2 * Copyright © 2008,2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
25 * Chris Wilson <chris@chris-wilson.co.uk>
29 #include <linux/dma_remapping.h>
30 #include <linux/reservation.h>
31 #include <linux/uaccess.h>
34 #include <drm/i915_drm.h>
37 #include "i915_gem_dmabuf.h"
38 #include "i915_trace.h"
39 #include "intel_drv.h"
40 #include "intel_frontbuffer.h"
42 #define __EXEC_OBJECT_HAS_PIN (1<<31)
43 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
44 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
45 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
46 #define __EXEC_OBJECT_INTERNAL_FLAGS (0xf<<28) /* all of the above */
48 #define BATCH_OFFSET_BIAS (256*1024)
50 struct i915_execbuffer_params {
51 struct drm_device *dev;
52 struct drm_file *file;
53 struct i915_vma *batch;
55 u32 args_batch_start_offset;
56 struct intel_engine_cs *engine;
57 struct i915_gem_context *ctx;
58 struct drm_i915_gem_request *request;
62 struct list_head vmas;
65 struct i915_vma *lut[0];
66 struct hlist_head buckets[0];
70 static struct eb_vmas *
71 eb_create(struct drm_i915_gem_execbuffer2 *args)
73 struct eb_vmas *eb = NULL;
75 if (args->flags & I915_EXEC_HANDLE_LUT) {
76 unsigned size = args->buffer_count;
77 size *= sizeof(struct i915_vma *);
78 size += sizeof(struct eb_vmas);
79 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
83 unsigned size = args->buffer_count;
84 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
85 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
86 while (count > 2*size)
88 eb = kzalloc(count*sizeof(struct hlist_head) +
89 sizeof(struct eb_vmas),
96 eb->and = -args->buffer_count;
98 INIT_LIST_HEAD(&eb->vmas);
103 eb_reset(struct eb_vmas *eb)
106 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
109 static struct i915_vma *
110 eb_get_batch(struct eb_vmas *eb)
112 struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
115 * SNA is doing fancy tricks with compressing batch buffers, which leads
116 * to negative relocation deltas. Usually that works out ok since the
117 * relocate address is still positive, except when the batch is placed
118 * very low in the GTT. Ensure this doesn't happen.
120 * Note that actual hangs have only been observed on gen7, but for
121 * paranoia do it everywhere.
123 if ((vma->exec_entry->flags & EXEC_OBJECT_PINNED) == 0)
124 vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
130 eb_lookup_vmas(struct eb_vmas *eb,
131 struct drm_i915_gem_exec_object2 *exec,
132 const struct drm_i915_gem_execbuffer2 *args,
133 struct i915_address_space *vm,
134 struct drm_file *file)
136 struct drm_i915_gem_object *obj;
137 struct list_head objects;
140 INIT_LIST_HEAD(&objects);
141 spin_lock(&file->table_lock);
142 /* Grab a reference to the object and release the lock so we can lookup
143 * or create the VMA without using GFP_ATOMIC */
144 for (i = 0; i < args->buffer_count; i++) {
145 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
147 spin_unlock(&file->table_lock);
148 DRM_DEBUG("Invalid object handle %d at index %d\n",
154 if (!list_empty(&obj->obj_exec_link)) {
155 spin_unlock(&file->table_lock);
156 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
157 obj, exec[i].handle, i);
162 i915_gem_object_get(obj);
163 list_add_tail(&obj->obj_exec_link, &objects);
165 spin_unlock(&file->table_lock);
168 while (!list_empty(&objects)) {
169 struct i915_vma *vma;
171 obj = list_first_entry(&objects,
172 struct drm_i915_gem_object,
176 * NOTE: We can leak any vmas created here when something fails
177 * later on. But that's no issue since vma_unbind can deal with
178 * vmas which are not actually bound. And since only
179 * lookup_or_create exists as an interface to get at the vma
180 * from the (obj, vm) we don't run the risk of creating
181 * duplicated vmas for the same vm.
183 vma = i915_gem_obj_lookup_or_create_vma(obj, vm, NULL);
184 if (unlikely(IS_ERR(vma))) {
185 DRM_DEBUG("Failed to lookup VMA\n");
190 /* Transfer ownership from the objects list to the vmas list. */
191 list_add_tail(&vma->exec_list, &eb->vmas);
192 list_del_init(&obj->obj_exec_link);
194 vma->exec_entry = &exec[i];
198 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
199 vma->exec_handle = handle;
200 hlist_add_head(&vma->exec_node,
201 &eb->buckets[handle & eb->and]);
210 while (!list_empty(&objects)) {
211 obj = list_first_entry(&objects,
212 struct drm_i915_gem_object,
214 list_del_init(&obj->obj_exec_link);
215 i915_gem_object_put(obj);
218 * Objects already transfered to the vmas list will be unreferenced by
225 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
228 if (handle >= -eb->and)
230 return eb->lut[handle];
232 struct hlist_head *head;
233 struct i915_vma *vma;
235 head = &eb->buckets[handle & eb->and];
236 hlist_for_each_entry(vma, head, exec_node) {
237 if (vma->exec_handle == handle)
245 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
247 struct drm_i915_gem_exec_object2 *entry;
248 struct drm_i915_gem_object *obj = vma->obj;
250 if (!drm_mm_node_allocated(&vma->node))
253 entry = vma->exec_entry;
255 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
256 i915_gem_object_unpin_fence(obj);
258 if (entry->flags & __EXEC_OBJECT_HAS_PIN)
259 __i915_vma_unpin(vma);
261 entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
264 static void eb_destroy(struct eb_vmas *eb)
266 while (!list_empty(&eb->vmas)) {
267 struct i915_vma *vma;
269 vma = list_first_entry(&eb->vmas,
272 list_del_init(&vma->exec_list);
273 i915_gem_execbuffer_unreserve_vma(vma);
279 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
281 return (HAS_LLC(obj->base.dev) ||
282 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
283 obj->cache_level != I915_CACHE_NONE);
286 /* Used to convert any address to canonical form.
