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/sync_file.h>
32 #include <linux/uaccess.h>
35 #include <drm/i915_drm.h>
38 #include "i915_gem_clflush.h"
39 #include "i915_trace.h"
40 #include "intel_drv.h"
41 #include "intel_frontbuffer.h"
43 #define DBG_USE_CPU_RELOC 0 /* -1 force GTT relocs; 1 force CPU relocs */
45 #define __EXEC_OBJECT_HAS_PIN (1<<31)
46 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
47 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
48 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
49 #define __EXEC_OBJECT_INTERNAL_FLAGS (0xf<<28) /* all of the above */
51 #define BATCH_OFFSET_BIAS (256*1024)
53 struct i915_execbuffer_params {
54 struct drm_device *dev;
55 struct drm_file *file;
56 struct i915_vma *batch;
58 u32 args_batch_start_offset;
59 struct intel_engine_cs *engine;
60 struct i915_gem_context *ctx;
61 struct drm_i915_gem_request *request;
65 struct drm_i915_private *i915;
66 struct list_head vmas;
69 struct i915_vma *lut[0];
70 struct hlist_head buckets[0];
74 static struct eb_vmas *
75 eb_create(struct drm_i915_private *i915,
76 struct drm_i915_gem_execbuffer2 *args)
78 struct eb_vmas *eb = NULL;
80 if (args->flags & I915_EXEC_HANDLE_LUT) {
81 unsigned size = args->buffer_count;
82 size *= sizeof(struct i915_vma *);
83 size += sizeof(struct eb_vmas);
84 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
88 unsigned size = args->buffer_count;
89 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
90 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
91 while (count > 2*size)
93 eb = kzalloc(count*sizeof(struct hlist_head) +
94 sizeof(struct eb_vmas),
101 eb->and = -args->buffer_count;
104 INIT_LIST_HEAD(&eb->vmas);
109 eb_reset(struct eb_vmas *eb)
112 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
115 static struct i915_vma *
116 eb_get_batch(struct eb_vmas *eb)
118 struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
121 * SNA is doing fancy tricks with compressing batch buffers, which leads
122 * to negative relocation deltas. Usually that works out ok since the
123 * relocate address is still positive, except when the batch is placed
124 * very low in the GTT. Ensure this doesn't happen.
126 * Note that actual hangs have only been observed on gen7, but for
127 * paranoia do it everywhere.
129 if ((vma->exec_entry->flags & EXEC_OBJECT_PINNED) == 0)
130 vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
136 eb_lookup_vmas(struct eb_vmas *eb,
137 struct drm_i915_gem_exec_object2 *exec,
138 const struct drm_i915_gem_execbuffer2 *args,
139 struct i915_address_space *vm,
140 struct drm_file *file)
142 struct drm_i915_gem_object *obj;
143 struct list_head objects;
146 INIT_LIST_HEAD(&objects);
147 spin_lock(&file->table_lock);
148 /* Grab a reference to the object and release the lock so we can lookup
149 * or create the VMA without using GFP_ATOMIC */
150 for (i = 0; i < args->buffer_count; i++) {
151 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
153 spin_unlock(&file->table_lock);
154 DRM_DEBUG("Invalid object handle %d at index %d\n",
160 if (!list_empty(&obj->obj_exec_link)) {
161 spin_unlock(&file->table_lock);
162 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
163 obj, exec[i].handle, i);
168 i915_gem_object_get(obj);
169 list_add_tail(&obj->obj_exec_link, &objects);
171 spin_unlock(&file->table_lock);
174 while (!list_empty(&objects)) {
175 struct i915_vma *vma;
177 obj = list_first_entry(&objects,
178 struct drm_i915_gem_object,
182 * NOTE: We can leak any vmas created here when something fails
183 * later on. But that's no issue since vma_unbind can deal with
184 * vmas which are not actually bound. And since only
185 * lookup_or_create exists as an interface to get at the vma
186 * from the (obj, vm) we don't run the risk of creating
187 * duplicated vmas for the same vm.
