n", ctx_id, ret);
948 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
949 struct intel_engine_cs *ring)
951 u32 seqno = intel_ring_get_seqno(ring);
952 struct i915_vma *vma;
954 list_for_each_entry(vma, vmas, exec_list) {
955 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
956 struct drm_i915_gem_object *obj = vma->obj;
957 u32 old_read = obj->base.read_domains;
958 u32 old_write = obj->base.write_domain;
960 obj->base.write_domain = obj->base.pending_write_domain;
961 if (obj->base.write_domain == 0)
962 obj->base.pending_read_domains |= obj->base.read_domains;
963 obj->base.read_domains = obj->base.pending_read_domains;
965 i915_vma_move_to_active(vma, ring);
966 if (obj->base.write_domain) {
968 obj->last_write_seqno = seqno;
970 intel_fb_obj_invalidate(obj, ring);
972 /* update for the implicit flush after a batch */
973 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
975 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
976 obj->last_fenced_seqno = seqno;
977 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
978 struct drm_i915_private *dev_priv = to_i915(ring->dev);
979 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
980 &dev_priv->mm.fence_list);
984 trace_i915_gem_object_change_domain(obj, old_read, old_write);
989 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
990 struct drm_file *file,
991 struct intel_engine_cs *ring,
992 struct drm_i915_gem_object *obj)
994 /* Unconditionally force add_request to emit a full flush. */
995 ring->gpu_caches_dirty = true;
997 /* Add a breadcrumb for the completion of the batch buffer */
998 (void)__i915_add_request(ring, file, obj, NULL);
1002 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1003 struct intel_engine_cs *ring)
1005 struct drm_i915_private *dev_priv = dev->dev_private;
1008 if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS]) {
1009 DRM_DEBUG("sol reset is gen7/rcs only\n");
1013 ret = intel_ring_begin(ring, 4 * 3);
1017 for (i = 0; i < 4; i++) {
1018 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1019 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1020 intel_ring_emit(ring, 0);
1023 intel_ring_advance(ring);
1029 i915_gem_ringbuffer_submission(struct drm_device *dev, struct drm_file *file,
1030 struct intel_engine_cs *ring,
1031 struct intel_context *ctx,
1032 struct drm_i915_gem_execbuffer2 *args,
1033 struct list_head *vmas,
1034 struct drm_i915_gem_object *batch_obj,
1035 u64 exec_start, u32 flags)
1037 struct drm_clip_rect *cliprects = NULL;
1038 struct drm_i915_private *dev_priv = dev->dev_private;
1044 if (args->num_cliprects != 0) {
1045 if (ring != &dev_priv->ring[RCS]) {
1046 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1050 if (INTEL_INFO(dev)->gen >= 5) {
1051 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1055 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1056 DRM_DEBUG("execbuf with %u cliprects\n",
1057 args->num_cliprects);
1061 cliprects = kcalloc(args->num_cliprects,
1064 if (cliprects == NULL) {
1069 if (copy_from_user(cliprects,
1070 to_user_ptr(args->cliprects_ptr),
1071 sizeof(*cliprects)*args->num_cliprects)) {
1076 if (args->DR4 == 0xffffffff) {
1077 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1081 if (args->DR1 || args->DR4 || args->cliprects_ptr) {
1082 DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