287 * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
288 * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
289 * addresses to be in a canonical form:
290 * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
291 * canonical form [63:48] == [47]."
293 #define GEN8_HIGH_ADDRESS_BIT 47
294 static inline uint64_t gen8_canonical_addr(uint64_t address)
296 return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
299 static inline uint64_t gen8_noncanonical_addr(uint64_t address)
301 return address & ((1ULL << (GEN8_HIGH_ADDRESS_BIT + 1)) - 1);
304 static inline uint64_t
305 relocation_target(struct drm_i915_gem_relocation_entry *reloc,
306 uint64_t target_offset)
308 return gen8_canonical_addr((int)reloc->delta + target_offset);
314 enum { KMAP, IOMAP } type;
317 static void reloc_cache_init(struct reloc_cache *cache)
323 static void reloc_cache_fini(struct reloc_cache *cache)
328 switch (cache->type) {
330 kunmap_atomic(cache->vaddr);
334 io_mapping_unmap_atomic(cache->vaddr);
339 static void *reloc_kmap(struct drm_i915_gem_object *obj,
340 struct reloc_cache *cache,
343 if (cache->page == page)
347 kunmap_atomic(cache->vaddr);
350 cache->vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj, page));
357 relocate_entry_cpu(struct drm_i915_gem_object *obj,
358 struct drm_i915_gem_relocation_entry *reloc,
359 struct reloc_cache *cache,
360 uint64_t target_offset)
362 struct drm_device *dev = obj->base.dev;
363 uint32_t page_offset = offset_in_page(reloc->offset);
364 uint64_t delta = relocation_target(reloc, target_offset);
368 ret = i915_gem_object_set_to_cpu_domain(obj, true);
372 vaddr = reloc_kmap(obj, cache, reloc->offset >> PAGE_SHIFT);
373 *(uint32_t *)(vaddr + page_offset) = lower_32_bits(delta);
375 if (INTEL_GEN(dev) >= 8) {
376 page_offset += sizeof(uint32_t);
377 if (page_offset == PAGE_SIZE) {
378 vaddr = reloc_kmap(obj, cache, cache->page + 1);
381 *(uint32_t *)(vaddr + page_offset) = upper_32_bits(delta);
387 static void *reloc_iomap(struct drm_i915_private *i915,
388 struct reloc_cache *cache,
391 if (cache->page == offset >> PAGE_SHIFT)
395 io_mapping_unmap_atomic(cache->vaddr);
397 cache->page = offset >> PAGE_SHIFT;
399 io_mapping_map_atomic_wc(i915->ggtt.mappable,
407 relocate_entry_gtt(struct drm_i915_gem_object *obj,
408 struct drm_i915_gem_relocation_entry *reloc,
409 struct reloc_cache *cache,
410 uint64_t target_offset)
412 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
413 struct i915_vma *vma;
414 uint64_t delta = relocation_target(reloc, target_offset);
416 void __iomem *reloc_page;
419 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, PIN_MAPPABLE);
423 ret = i915_gem_object_set_to_gtt_domain(obj, true);
427 ret = i915_gem_object_put_fence(obj);
431 /* Map the page containing the relocation we're going to perform. */
432 offset = vma->node.start + reloc->offset;
433 reloc_page = reloc_iomap(dev_priv, cache, offset);
434 iowrite32(lower_32_bits(delta), reloc_page + offset_in_page(offset));
436 if (INTEL_GEN(dev_priv) >= 8) {
437 offset += sizeof(uint32_t);
438 if (offset_in_page(offset) == 0)
439 reloc_page = reloc_iomap(dev_priv, cache, offset);
440 iowrite32(upper_32_bits(delta),
441 reloc_page + offset_in_page(offset));
445 __i915_vma_unpin(vma);
450 clflush_write32(void *addr, uint32_t value)
452 /* This is not a fast path, so KISS. */
453 drm_clflush_virt_range(addr, sizeof(uint32_t));
454 *(uint32_t *)addr = value;
455 drm_clflush_virt_range(addr, sizeof(uint32_t));
459 relocate_entry_clflush(struct drm_i915_gem_object *obj,
460 struct drm_i915_gem_relocation_entry *reloc,
461 struct reloc_cache *cache,
462 uint64_t target_offset)
464 struct drm_device *dev = obj->base.dev;
465 uint32_t page_offset = offset_in_page(reloc->offset);
466 uint64_t delta = relocation_target(reloc, target_offset);
470 ret = i915_gem_object_set_to_gtt_domain(obj, true);
474 vaddr = reloc_kmap(obj, cache, reloc->offset >> PAGE_SHIFT);
475 clflush_write32(vaddr + page_offset, lower_32_bits(delta));
477 if (INTEL_GEN(dev) >= 8) {
478 page_offset += sizeof(uint32_t);
479 if (page_offset == PAGE_SIZE) {
480 vaddr = reloc_kmap(obj, cache, cache->page + 1);
483 clflush_write32(vaddr + page_offset, upper_32_bits(delta));
489 static bool object_is_idle(struct drm_i915_gem_object *obj)
491 unsigned long active = i915_gem_object_get_active(obj);
494 for_each_active(active, idx) {
495 if (!i915_gem_active_is_idle(&obj->last_read[idx],
496 &obj->base.dev->struct_mutex))
504 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
506 struct drm_i915_gem_relocation_entry *reloc,
507 struct reloc_cache *cache)
509 struct drm_device *dev = obj->base.dev;
510 struct drm_gem_object *target_obj;
511 struct drm_i915_gem_object *target_i915_obj;
512 struct i915_vma *target_vma;
513 uint64_t target_offset;
516 /* we've already hold a reference to all valid objects */
517 target_vma = eb_get_vma(eb, reloc->target_handle);
518 if (unlikely(target_vma == NULL))
520 target_i915_obj = target_vma->obj;
521 target_obj = &target_vma->obj->base;
523 target_offset = gen8_canonical_addr(target_vma->node.start);
525 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
526 * pipe_control writes because the gpu doesn't properly redirect them
527 * through the ppgtt for non_secure batchbuffers. */
528 if (unlikely(IS_GEN6(dev) &&
529 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
530 ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
532 if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
536 /* Validate that the target is in a valid r/w GPU domain */
537 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
538 DRM_DEBUG("reloc with multiple write domains: "
539 "obj %p target %d offset %d "
540 "read %08x write %08x",
541 obj, reloc->target_handle,
544 reloc->write_domain);
547 if (unlikely((reloc->write_domain | reloc->read_domains)
548 & ~I915_GEM_GPU_DOMAINS)) {
549 DRM_DEBUG("reloc with read/write non-GPU domains: "
550 "obj %p target %d offset %d "
551 "read %08x write %08x",
552 obj, reloc->target_handle,
555 reloc->write_domain);
559 target_obj->pending_read_domains |= reloc->read_domains;
560 target_obj->pending_write_domain |= reloc->write_domain;
562 /* If the relocation already has the right value in it, no
563 * more work needs to be done.