189 vma = i915_vma_instance(obj, vm, NULL);
190 if (unlikely(IS_ERR(vma))) {
191 DRM_DEBUG("Failed to lookup VMA\n");
196 /* Transfer ownership from the objects list to the vmas list. */
197 list_add_tail(&vma->exec_list, &eb->vmas);
198 list_del_init(&obj->obj_exec_link);
200 vma->exec_entry = &exec[i];
204 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
205 vma->exec_handle = handle;
206 hlist_add_head(&vma->exec_node,
207 &eb->buckets[handle & eb->and]);
216 while (!list_empty(&objects)) {
217 obj = list_first_entry(&objects,
218 struct drm_i915_gem_object,
220 list_del_init(&obj->obj_exec_link);
221 i915_gem_object_put(obj);
224 * Objects already transfered to the vmas list will be unreferenced by
231 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
234 if (handle >= -eb->and)
236 return eb->lut[handle];
238 struct hlist_head *head;
239 struct i915_vma *vma;
241 head = &eb->buckets[handle & eb->and];
242 hlist_for_each_entry(vma, head, exec_node) {
243 if (vma->exec_handle == handle)
251 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
253 struct drm_i915_gem_exec_object2 *entry;
255 if (!drm_mm_node_allocated(&vma->node))
258 entry = vma->exec_entry;
260 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
261 i915_vma_unpin_fence(vma);
263 if (entry->flags & __EXEC_OBJECT_HAS_PIN)
264 __i915_vma_unpin(vma);
266 entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
269 static void eb_destroy(struct eb_vmas *eb)
271 while (!list_empty(&eb->vmas)) {
272 struct i915_vma *vma;
274 vma = list_first_entry(&eb->vmas,
277 list_del_init(&vma->exec_list);
278 i915_gem_execbuffer_unreserve_vma(vma);
279 vma->exec_entry = NULL;
285 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
287 if (!i915_gem_object_has_struct_page(obj))
290 if (DBG_USE_CPU_RELOC)
291 return DBG_USE_CPU_RELOC > 0;
293 return (HAS_LLC(to_i915(obj->base.dev)) ||
294 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
295 obj->cache_level != I915_CACHE_NONE);
298 /* Used to convert any address to canonical form.
299 * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
300 * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
301 * addresses to be in a canonical form:
302 * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
303 * canonical form [63:48] == [47]."
305 #define GEN8_HIGH_ADDRESS_BIT 47
306 static inline uint64_t gen8_canonical_addr(uint64_t address)
308 return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
311 static inline uint64_t gen8_noncanonical_addr(uint64_t address)
313 return address & ((1ULL << (GEN8_HIGH_ADDRESS_BIT + 1)) - 1);
316 static inline uint64_t
317 relocation_target(const struct drm_i915_gem_relocation_entry *reloc,
318 uint64_t target_offset)
320 return gen8_canonical_addr((int)reloc->delta + target_offset);
324 struct drm_i915_private *i915;
325 struct drm_mm_node node;
328 bool use_64bit_reloc;
331 static void reloc_cache_init(struct reloc_cache *cache,
332 struct drm_i915_private *i915)
337 /* Must be a variable in the struct to allow GCC to unroll. */
338 cache->use_64bit_reloc = HAS_64BIT_RELOC(i915);
339 cache->node.allocated = false;
342 static inline void *unmask_page(unsigned long p)
344 return (void *)(uintptr_t)(p & PAGE_MASK);
347 static inline unsigned int unmask_flags(unsigned long p)
349 return p & ~PAGE_MASK;
352 #define KMAP 0x4 /* after CLFLUSH_FLAGS */
354 static void reloc_cache_fini(struct reloc_cache *cache)
361 vaddr = unmask_page(cache->vaddr);
362 if (cache->vaddr & KMAP) {
363 if (cache->vaddr & CLFLUSH_AFTER)
366 kunmap_atomic(vaddr);
367 i915_gem_obj_finish_shmem_access((struct drm_i915_gem_object *)cache->node.mm);
370 io_mapping_unmap_atomic((void __iomem *)vaddr);
371 if (cache->node.allocated) {
372 struct i915_ggtt *ggtt = &cache->i915->ggtt;
374 ggtt->base.clear_range(&ggtt->base,
377 drm_mm_remove_node(&cache->node);
379 i915_vma_unpin((struct i915_vma *)cache->node.mm);
384 static void *reloc_kmap(struct drm_i915_gem_object *obj,
385 struct reloc_cache *cache,
391 kunmap_atomic(unmask_page(cache->vaddr));
393 unsigned int flushes;
396 ret = i915_gem_obj_prepare_shmem_write(obj, &flushes);
400 BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS);
401 BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK);
403 cache->vaddr = flushes | KMAP;
404 cache->node.mm = (void *)obj;
409 vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj, page));