1087 ret = i915_gem_execbuffer_move_to_gpu(ring, vmas);
1091 ret = i915_switch_context(ring, ctx);
1095 instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1096 instp_mask = I915_EXEC_CONSTANTS_MASK;
1097 switch (instp_mode) {
1098 case I915_EXEC_CONSTANTS_REL_GENERAL:
1099 case I915_EXEC_CONSTANTS_ABSOLUTE:
1100 case I915_EXEC_CONSTANTS_REL_SURFACE:
1101 if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
1102 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1107 if (instp_mode != dev_priv->relative_constants_mode) {
1108 if (INTEL_INFO(dev)->gen < 4) {
1109 DRM_DEBUG("no rel constants on pre-gen4\n");
1114 if (INTEL_INFO(dev)->gen > 5 &&
1115 instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1116 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1121 /* The HW changed the meaning on this bit on gen6 */
1122 if (INTEL_INFO(dev)->gen >= 6)
1123 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1127 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1132 if (ring == &dev_priv->ring[RCS] &&
1133 instp_mode != dev_priv->relative_constants_mode) {
1134 ret = intel_ring_begin(ring, 4);
1138 intel_ring_emit(ring, MI_NOOP);
1139 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1140 intel_ring_emit(ring, INSTPM);
1141 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1142 intel_ring_advance(ring);
1144 dev_priv->relative_constants_mode = instp_mode;
1147 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1148 ret = i915_reset_gen7_sol_offsets(dev, ring);
1153 exec_len = args->batch_len;
1155 for (i = 0; i < args->num_cliprects; i++) {
1156 ret = i915_emit_box(dev, &cliprects[i],
1157 args->DR1, args->DR4);
1161 ret = ring->dispatch_execbuffer(ring,
1162 exec_start, exec_len,
1168 ret = ring->dispatch_execbuffer(ring,
1169 exec_start, exec_len,
1175 trace_i915_gem_ring_dispatch(ring, intel_ring_get_seqno(ring), flags);
1177 i915_gem_execbuffer_move_to_active(vmas, ring);
1178 i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
1186 * Find one BSD ring to dispatch the corresponding BSD command.
1187 * The Ring ID is returned.
1189 static int gen8_dispatch_bsd_ring(struct drm_device *dev,
1190 struct drm_file *file)
1192 struct drm_i915_private *dev_priv = dev->dev_private;
1193 struct drm_i915_file_private *file_priv = file->driver_priv;
1195 /* Check whether the file_priv is using one ring */
1196 if (file_priv->bsd_ring)
1197 return file_priv->bsd_ring->id;
1199 /* If no, use the ping-pong mechanism to select one ring */
1202 mutex_lock(&dev->struct_mutex);
1203 if (dev_priv->mm.bsd_ring_dispatch_index == 0) {
1205 dev_priv->mm.bsd_ring_dispatch_index = 1;
1208 dev_priv->mm.bsd_ring_dispatch_index = 0;
1210 file_priv->bsd_ring = &dev_priv->ring[ring_id];
1211 mutex_unlock(&dev->struct_mutex);
1216 static struct drm_i915_gem_object *
1217 eb_get_batch(struct eb_vmas *eb)
1219 struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
1222 * SNA is doing fancy tricks with compressing batch buffers, which leads
1223 * to negative relocation deltas. Usually that works out ok since the
1224 * relocate address is still positive, except when the batch is placed
1225 * very low in the GTT. Ensure this doesn't happen.
1227 * Note that actual hangs have only been observed on gen7, but for
1228 * paranoia do it everywhere.