565 if (target_offset == reloc->presumed_offset)
568 /* Check that the relocation address is valid... */
569 if (unlikely(reloc->offset >
570 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
571 DRM_DEBUG("Relocation beyond object bounds: "
572 "obj %p target %d offset %d size %d.\n",
573 obj, reloc->target_handle,
575 (int) obj->base.size);
578 if (unlikely(reloc->offset & 3)) {
579 DRM_DEBUG("Relocation not 4-byte aligned: "
580 "obj %p target %d offset %d.\n",
581 obj, reloc->target_handle,
582 (int) reloc->offset);
586 /* We can't wait for rendering with pagefaults disabled */
587 if (pagefault_disabled() && !object_is_idle(obj))
590 if (use_cpu_reloc(obj))
591 ret = relocate_entry_cpu(obj, reloc, cache, target_offset);
592 else if (obj->map_and_fenceable)
593 ret = relocate_entry_gtt(obj, reloc, cache, target_offset);
594 else if (static_cpu_has(X86_FEATURE_CLFLUSH))
595 ret = relocate_entry_clflush(obj, reloc, cache, target_offset);
597 WARN_ONCE(1, "Impossible case in relocation handling\n");
604 /* and update the user's relocation entry */
605 reloc->presumed_offset = target_offset;
611 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
614 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
615 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
616 struct drm_i915_gem_relocation_entry __user *user_relocs;
617 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
618 struct reloc_cache cache;
621 user_relocs = u64_to_user_ptr(entry->relocs_ptr);
622 reloc_cache_init(&cache);
624 remain = entry->relocation_count;
626 struct drm_i915_gem_relocation_entry *r = stack_reloc;
628 if (count > ARRAY_SIZE(stack_reloc))
629 count = ARRAY_SIZE(stack_reloc);
632 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0]))) {
638 u64 offset = r->presumed_offset;
640 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r, &cache);
644 if (r->presumed_offset != offset &&
645 __put_user(r->presumed_offset,
646 &user_relocs->presumed_offset)) {
657 reloc_cache_fini(&cache);
663 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
665 struct drm_i915_gem_relocation_entry *relocs)
667 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
668 struct reloc_cache cache;
671 reloc_cache_init(&cache);
672 for (i = 0; i < entry->relocation_count; i++) {
673 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i], &cache);
677 reloc_cache_fini(&cache);
683 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
685 struct i915_vma *vma;
688 /* This is the fast path and we cannot handle a pagefault whilst
689 * holding the struct mutex lest the user pass in the relocations
690 * contained within a mmaped bo. For in such a case we, the page
691 * fault handler would call i915_gem_fault() and we would try to
692 * acquire the struct mutex again. Obviously this is bad and so
693 * lockdep complains vehemently.
696 list_for_each_entry(vma, &eb->vmas, exec_list) {
697 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
706 static bool only_mappable_for_reloc(unsigned int flags)
708 return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
709 __EXEC_OBJECT_NEEDS_MAP;
713 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
714 struct intel_engine_cs *engine,
717 struct drm_i915_gem_object *obj = vma->obj;
718 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
723 if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
726 if (!drm_mm_node_allocated(&vma->node)) {
727 /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
728 * limit address to the first 4GBs for unflagged objects.