410 cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr;
416 static void *reloc_iomap(struct drm_i915_gem_object *obj,
417 struct reloc_cache *cache,
420 struct i915_ggtt *ggtt = &cache->i915->ggtt;
421 unsigned long offset;
425 io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr));
427 struct i915_vma *vma;
430 if (use_cpu_reloc(obj))
433 ret = i915_gem_object_set_to_gtt_domain(obj, true);
437 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
438 PIN_MAPPABLE | PIN_NONBLOCK);
440 memset(&cache->node, 0, sizeof(cache->node));
441 ret = drm_mm_insert_node_in_range
442 (&ggtt->base.mm, &cache->node,
443 PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE,
444 0, ggtt->mappable_end,
446 if (ret) /* no inactive aperture space, use cpu reloc */
449 ret = i915_vma_put_fence(vma);
455 cache->node.start = vma->node.start;
456 cache->node.mm = (void *)vma;
460 offset = cache->node.start;
461 if (cache->node.allocated) {
463 ggtt->base.insert_page(&ggtt->base,
464 i915_gem_object_get_dma_address(obj, page),
465 offset, I915_CACHE_NONE, 0);
467 offset += page << PAGE_SHIFT;
470 vaddr = (void __force *) io_mapping_map_atomic_wc(&cache->i915->ggtt.mappable, offset);
472 cache->vaddr = (unsigned long)vaddr;
477 static void *reloc_vaddr(struct drm_i915_gem_object *obj,
478 struct reloc_cache *cache,
483 if (cache->page == page) {
484 vaddr = unmask_page(cache->vaddr);
487 if ((cache->vaddr & KMAP) == 0)
488 vaddr = reloc_iomap(obj, cache, page);
490 vaddr = reloc_kmap(obj, cache, page);
496 static void clflush_write32(u32 *addr, u32 value, unsigned int flushes)
498 if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) {
499 if (flushes & CLFLUSH_BEFORE) {
506 /* Writes to the same cacheline are serialised by the CPU
507 * (including clflush). On the write path, we only require
508 * that it hits memory in an orderly fashion and place
509 * mb barriers at the start and end of the relocation phase
510 * to ensure ordering of clflush wrt to the system.
512 if (flushes & CLFLUSH_AFTER)
519 relocate_entry(struct drm_i915_gem_object *obj,
520 const struct drm_i915_gem_relocation_entry *reloc,
521 struct reloc_cache *cache,
524 u64 offset = reloc->offset;
525 bool wide = cache->use_64bit_reloc;
528 target_offset = relocation_target(reloc, target_offset);
530 vaddr = reloc_vaddr(obj, cache, offset >> PAGE_SHIFT);
532 return PTR_ERR(vaddr);
534 clflush_write32(vaddr + offset_in_page(offset),
535 lower_32_bits(target_offset),
539 offset += sizeof(u32);
540 target_offset >>= 32;
549 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
551 struct drm_i915_gem_relocation_entry *reloc,
552 struct reloc_cache *cache)
554 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
555 struct drm_gem_object *target_obj;
556 struct drm_i915_gem_object *target_i915_obj;
557 struct i915_vma *target_vma;
558 uint64_t target_offset;
561 /* we've already hold a reference to all valid objects */
562 target_vma = eb_get_vma(eb, reloc->target_handle);
563 if (unlikely(target_vma == NULL))
565 target_i915_obj = target_vma->obj;
566 target_obj = &target_vma->obj->base;
568 target_offset = gen8_canonical_addr(target_vma->node.start);
570 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
571 * pipe_control writes because the gpu doesn't properly redirect them
572 * through the ppgtt for non_secure batchbuffers. */
573 if (unlikely(IS_GEN6(dev_priv) &&
574 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
575 ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
577 if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
581 /* Validate that the target is in a valid r/w GPU domain */
582 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
583 DRM_DEBUG("reloc with multiple write domains: "
584 "obj %p target %d offset %d "
585 "read %08x write %08x",
586 obj, reloc->target_handle,
589 reloc->write_domain);
592 if (unlikely((reloc->write_domain | reloc->read_domains)
593 & ~I915_GEM_GPU_DOMAINS)) {
594 DRM_DEBUG("reloc with read/write non-GPU domains: "
595 "obj %p target %d offset %d "
596 "read %08x write %08x",
597 obj, reloc->target_handle,
600 reloc->write_domain);
604 target_obj->pending_read_domains |= reloc->read_domains;
605 target_obj->pending_write_domain |= reloc->write_domain;
607 /* If the relocation already has the right value in it, no
608 * more work needs to be done.