1230 vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
1236 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1237 struct drm_file *file,
1238 struct drm_i915_gem_execbuffer2 *args,
1239 struct drm_i915_gem_exec_object2 *exec)
1241 struct drm_i915_private *dev_priv = dev->dev_private;
1243 struct drm_i915_gem_object *batch_obj;
1244 struct intel_engine_cs *ring;
1245 struct intel_context *ctx;
1246 struct i915_address_space *vm;
1247 const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1248 u64 exec_start = args->batch_start_offset;
1253 if (!i915_gem_check_execbuffer(args))
1256 ret = validate_exec_list(dev, exec, args->buffer_count);
1261 if (args->flags & I915_EXEC_SECURE) {
1262 flags |= I915_DISPATCH_SECURE;
1264 if (args->flags & I915_EXEC_IS_PINNED)
1265 flags |= I915_DISPATCH_PINNED;
1267 if ((args->flags & I915_EXEC_RING_MASK) > LAST_USER_RING) {
1268 DRM_DEBUG("execbuf with unknown ring: %d\n",
1269 (int)(args->flags & I915_EXEC_RING_MASK));
1273 if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_DEFAULT)
1274 ring = &dev_priv->ring[RCS];
1275 else if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_BSD) {
1276 if (HAS_BSD2(dev)) {
1278 ring_id = gen8_dispatch_bsd_ring(dev, file);
1279 ring = &dev_priv->ring[ring_id];
1281 ring = &dev_priv->ring[VCS];
1283 ring = &dev_priv->ring[(args->flags & I915_EXEC_RING_MASK) - 1];
1285 if (!intel_ring_initialized(ring)) {
1286 DRM_DEBUG("execbuf with invalid ring: %d\n",
1287 (int)(args->flags & I915_EXEC_RING_MASK));
1291 if (args->buffer_count < 1) {
1292 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1296 intel_runtime_pm_get(dev_priv);
1298 ret = i915_mutex_lock_interruptible(dev);
1302 if (dev_priv->ums.mm_suspended) {
1303 mutex_unlock(&dev->struct_mutex);
1308 ctx = i915_gem_validate_context(dev, file, ring, ctx_id);
1310 mutex_unlock(&dev->struct_mutex);
1315 i915_gem_context_reference(ctx);
1318 vm = &ctx->ppgtt->base;
1320 vm = &dev_priv->gtt.base;
1322 eb = eb_create(args);
1324 i915_gem_context_unreference(ctx);
1325 mutex_unlock(&dev->struct_mutex);
1330 /* Look up object handles */
1331 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1335 /* take note of the batch buffer before we might reorder the lists */
1336 batch_obj = eb_get_batch(eb);
1338 /* Move the objects en-masse into the GTT, evicting if necessary. */
1339 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1340 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
1344 /* The objects are in their final locations, apply the relocations. */
1346 ret = i915_gem_execbuffer_relocate(eb);
1348 if (ret == -EFAULT) {
1349 ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1351 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1357 /* Set the pending read domains for the batch buffer to COMMAND */
1358 if (batch_obj->base.pending_write_domain) {
1359 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1363 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1365 if (i915_needs_cmd_parser(ring)) {
1366 ret = i915_parse_cmds(ring,
1368 args->batch_start_offset,
1374 * XXX: Actually do this when enabling batch copy...
1376 * Set the DISPATCH_SECURE bit to remove the NON_SECURE bit
1377 * from MI_BATCH_BUFFER_START commands issued in the
1378 * dispatch_execbuffer implementations. We specifically don't
1379 * want that set when the command parser is enabled.
1383 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1384 * batch" bit. Hence we need to pin secure batches into the global gtt.
1385 * hsw should have this fixed, but bdw mucks it up again. */
1386 if (flags & I915_DISPATCH_SECURE) {
1388 * So on first glance it looks freaky that we pin the batch here
1389 * outside of the reservation loop. But:
1390 * - The batch is already pinned into the relevant ppgtt, so we
1391 * already have the backing storage fully allocated.
1392 * - No other BO uses the global gtt (well contexts, but meh),
1393 * so we don't really have issues with mutliple objects not
1394 * fitting due to fragmentation.
1395 * So this is actually safe.
1397 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1401 exec_start += i915_gem_obj_ggtt_offset(batch_obj);
1403 exec_start += i915_gem_obj_offset(batch_obj, vm);
1405 ret = dev_priv->gt.do_execbuf(dev, file, ring, ctx, args,
1406 &eb->vmas, batch_obj, exec_start, flags);
1409 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1410 * batch vma for correctness. For less ugly and less fragility this
1411 * needs to be adjusted to also track the ggtt batch vma properly as
1414 if (flags & I915_DISPATCH_SECURE)
1415 i915_gem_object_ggtt_unpin(batch_obj);
1417 /* the request owns the ref now */
1418 i915_gem_context_unreference(ctx);
1421 mutex_unlock(&dev->struct_mutex);
1424 /* intel_gpu_busy should also get a ref, so it will free when the device
1425 * is really idle. */
1426 intel_runtime_pm_put(dev_priv);
1431 * Legacy execbuffer just creates an exec2 list from the original exec object
1432 * list array and passes it to the real function.