730 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
731 flags |= PIN_ZONE_4G;
732 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
733 flags |= PIN_GLOBAL | PIN_MAPPABLE;
734 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
735 flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
736 if (entry->flags & EXEC_OBJECT_PINNED)
737 flags |= entry->offset | PIN_OFFSET_FIXED;
738 if ((flags & PIN_MAPPABLE) == 0)
742 ret = i915_vma_pin(vma,
746 if ((ret == -ENOSPC || ret == -E2BIG) &&
747 only_mappable_for_reloc(entry->flags))
748 ret = i915_vma_pin(vma,
751 flags & ~PIN_MAPPABLE);
755 entry->flags |= __EXEC_OBJECT_HAS_PIN;
757 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
758 ret = i915_gem_object_get_fence(obj);
762 if (i915_gem_object_pin_fence(obj))
763 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
766 if (entry->offset != vma->node.start) {
767 entry->offset = vma->node.start;
771 if (entry->flags & EXEC_OBJECT_WRITE) {
772 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
773 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
780 need_reloc_mappable(struct i915_vma *vma)
782 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
784 if (entry->relocation_count == 0)
787 if (!i915_vma_is_ggtt(vma))
790 /* See also use_cpu_reloc() */
791 if (HAS_LLC(vma->obj->base.dev))
794 if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
801 eb_vma_misplaced(struct i915_vma *vma)
803 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
804 struct drm_i915_gem_object *obj = vma->obj;
806 WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
807 !i915_vma_is_ggtt(vma));
809 if (entry->alignment &&
810 vma->node.start & (entry->alignment - 1))
813 if (vma->node.size < entry->pad_to_size)
816 if (entry->flags & EXEC_OBJECT_PINNED &&
817 vma->node.start != entry->offset)
820 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
821 vma->node.start < BATCH_OFFSET_BIAS)
824 /* avoid costly ping-pong once a batch bo ended up non-mappable */
825 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
826 return !only_mappable_for_reloc(entry->flags);
828 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
829 (vma->node.start + vma->node.size - 1) >> 32)
836 i915_gem_execbuffer_reserve(struct intel_engine_cs *engine,
837 struct list_head *vmas,
838 struct i915_gem_context *ctx,
841 struct drm_i915_gem_object *obj;
842 struct i915_vma *vma;
843 struct i915_address_space *vm;
844 struct list_head ordered_vmas;
845 struct list_head pinned_vmas;
846 bool has_fenced_gpu_access = INTEL_GEN(engine->i915) < 4;
849 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
851 INIT_LIST_HEAD(&ordered_vmas);
852 INIT_LIST_HEAD(&pinned_vmas);
853 while (!list_empty(vmas)) {
854 struct drm_i915_gem_exec_object2 *entry;
855 bool need_fence, need_mappable;
857 vma = list_first_entry(vmas, struct i915_vma, exec_list);
859 entry = vma->exec_entry;
861 if (ctx->flags & CONTEXT_NO_ZEROMAP)
862 entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
864 if (!has_fenced_gpu_access)
865 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
867 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
868 i915_gem_object_is_tiled(obj);
869 need_mappable = need_fence || need_reloc_mappable(vma);
871 if (entry->flags & EXEC_OBJECT_PINNED)
872 list_move_tail(&vma->exec_list, &pinned_vmas);
873 else if (need_mappable) {
874 entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
875 list_move(&vma->exec_list, &ordered_vmas);
877 list_move_tail(&vma->exec_list, &ordered_vmas);
879 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
880 obj->base.pending_write_domain = 0;
882 list_splice(&ordered_vmas, vmas);
883 list_splice(&pinned_vmas, vmas);
885 /* Attempt to pin all of the buffers into the GTT.
886 * This is done in 3 phases:
888 * 1a. Unbind all objects that do not match the GTT constraints for
889 * the execbuffer (fenceable, mappable, alignment etc).
890 * 1b. Increment pin count for already bound objects.
891 * 2. Bind new objects.
892 * 3. Decrement pin count.
894 * This avoid unnecessary unbinding of later objects in order to make
895 * room for the earlier objects *unless* we need to defragment.
901 /* Unbind any ill-fitting objects or pin. */
902 list_for_each_entry(vma, vmas, exec_list) {
903 if (!drm_mm_node_allocated(&vma->node))
906 if (eb_vma_misplaced(vma))
907 ret = i915_vma_unbind(vma);
909 ret = i915_gem_execbuffer_reserve_vma(vma,
916 /* Bind fresh objects */
917 list_for_each_entry(vma, vmas, exec_list) {
918 if (drm_mm_node_allocated(&vma->node))
921 ret = i915_gem_execbuffer_reserve_vma(vma, engine,
928 if (ret != -ENOSPC || retry++)
931 /* Decrement pin count for bound objects */
932 list_for_each_entry(vma, vmas, exec_list)
933 i915_gem_execbuffer_unreserve_vma(vma);
935 ret = i915_gem_evict_vm(vm, true);
942 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
943 struct drm_i915_gem_execbuffer2 *args,
944 struct drm_file *file,
945 struct intel_engine_cs *engine,
947 struct drm_i915_gem_exec_object2 *exec,
948 struct i915_gem_context *ctx)
950 struct drm_i915_gem_relocation_entry *reloc;
951 struct i915_address_space *vm;
952 struct i915_vma *vma;
956 unsigned count = args->buffer_count;
958 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
960 /* We may process another execbuffer during the unlock... */
961 while (!list_empty(&eb->vmas)) {
962 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
963 list_del_init(&vma->exec_list);
964 i915_gem_execbuffer_unreserve_vma(vma);
968 mutex_unlock(&dev->struct_mutex);
971 for (i = 0; i < count; i++)
972 total += exec[i].relocation_count;
974 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
975 reloc = drm_malloc_ab(total, sizeof(*reloc));
976 if (reloc == NULL || reloc_offset == NULL) {
977 drm_free_large(reloc);
978 drm_free_large(reloc_offset);
979 mutex_lock(&dev->struct_mutex);
984 for (i = 0; i < count; i++) {
985 struct drm_i915_gem_relocation_entry __user *user_relocs;
986 u64 invalid_offset = (u64)-1;
989 user_relocs = u64_to_user_ptr(exec[i].relocs_ptr);
991 if (copy_from_user(reloc+total, user_relocs,
992 exec[i].relocation_count * sizeof(*reloc))) {
994 mutex_lock(&dev->struct_mutex);
998 /* As we do not update the known relocation offsets after
999 * relocating (due to the complexities in lock handling),
1000 * we need to mark them as invalid now so that we force the
1001 * relocation processing next time. Just in case the target
1002 * object is evicted and then rebound into its old
1003 * presumed_offset before the next execbuffer - if that
1004 * happened we would make the mistake of assuming that the
1005 * relocations were valid.
1007 for (j = 0; j < exec[i].relocation_count; j++) {
1008 if (__copy_to_user(&user_relocs[j].presumed_offset,
1010 sizeof(invalid_offset))) {
1012 mutex_lock(&dev->struct_mutex);
1017 reloc_offset[i] = total;
1018 total += exec[i].relocation_count;
1021 ret = i915_mutex_lock_interruptible(dev);
1023 mutex_lock(&dev->struct_mutex);
1027 /* reacquire the objects */
1029 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1033 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1034 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1039 list_for_each_entry(vma, &eb->vmas, exec_list) {
1040 int offset = vma->exec_entry - exec;
1041 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
1042 reloc + reloc_offset[offset]);
1047 /* Leave the user relocations as are, this is the painfully slow path,
1048 * and we want to avoid the complication of dropping the lock whilst
1049 * having buffers reserved in the aperture and so causing spurious
1050 * ENOSPC for random operations.