610 if (target_offset == reloc->presumed_offset)
613 /* Check that the relocation address is valid... */
614 if (unlikely(reloc->offset >
615 obj->base.size - (cache->use_64bit_reloc ? 8 : 4))) {
616 DRM_DEBUG("Relocation beyond object bounds: "
617 "obj %p target %d offset %d size %d.\n",
618 obj, reloc->target_handle,
620 (int) obj->base.size);
623 if (unlikely(reloc->offset & 3)) {
624 DRM_DEBUG("Relocation not 4-byte aligned: "
625 "obj %p target %d offset %d.\n",
626 obj, reloc->target_handle,
627 (int) reloc->offset);
631 ret = relocate_entry(obj, reloc, cache, target_offset);
635 /* and update the user's relocation entry */
636 reloc->presumed_offset = target_offset;
641 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
644 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
645 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
646 struct drm_i915_gem_relocation_entry __user *user_relocs;
647 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
648 struct reloc_cache cache;
651 user_relocs = u64_to_user_ptr(entry->relocs_ptr);
652 reloc_cache_init(&cache, eb->i915);
654 remain = entry->relocation_count;
656 struct drm_i915_gem_relocation_entry *r = stack_reloc;
657 unsigned long unwritten;
660 count = min_t(unsigned int, remain, ARRAY_SIZE(stack_reloc));
663 /* This is the fast path and we cannot handle a pagefault
664 * whilst holding the struct mutex lest the user pass in the
665 * relocations contained within a mmaped bo. For in such a case
666 * we, the page fault handler would call i915_gem_fault() and
667 * we would try to acquire the struct mutex again. Obviously
668 * this is bad and so lockdep complains vehemently.
671 unwritten = __copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0]));
673 if (unlikely(unwritten)) {
679 u64 offset = r->presumed_offset;
681 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r, &cache);
685 if (r->presumed_offset != offset) {
687 unwritten = __put_user(r->presumed_offset,
688 &user_relocs->presumed_offset);
690 if (unlikely(unwritten)) {
691 /* Note that reporting an error now
692 * leaves everything in an inconsistent
693 * state as we have *already* changed
694 * the relocation value inside the
695 * object. As we have not changed the
696 * reloc.presumed_offset or will not
697 * change the execobject.offset, on the
698 * call we may not rewrite the value
699 * inside the object, leaving it
700 * dangling and causing a GPU hang.
713 reloc_cache_fini(&cache);
719 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
721 struct drm_i915_gem_relocation_entry *relocs)
723 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
724 struct reloc_cache cache;
727 reloc_cache_init(&cache, eb->i915);
728 for (i = 0; i < entry->relocation_count; i++) {
729 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i], &cache);
733 reloc_cache_fini(&cache);
739 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
741 struct i915_vma *vma;
744 list_for_each_entry(vma, &eb->vmas, exec_list) {
745 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
753 static bool only_mappable_for_reloc(unsigned int flags)
755 return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
756 __EXEC_OBJECT_NEEDS_MAP;
760 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
761 struct intel_engine_cs *engine,
764 struct drm_i915_gem_object *obj = vma->obj;
765 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
770 if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
773 if (!drm_mm_node_allocated(&vma->node)) {
774 /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
775 * limit address to the first 4GBs for unflagged objects.