1435 i915_gem_execbuffer(struct drm_device *dev, void *data,
1436 struct drm_file *file)
1438 struct drm_i915_gem_execbuffer *args = data;
1439 struct drm_i915_gem_execbuffer2 exec2;
1440 struct drm_i915_gem_exec_object *exec_list = NULL;
1441 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1444 if (args->buffer_count < 1) {
1445 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1449 /* Copy in the exec list from userland */
1450 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1451 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1452 if (exec_list == NULL || exec2_list == NULL) {
1453 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1454 args->buffer_count);
1455 drm_free_large(exec_list);
1456 drm_free_large(exec2_list);
1459 ret = copy_from_user(exec_list,
1460 to_user_ptr(args->buffers_ptr),
1461 sizeof(*exec_list) * args->buffer_count);
1463 DRM_DEBUG("copy %d exec entries failed %d\n",
1464 args->buffer_count, ret);
1465 drm_free_large(exec_list);
1466 drm_free_large(exec2_list);
1470 for (i = 0; i < args->buffer_count; i++) {
1471 exec2_list[i].handle = exec_list[i].handle;
1472 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1473 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1474 exec2_list[i].alignment = exec_list[i].alignment;
1475 exec2_list[i].offset = exec_list[i].offset;
1476 if (INTEL_INFO(dev)->gen < 4)
1477 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1479 exec2_list[i].flags = 0;
1482 exec2.buffers_ptr = args->buffers_ptr;
1483 exec2.buffer_count = args->buffer_count;
1484 exec2.batch_start_offset = args->batch_start_offset;
1485 exec2.batch_len = args->batch_len;
1486 exec2.DR1 = args->DR1;
1487 exec2.DR4 = args->DR4;
1488 exec2.num_cliprects = args->num_cliprects;
1489 exec2.cliprects_ptr = args->cliprects_ptr;
1490 exec2.flags = I915_EXEC_RENDER;
1491 i915_execbuffer2_set_context_id(exec2, 0);
1493 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1495 struct drm_i915_gem_exec_object __user *user_exec_list =
1496 to_user_ptr(args->buffers_ptr);
1498 /* Copy the new buffer offsets back to the user's exec list. */
1499 for (i = 0; i < args->buffer_count; i++) {
1500 ret = __copy_to_user(&user_exec_list[i].offset,
1501 &exec2_list[i].offset,
1502 sizeof(user_exec_list[i].offset));
1505 DRM_DEBUG("failed to copy %d exec entries "
1506 "back to user (%d)\n",
1507 args->buffer_count, ret);
1513 drm_free_large(exec_list);
1514 drm_free_large(exec2_list);
1519 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1520 struct drm_file *file)
1522 struct drm_i915_gem_execbuffer2 *args = data;
1523 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1526 if (args->buffer_count < 1 ||
1527 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1528 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1532 if (args->rsvd2 != 0) {
1533 DRM_DEBUG("dirty rvsd2 field\n");
1537 exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1539 if (exec2_list == NULL)
1540 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1541 args->buffer_count);
1542 if (exec2_list == NULL) {
1543 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1544 args->buffer_count);
1547 ret = copy_from_user(exec2_list,
1548 to_user_ptr(args->buffers_ptr),
1549 sizeof(*exec2_list) * args->buffer_count);
1551 DRM_DEBUG("copy %d exec entries failed %d\n",
1552 args->buffer_count, ret);
1553 drm_free_large(exec2_list);
1557 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1559 /* Copy the new buffer offsets back to the user's exec list. */
1560 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1561 to_user_ptr(args->buffers_ptr);
1564 for (i = 0; i < args->buffer_count; i++) {
1565 ret = __copy_to_user(&user_exec_list[i].offset,
1566 &exec2_list[i].offset,
1567 sizeof(user_exec_list[i].offset));
1570 DRM_DEBUG("failed to copy %d exec entries "
1572 args->buffer_count);
1578 drm_free_large(exec2_list);