1054 drm_free_large(reloc);
1055 drm_free_large(reloc_offset);
1059 static unsigned int eb_other_engines(struct drm_i915_gem_request *req)
1063 mask = ~intel_engine_flag(req->engine) & I915_BO_ACTIVE_MASK;
1064 mask <<= I915_BO_ACTIVE_SHIFT;
1070 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
1071 struct list_head *vmas)
1073 const unsigned int other_rings = eb_other_engines(req);
1074 struct i915_vma *vma;
1077 list_for_each_entry(vma, vmas, exec_list) {
1078 struct drm_i915_gem_object *obj = vma->obj;
1080 if (obj->flags & other_rings) {
1081 ret = i915_gem_object_sync(obj, req);
1086 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
1087 i915_gem_clflush_object(obj, false);
1090 /* Unconditionally flush any chipset caches (for streaming writes). */
1091 i915_gem_chipset_flush(req->engine->i915);
1093 /* Unconditionally invalidate GPU caches and TLBs. */
1094 return req->engine->emit_flush(req, EMIT_INVALIDATE);
1098 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
1100 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
1103 /* Kernel clipping was a DRI1 misfeature */
1104 if (exec->num_cliprects || exec->cliprects_ptr)
1107 if (exec->DR4 == 0xffffffff) {
1108 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1111 if (exec->DR1 || exec->DR4)
1114 if ((exec->batch_start_offset | exec->batch_len) & 0x7)
1121 validate_exec_list(struct drm_device *dev,
1122 struct drm_i915_gem_exec_object2 *exec,
1125 unsigned relocs_total = 0;
1126 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
1127 unsigned invalid_flags;
1130 /* INTERNAL flags must not overlap with external ones */
1131 BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & ~__EXEC_OBJECT_UNKNOWN_FLAGS);
1133 invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
1134 if (USES_FULL_PPGTT(dev))
1135 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
1137 for (i = 0; i < count; i++) {
1138 char __user *ptr = u64_to_user_ptr(exec[i].relocs_ptr);
1139 int length; /* limited by fault_in_pages_readable() */
1141 if (exec[i].flags & invalid_flags)
1144 /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
1145 * any non-page-aligned or non-canonical addresses.
1147 if (exec[i].flags & EXEC_OBJECT_PINNED) {
1148 if (exec[i].offset !=
1149 gen8_canonical_addr(exec[i].offset & PAGE_MASK))
1152 /* From drm_mm perspective address space is continuous,
1153 * so from this point we're always using non-canonical
1156 exec[i].offset = gen8_noncanonical_addr(exec[i].offset);
1159 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
1162 /* pad_to_size was once a reserved field, so sanitize it */
1163 if (exec[i].flags & EXEC_OBJECT_PAD_TO_SIZE) {
1164 if (offset_in_page(exec[i].pad_to_size))
1167 exec[i].pad_to_size = 0;
1170 /* First check for malicious input causing overflow in
1171 * the worst case where we need to allocate the entire
1172 * relocation tree as a single array.
1174 if (exec[i].relocation_count > relocs_max - relocs_total)
1176 relocs_total += exec[i].relocation_count;
1178 length = exec[i].relocation_count *
1179 sizeof(struct drm_i915_gem_relocation_entry);
1181 * We must check that the entire relocation array is safe
1182 * to read, but since we may need to update the presumed
1183 * offsets during execution, check for full write access.
1185 if (!access_ok(VERIFY_WRITE, ptr, length))
1188 if (likely(!i915.prefault_disable)) {
1189 if (fault_in_multipages_readable(ptr, length))
1197 static struct i915_gem_context *
1198 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
1199 struct intel_engine_cs *engine, const u32 ctx_id)
1201 struct i915_gem_context *ctx = NULL;
1202 struct i915_ctx_hang_stats *hs;
1204 if (engine->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
1205 return ERR_PTR(-EINVAL);
1207 ctx = i915_gem_context_lookup(file->driver_priv, ctx_id);
1211 hs = &ctx->hang_stats;
1213 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1214 return ERR_PTR(-EIO);
1220 void i915_vma_move_to_active(struct i915_vma *vma,
1221 struct drm_i915_gem_request *req,
1224 struct drm_i915_gem_object *obj = vma->obj;
1225 const unsigned int idx = req->engine->id;
1227 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1229 obj->dirty = 1; /* be paranoid */
1231 /* Add a reference if we're newly entering the active list.
1232 * The order in which we add operations to the retirement queue is
1233 * vital here: mark_active adds to the start of the callback list,
1234 * such that subsequent callbacks are called first. Therefore we
1235 * add the active reference first and queue for it to be dropped
1238 if (!i915_gem_object_is_active(obj))
1239 i915_gem_object_get(obj);
1240 i915_gem_object_set_active(obj, idx);
1241 i915_gem_active_set(&obj->last_read[idx], req);
1243 if (flags & EXEC_OBJECT_WRITE) {
1244 i915_gem_active_set(&obj->last_write, req);
1246 intel_fb_obj_invalidate(obj, ORIGIN_CS);
1248 /* update for the implicit flush after a batch */
1249 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1252 if (flags & EXEC_OBJECT_NEEDS_FENCE) {
1253 i915_gem_active_set(&obj->last_fence, req);
1254 if (flags & __EXEC_OBJECT_HAS_FENCE) {
1255 struct drm_i915_private *dev_priv = req->i915;
1257 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
1258 &dev_priv->mm.fence_list);
1262 i915_vma_set_active(vma, idx);
1263 i915_gem_active_set(&vma->last_read[idx], req);
1264 list_move_tail(&vma->vm_link, &vma->vm->active_list);
1267 static void eb_export_fence(struct drm_i915_gem_object *obj,
1268 struct drm_i915_gem_request *req,
1271 struct reservation_object *resv;
1273 resv = i915_gem_object_get_dmabuf_resv(obj);
1277 /* Ignore errors from failing to allocate the new fence, we can't
1278 * handle an error right now. Worst case should be missed
1279 * synchronisation leading to rendering corruption.