777 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
778 flags |= PIN_ZONE_4G;
779 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
780 flags |= PIN_GLOBAL | PIN_MAPPABLE;
781 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
782 flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
783 if (entry->flags & EXEC_OBJECT_PINNED)
784 flags |= entry->offset | PIN_OFFSET_FIXED;
785 if ((flags & PIN_MAPPABLE) == 0)
789 ret = i915_vma_pin(vma,
793 if ((ret == -ENOSPC || ret == -E2BIG) &&
794 only_mappable_for_reloc(entry->flags))
795 ret = i915_vma_pin(vma,
798 flags & ~PIN_MAPPABLE);
802 entry->flags |= __EXEC_OBJECT_HAS_PIN;
804 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
805 ret = i915_vma_get_fence(vma);
809 if (i915_vma_pin_fence(vma))
810 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
813 if (entry->offset != vma->node.start) {
814 entry->offset = vma->node.start;
818 if (entry->flags & EXEC_OBJECT_WRITE) {
819 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
820 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
827 need_reloc_mappable(struct i915_vma *vma)
829 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
831 if (entry->relocation_count == 0)
834 if (!i915_vma_is_ggtt(vma))
837 /* See also use_cpu_reloc() */
838 if (HAS_LLC(to_i915(vma->obj->base.dev)))
841 if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
848 eb_vma_misplaced(struct i915_vma *vma)
850 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
852 WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
853 !i915_vma_is_ggtt(vma));
855 if (entry->alignment && !IS_ALIGNED(vma->node.start, entry->alignment))
858 if (vma->node.size < entry->pad_to_size)
861 if (entry->flags & EXEC_OBJECT_PINNED &&
862 vma->node.start != entry->offset)
865 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
866 vma->node.start < BATCH_OFFSET_BIAS)
869 /* avoid costly ping-pong once a batch bo ended up non-mappable */
870 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
871 !i915_vma_is_map_and_fenceable(vma))
872 return !only_mappable_for_reloc(entry->flags);
874 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
875 (vma->node.start + vma->node.size - 1) >> 32)
882 i915_gem_execbuffer_reserve(struct intel_engine_cs *engine,
883 struct list_head *vmas,
884 struct i915_gem_context *ctx,
887 struct drm_i915_gem_object *obj;
888 struct i915_vma *vma;
889 struct i915_address_space *vm;
890 struct list_head ordered_vmas;
891 struct list_head pinned_vmas;
892 bool has_fenced_gpu_access = INTEL_GEN(engine->i915) < 4;
895 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
897 INIT_LIST_HEAD(&ordered_vmas);
898 INIT_LIST_HEAD(&pinned_vmas);
899 while (!list_empty(vmas)) {
900 struct drm_i915_gem_exec_object2 *entry;
901 bool need_fence, need_mappable;
903 vma = list_first_entry(vmas, struct i915_vma, exec_list);
905 entry = vma->exec_entry;
907 if (ctx->flags & CONTEXT_NO_ZEROMAP)
908 entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
910 if (!has_fenced_gpu_access)
911 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
913 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
914 i915_gem_object_is_tiled(obj);
915 need_mappable = need_fence || need_reloc_mappable(vma);
917 if (entry->flags & EXEC_OBJECT_PINNED)
918 list_move_tail(&vma->exec_list, &pinned_vmas);
919 else if (need_mappable) {
920 entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
921 list_move(&vma->exec_list, &ordered_vmas);
923 list_move_tail(&vma->exec_list, &ordered_vmas);
925 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
926 obj->base.pending_write_domain = 0;
928 list_splice(&ordered_vmas, vmas);
929 list_splice(&pinned_vmas, vmas);
931 /* Attempt to pin all of the buffers into the GTT.
932 * This is done in 3 phases:
934 * 1a. Unbind all objects that do not match the GTT constraints for
935 * the execbuffer (fenceable, mappable, alignment etc).
936 * 1b. Increment pin count for already bound objects.
937 * 2. Bind new objects.
938 * 3. Decrement pin count.
940 * This avoid unnecessary unbinding of later objects in order to make
941 * room for the earlier objects *unless* we need to defragment.