1281 ww_mutex_lock(&resv->lock, NULL);
1282 if (flags & EXEC_OBJECT_WRITE)
1283 reservation_object_add_excl_fence(resv, &req->fence);
1284 else if (reservation_object_reserve_shared(resv) == 0)
1285 reservation_object_add_shared_fence(resv, &req->fence);
1286 ww_mutex_unlock(&resv->lock);
1290 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1291 struct drm_i915_gem_request *req)
1293 struct i915_vma *vma;
1295 list_for_each_entry(vma, vmas, exec_list) {
1296 struct drm_i915_gem_object *obj = vma->obj;
1297 u32 old_read = obj->base.read_domains;
1298 u32 old_write = obj->base.write_domain;
1300 obj->base.write_domain = obj->base.pending_write_domain;
1301 if (obj->base.write_domain)
1302 vma->exec_entry->flags |= EXEC_OBJECT_WRITE;
1304 obj->base.pending_read_domains |= obj->base.read_domains;
1305 obj->base.read_domains = obj->base.pending_read_domains;
1307 i915_vma_move_to_active(vma, req, vma->exec_entry->flags);
1308 eb_export_fence(obj, req, vma->exec_entry->flags);
1309 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1314 i915_reset_gen7_sol_offsets(struct drm_i915_gem_request *req)
1316 struct intel_ring *ring = req->ring;
1319 if (!IS_GEN7(req->i915) || req->engine->id != RCS) {
1320 DRM_DEBUG("sol reset is gen7/rcs only\n");
1324 ret = intel_ring_begin(req, 4 * 3);
1328 for (i = 0; i < 4; i++) {
1329 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1330 intel_ring_emit_reg(ring, GEN7_SO_WRITE_OFFSET(i));
1331 intel_ring_emit(ring, 0);
1334 intel_ring_advance(ring);
1339 static struct i915_vma *
1340 i915_gem_execbuffer_parse(struct intel_engine_cs *engine,
1341 struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1342 struct drm_i915_gem_object *batch_obj,
1344 u32 batch_start_offset,
1348 struct drm_i915_gem_object *shadow_batch_obj;
1349 struct i915_vma *vma;
1352 shadow_batch_obj = i915_gem_batch_pool_get(&engine->batch_pool,
1353 PAGE_ALIGN(batch_len));
1354 if (IS_ERR(shadow_batch_obj))
1355 return ERR_CAST(shadow_batch_obj);
1357 ret = intel_engine_cmd_parser(engine,
1364 if (ret == -EACCES) /* unhandled chained batch */
1371 vma = i915_gem_object_ggtt_pin(shadow_batch_obj, NULL, 0, 0, 0);
1375 memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1377 vma->exec_entry = shadow_exec_entry;
1378 vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1379 i915_gem_object_get(shadow_batch_obj);
1380 list_add_tail(&vma->exec_list, &eb->vmas);
1383 i915_gem_object_unpin_pages(shadow_batch_obj);
1388 execbuf_submit(struct i915_execbuffer_params *params,
1389 struct drm_i915_gem_execbuffer2 *args,
1390 struct list_head *vmas)
1392 struct drm_i915_private *dev_priv = params->request->i915;
1393 u64 exec_start, exec_len;
1398 ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1402 ret = i915_switch_context(params->request);
1406 instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1407 instp_mask = I915_EXEC_CONSTANTS_MASK;
1408 switch (instp_mode) {
1409 case I915_EXEC_CONSTANTS_REL_GENERAL:
1410 case I915_EXEC_CONSTANTS_ABSOLUTE:
1411 case I915_EXEC_CONSTANTS_REL_SURFACE:
1412 if (instp_mode != 0 && params->engine->id != RCS) {
1413 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1417 if (instp_mode != dev_priv->relative_constants_mode) {
1418 if (INTEL_INFO(dev_priv)->gen < 4) {
1419 DRM_DEBUG("no rel constants on pre-gen4\n");
1423 if (INTEL_INFO(dev_priv)->gen > 5 &&
1424 instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1425 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1429 /* The HW changed the meaning on this bit on gen6 */
1430 if (INTEL_INFO(dev_priv)->gen >= 6)
1431 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1435 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1439 if (params->engine->id == RCS &&
1440 instp_mode != dev_priv->relative_constants_mode) {
1441 struct intel_ring *ring = params->request->ring;
1443 ret = intel_ring_begin(params->request, 4);
1447 intel_ring_emit(ring, MI_NOOP);
1448 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1449 intel_ring_emit_reg(ring, INSTPM);
1450 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1451 intel_ring_advance(ring);
1453 dev_priv->relative_constants_mode = instp_mode;
1456 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1457 ret = i915_reset_gen7_sol_offsets(params->request);
1462 exec_len = args->batch_len;
1463 exec_start = params->batch->node.start +
1464 params->args_batch_start_offset;
1467 exec_len = params->batch->size;
1469 ret = params->engine->emit_bb_start(params->request,
1470 exec_start, exec_len,
1471 params->dispatch_flags);
1475 trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1477 i915_gem_execbuffer_move_to_active(vmas, params->request);
1483 * Find one BSD ring to dispatch the corresponding BSD command.
1484 * The engine index is returned.
1487 gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
1488 struct drm_file *file)
1490 struct drm_i915_file_private *file_priv = file->driver_priv;
1492 /* Check whether the file_priv has already selected one ring. */
1493 if ((int)file_priv->bsd_engine < 0) {
1494 /* If not, use the ping-pong mechanism to select one. */
1495 mutex_lock(&dev_priv->drm.struct_mutex);
1496 file_priv->bsd_engine = dev_priv->mm.bsd_engine_dispatch_index;
1497 dev_priv->mm.bsd_engine_dispatch_index ^= 1;
1498 mutex_unlock(&dev_priv->drm.struct_mutex);
1501 return file_priv->bsd_engine;
1504 #define I915_USER_RINGS (4)
1506 static const enum intel_engine_id user_ring_map[I915_USER_RINGS + 1] = {
1507 [I915_EXEC_DEFAULT] = RCS,
1508 [I915_EXEC_RENDER] = RCS,
1509 [I915_EXEC_BLT] = BCS,
1510 [I915_EXEC_BSD] = VCS,
1511 [I915_EXEC_VEBOX] = VECS
1514 static struct intel_engine_cs *
1515 eb_select_engine(struct drm_i915_private *dev_priv,
1516 struct drm_file *file,
1517 struct drm_i915_gem_execbuffer2 *args)
1519 unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
1520 struct intel_engine_cs *engine;
1522 if (user_ring_id > I915_USER_RINGS) {
1523 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id);
1527 if ((user_ring_id != I915_EXEC_BSD) &&
1528 ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1529 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1530 "bsd dispatch flags: %d\n", (int)(args->flags));
1534 if (user_ring_id == I915_EXEC_BSD && HAS_BSD2(dev_priv)) {
1535 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
1537 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
1538 bsd_idx = gen8_dispatch_bsd_engine(dev_priv, file);
1539 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
1540 bsd_idx <= I915_EXEC_BSD_RING2) {
1541 bsd_idx >>= I915_EXEC_BSD_SHIFT;
1544 DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
1549 engine = &dev_priv->engine[_VCS(bsd_idx)];
1551 engine = &dev_priv->engine[user_ring_map[user_ring_id]];
1554 if (!intel_engine_initialized(engine)) {
1555 DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id);
1563 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1564 struct drm_file *file,
1565 struct drm_i915_gem_execbuffer2 *args,
1566 struct drm_i915_gem_exec_object2 *exec)
1568 struct drm_i915_private *dev_priv = to_i915(dev);
1569 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1571 struct drm_i915_gem_exec_object2 shadow_exec_entry;
1572 struct intel_engine_cs *engine;
1573 struct i915_gem_context *ctx;
1574 struct i915_address_space *vm;
1575 struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1576 struct i915_execbuffer_params *params = ¶ms_master;
1577 const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1582 if (!i915_gem_check_execbuffer(args))
1585 ret = validate_exec_list(dev, exec, args->buffer_count);
1590 if (args->flags & I915_EXEC_SECURE) {
1591 if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
1594 dispatch_flags |= I915_DISPATCH_SECURE;
1596 if (args->flags & I915_EXEC_IS_PINNED)
1597 dispatch_flags |= I915_DISPATCH_PINNED;
1599 engine = eb_select_engine(dev_priv, file, args);
1603 if (args->buffer_count < 1) {
1604 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1608 if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1609 if (!HAS_RESOURCE_STREAMER(dev)) {
1610 DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1613 if (engine->id != RCS) {
1614 DRM_DEBUG("RS is not available on %s\n",
1619 dispatch_flags |= I915_DISPATCH_RS;
1622 /* Take a local wakeref for preparing to dispatch the execbuf as
1623 * we expect to access the hardware fairly frequently in the
1624 * process. Upon first dispatch, we acquire another prolonged
1625 * wakeref that we hold until the GPU has been idle for at least
1628 intel_runtime_pm_get(dev_priv);
1630 ret = i915_mutex_lock_interruptible(dev);
1634 ctx = i915_gem_validate_context(dev, file, engine, ctx_id);
1636 mutex_unlock(&dev->struct_mutex);
1641 i915_gem_context_get(ctx);
1644 vm = &ctx->ppgtt->base;
1648 memset(¶ms_master, 0x00, sizeof(params_master));
1650 eb = eb_create(args);
1652 i915_gem_context_put(ctx);
1653 mutex_unlock(&dev->struct_mutex);
1658 /* Look up object handles */
1659 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1663 /* take note of the batch buffer before we might reorder the lists */
1664 params->batch = eb_get_batch(eb);
1666 /* Move the objects en-masse into the GTT, evicting if necessary. */
1667 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1668 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1673 /* The objects are in their final locations, apply the relocations. */
1675 ret = i915_gem_execbuffer_relocate(eb);
1677 if (ret == -EFAULT) {
1678 ret = i915_gem_execbuffer_relocate_slow(dev, args, file,
1681 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1687 /* Set the pending read domains for the batch buffer to COMMAND */
1688 if (params->batch->obj->base.pending_write_domain) {
1689 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1694 params->args_batch_start_offset = args->batch_start_offset;
1695 if (intel_engine_needs_cmd_parser(engine) && args->batch_len) {
1696 struct i915_vma *vma;
1698 vma = i915_gem_execbuffer_parse(engine, &shadow_exec_entry,
1701 args->batch_start_offset,
1703 drm_is_current_master(file));
1711 * Batch parsed and accepted:
1713 * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1714 * bit from MI_BATCH_BUFFER_START commands issued in
1715 * the dispatch_execbuffer implementations. We
1716 * specifically don't want that set on batches the
1717 * command parser has accepted.
1719 dispatch_flags |= I915_DISPATCH_SECURE;
1720 params->args_batch_start_offset = 0;
1721 params->batch = vma;
1725 params->batch->obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1727 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1728 * batch" bit. Hence we need to pin secure batches into the global gtt.
1729 * hsw should have this fixed, but bdw mucks it up again. */
1730 if (dispatch_flags & I915_DISPATCH_SECURE) {
1731 struct drm_i915_gem_object *obj = params->batch->obj;
1732 struct i915_vma *vma;
1735 * So on first glance it looks freaky that we pin the batch here
1736 * outside of the reservation loop. But:
1737 * - The batch is already pinned into the relevant ppgtt, so we
1738 * already have the backing storage fully allocated.
1739 * - No other BO uses the global gtt (well contexts, but meh),
1740 * so we don't really have issues with multiple objects not
1741 * fitting due to fragmentation.
1742 * So this is actually safe.