947 /* Unbind any ill-fitting objects or pin. */
948 list_for_each_entry(vma, vmas, exec_list) {
949 if (!drm_mm_node_allocated(&vma->node))
952 if (eb_vma_misplaced(vma))
953 ret = i915_vma_unbind(vma);
955 ret = i915_gem_execbuffer_reserve_vma(vma,
962 /* Bind fresh objects */
963 list_for_each_entry(vma, vmas, exec_list) {
964 if (drm_mm_node_allocated(&vma->node))
967 ret = i915_gem_execbuffer_reserve_vma(vma, engine,
974 if (ret != -ENOSPC || retry++)
977 /* Decrement pin count for bound objects */
978 list_for_each_entry(vma, vmas, exec_list)
979 i915_gem_execbuffer_unreserve_vma(vma);
981 ret = i915_gem_evict_vm(vm, true);
988 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
989 struct drm_i915_gem_execbuffer2 *args,
990 struct drm_file *file,
991 struct intel_engine_cs *engine,
993 struct drm_i915_gem_exec_object2 *exec,
994 struct i915_gem_context *ctx)
996 struct drm_i915_gem_relocation_entry *reloc;
997 struct i915_address_space *vm;
998 struct i915_vma *vma;
1002 unsigned count = args->buffer_count;
1004 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
1006 /* We may process another execbuffer during the unlock... */
1007 while (!list_empty(&eb->vmas)) {
1008 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
1009 list_del_init(&vma->exec_list);
1010 i915_gem_execbuffer_unreserve_vma(vma);
1014 mutex_unlock(&dev->struct_mutex);
1017 for (i = 0; i < count; i++)
1018 total += exec[i].relocation_count;
1020 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
1021 reloc = drm_malloc_ab(total, sizeof(*reloc));
1022 if (reloc == NULL || reloc_offset == NULL) {
1023 drm_free_large(reloc);
1024 drm_free_large(reloc_offset);
1025 mutex_lock(&dev->struct_mutex);
1030 for (i = 0; i < count; i++) {
1031 struct drm_i915_gem_relocation_entry __user *user_relocs;
1032 u64 invalid_offset = (u64)-1;
1035 user_relocs = u64_to_user_ptr(exec[i].relocs_ptr);
1037 if (copy_from_user(reloc+total, user_relocs,
1038 exec[i].relocation_count * sizeof(*reloc))) {
1040 mutex_lock(&dev->struct_mutex);
1044 /* As we do not update the known relocation offsets after
1045 * relocating (due to the complexities in lock handling),
1046 * we need to mark them as invalid now so that we force the
1047 * relocation processing next time. Just in case the target
1048 * object is evicted and then rebound into its old
1049 * presumed_offset before the next execbuffer - if that
1050 * happened we would make the mistake of assuming that the
1051 * relocations were valid.
1053 for (j = 0; j < exec[i].relocation_count; j++) {
1054 if (__copy_to_user(&user_relocs[j].presumed_offset,
1056 sizeof(invalid_offset))) {
1058 mutex_lock(&dev->struct_mutex);
1063 reloc_offset[i] = total;
1064 total += exec[i].relocation_count;
1067 ret = i915_mutex_lock_interruptible(dev);
1069 mutex_lock(&dev->struct_mutex);
1073 /* reacquire the objects */
1075 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1079 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1080 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1085 list_for_each_entry(vma, &eb->vmas, exec_list) {
1086 int offset = vma->exec_entry - exec;
1087 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
1088 reloc + reloc_offset[offset]);
1093 /* Leave the user relocations as are, this is the painfully slow path,
1094 * and we want to avoid the complication of dropping the lock whilst
1095 * having buffers reserved in the aperture and so causing spurious
1096 * ENOSPC for random operations.
1100 drm_free_large(reloc);
1101 drm_free_large(reloc_offset);
1106 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
1107 struct list_head *vmas)
1109 struct i915_vma *vma;
1112 list_for_each_entry(vma, vmas, exec_list) {
1113 struct drm_i915_gem_object *obj = vma->obj;
1115 if (vma->exec_entry->flags & EXEC_OBJECT_ASYNC)
1118 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU) {
1119 i915_gem_clflush_object(obj, 0);
1120 obj->base.write_domain = 0;
1123 ret = i915_gem_request_await_object
1124 (req, obj, obj->base.pending_write_domain);
1129 /* Unconditionally flush any chipset caches (for streaming writes). */
1130 i915_gem_chipset_flush(req->engine->i915);
1132 /* Unconditionally invalidate GPU caches and TLBs. */
1133 return req->engine->emit_flush(req, EMIT_INVALIDATE);
1137 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
1139 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
1142 /* Kernel clipping was a DRI1 misfeature */
1143 if (exec->num_cliprects || exec->cliprects_ptr)
1146 if (exec->DR4 == 0xffffffff) {
1147 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1150 if (exec->DR1 || exec->DR4)
1153 if ((exec->batch_start_offset | exec->batch_len) & 0x7)
1160 validate_exec_list(struct drm_device *dev,
1161 struct drm_i915_gem_exec_object2 *exec,
1164 unsigned relocs_total = 0;
1165 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
1166 unsigned invalid_flags;
1169 /* INTERNAL flags must not overlap with external ones */
1170 BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & ~__EXEC_OBJECT_UNKNOWN_FLAGS);
1172 invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
1173 if (USES_FULL_PPGTT(dev))
1174 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
1176 for (i = 0; i < count; i++) {
1177 char __user *ptr = u64_to_user_ptr(exec[i].relocs_ptr);
1178 int length; /* limited by fault_in_pages_readable() */
1180 if (exec[i].flags & invalid_flags)
1183 /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
1184 * any non-page-aligned or non-canonical addresses.