1744 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, 0);
1750 params->batch = vma;
1753 /* Allocate a request for this batch buffer nice and early. */
1754 params->request = i915_gem_request_alloc(engine, ctx);
1755 if (IS_ERR(params->request)) {
1756 ret = PTR_ERR(params->request);
1757 goto err_batch_unpin;
1760 /* Whilst this request exists, batch_obj will be on the
1761 * active_list, and so will hold the active reference. Only when this
1762 * request is retired will the the batch_obj be moved onto the
1763 * inactive_list and lose its active reference. Hence we do not need
1764 * to explicitly hold another reference here.
1766 params->request->batch = params->batch;
1768 ret = i915_gem_request_add_to_client(params->request, file);
1773 * Save assorted stuff away to pass through to *_submission().
1774 * NB: This data should be 'persistent' and not local as it will
1775 * kept around beyond the duration of the IOCTL once the GPU
1776 * scheduler arrives.
1779 params->file = file;
1780 params->engine = engine;
1781 params->dispatch_flags = dispatch_flags;
1784 ret = execbuf_submit(params, args, &eb->vmas);
1786 __i915_add_request(params->request, ret == 0);
1790 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1791 * batch vma for correctness. For less ugly and less fragility this
1792 * needs to be adjusted to also track the ggtt batch vma properly as
1795 if (dispatch_flags & I915_DISPATCH_SECURE)
1796 i915_vma_unpin(params->batch);
1798 /* the request owns the ref now */
1799 i915_gem_context_put(ctx);
1802 mutex_unlock(&dev->struct_mutex);
1805 /* intel_gpu_busy should also get a ref, so it will free when the device
1806 * is really idle. */
1807 intel_runtime_pm_put(dev_priv);
1812 * Legacy execbuffer just creates an exec2 list from the original exec object
1813 * list array and passes it to the real function.
1816 i915_gem_execbuffer(struct drm_device *dev, void *data,
1817 struct drm_file *file)
1819 struct drm_i915_gem_execbuffer *args = data;
1820 struct drm_i915_gem_execbuffer2 exec2;
1821 struct drm_i915_gem_exec_object *exec_list = NULL;
1822 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1825 if (args->buffer_count < 1) {
1826 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1830 /* Copy in the exec list from userland */
1831 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1832 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1833 if (exec_list == NULL || exec2_list == NULL) {
1834 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1835 args->buffer_count);
1836 drm_free_large(exec_list);
1837 drm_free_large(exec2_list);
1840 ret = copy_from_user(exec_list,
1841 u64_to_user_ptr(args->buffers_ptr),
1842 sizeof(*exec_list) * args->buffer_count);
1844 DRM_DEBUG("copy %d exec entries failed %d\n",
1845 args->buffer_count, ret);
1846 drm_free_large(exec_list);
1847 drm_free_large(exec2_list);
1851 for (i = 0; i < args->buffer_count; i++) {
1852 exec2_list[i].handle = exec_list[i].handle;
1853 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1854 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1855 exec2_list[i].alignment = exec_list[i].alignment;
1856 exec2_list[i].offset = exec_list[i].offset;
1857 if (INTEL_INFO(dev)->gen < 4)
1858 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1860 exec2_list[i].flags = 0;
1863 exec2.buffers_ptr = args->buffers_ptr;
1864 exec2.buffer_count = args->buffer_count;
1865 exec2.batch_start_offset = args->batch_start_offset;
1866 exec2.batch_len = args->batch_len;
1867 exec2.DR1 = args->DR1;
1868 exec2.DR4 = args->DR4;
1869 exec2.num_cliprects = args->num_cliprects;
1870 exec2.cliprects_ptr = args->cliprects_ptr;
1871 exec2.flags = I915_EXEC_RENDER;
1872 i915_execbuffer2_set_context_id(exec2, 0);
1874 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1876 struct drm_i915_gem_exec_object __user *user_exec_list =
1877 u64_to_user_ptr(args->buffers_ptr);
1879 /* Copy the new buffer offsets back to the user's exec list. */
1880 for (i = 0; i < args->buffer_count; i++) {
1881 exec2_list[i].offset =
1882 gen8_canonical_addr(exec2_list[i].offset);
1883 ret = __copy_to_user(&user_exec_list[i].offset,
1884 &exec2_list[i].offset,
1885 sizeof(user_exec_list[i].offset));
1888 DRM_DEBUG("failed to copy %d exec entries "
1889 "back to user (%d)\n",
1890 args->buffer_count, ret);
1896 drm_free_large(exec_list);
1897 drm_free_large(exec2_list);
1902 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1903 struct drm_file *file)
1905 struct drm_i915_gem_execbuffer2 *args = data;
1906 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1909 if (args->buffer_count < 1 ||
1910 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1911 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1915 if (args->rsvd2 != 0) {
1916 DRM_DEBUG("dirty rvsd2 field\n");
1920 exec2_list = drm_malloc_gfp(args->buffer_count,
1921 sizeof(*exec2_list),
1923 if (exec2_list == NULL) {
1924 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1925 args->buffer_count);
1928 ret = copy_from_user(exec2_list,
1929 u64_to_user_ptr(args->buffers_ptr),
1930 sizeof(*exec2_list) * args->buffer_count);
1932 DRM_DEBUG("copy %d exec entries failed %d\n",
1933 args->buffer_count, ret);
1934 drm_free_large(exec2_list);
1938 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1940 /* Copy the new buffer offsets back to the user's exec list. */
1941 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1942 u64_to_user_ptr(args->buffers_ptr);
1945 for (i = 0; i < args->buffer_count; i++) {
1946 exec2_list[i].offset =
1947 gen8_canonical_addr(exec2_list[i].offset);
1948 ret = __copy_to_user(&user_exec_list[i].offset,
1949 &exec2_list[i].offset,
1950 sizeof(user_exec_list[i].offset));
1953 DRM_DEBUG("failed to copy %d exec entries "
1955 args->buffer_count);
1961 drm_free_large(exec2_list);