1186 if (exec[i].flags & EXEC_OBJECT_PINNED) {
1187 if (exec[i].offset !=
1188 gen8_canonical_addr(exec[i].offset & PAGE_MASK))
1192 /* From drm_mm perspective address space is continuous,
1193 * so from this point we're always using non-canonical
1196 exec[i].offset = gen8_noncanonical_addr(exec[i].offset);
1198 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
1201 /* pad_to_size was once a reserved field, so sanitize it */
1202 if (exec[i].flags & EXEC_OBJECT_PAD_TO_SIZE) {
1203 if (offset_in_page(exec[i].pad_to_size))
1206 exec[i].pad_to_size = 0;
1209 /* First check for malicious input causing overflow in
1210 * the worst case where we need to allocate the entire
1211 * relocation tree as a single array.
1213 if (exec[i].relocation_count > relocs_max - relocs_total)
1215 relocs_total += exec[i].relocation_count;
1217 length = exec[i].relocation_count *
1218 sizeof(struct drm_i915_gem_relocation_entry);
1220 * We must check that the entire relocation array is safe
1221 * to read, but since we may need to update the presumed
1222 * offsets during execution, check for full write access.
1224 if (!access_ok(VERIFY_WRITE, ptr, length))
1227 if (likely(!i915.prefault_disable)) {
1228 if (fault_in_pages_readable(ptr, length))
1236 static struct i915_gem_context *
1237 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
1238 struct intel_engine_cs *engine, const u32 ctx_id)
1240 struct i915_gem_context *ctx;
1242 ctx = i915_gem_context_lookup(file->driver_priv, ctx_id);
1246 if (i915_gem_context_is_banned(ctx)) {
1247 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1248 return ERR_PTR(-EIO);
1254 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
1256 return !(obj->cache_level == I915_CACHE_NONE ||
1257 obj->cache_level == I915_CACHE_WT);
1260 void i915_vma_move_to_active(struct i915_vma *vma,
1261 struct drm_i915_gem_request *req,
1264 struct drm_i915_gem_object *obj = vma->obj;
1265 const unsigned int idx = req->engine->id;
1267 lockdep_assert_held(&req->i915->drm.struct_mutex);
1268 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1270 /* Add a reference if we're newly entering the active list.
1271 * The order in which we add operations to the retirement queue is
1272 * vital here: mark_active adds to the start of the callback list,
1273 * such that subsequent callbacks are called first. Therefore we
1274 * add the active reference first and queue for it to be dropped
1277 if (!i915_vma_is_active(vma))
1278 obj->active_count++;
1279 i915_vma_set_active(vma, idx);
1280 i915_gem_active_set(&vma->last_read[idx], req);
1281 list_move_tail(&vma->vm_link, &vma->vm->active_list);
1283 if (flags & EXEC_OBJECT_WRITE) {
1284 if (intel_fb_obj_invalidate(obj, ORIGIN_CS))
1285 i915_gem_active_set(&obj->frontbuffer_write, req);
1287 /* update for the implicit flush after a batch */
1288 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1289 if (!obj->cache_dirty && gpu_write_needs_clflush(obj))
1290 obj->cache_dirty = true;
1293 if (flags & EXEC_OBJECT_NEEDS_FENCE)
1294 i915_gem_active_set(&vma->last_fence, req);
1297 static void eb_export_fence(struct drm_i915_gem_object *obj,
1298 struct drm_i915_gem_request *req,
1301 struct reservation_object *resv = obj->resv;
1303 /* Ignore errors from failing to allocate the new fence, we can't
1304 * handle an error right now. Worst case should be missed
1305 * synchronisation leading to rendering corruption.
1307 reservation_object_lock(resv, NULL);
1308 if (flags & EXEC_OBJECT_WRITE)
1309 reservation_object_add_excl_fence(resv, &req->fence);
1310 else if (reservation_object_reserve_shared(resv) == 0)
1311 reservation_object_add_shared_fence(resv, &req->fence);
1312 reservation_object_unlock(resv);
1316 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1317 struct drm_i915_gem_request *req)
1319 struct i915_vma *vma;
1321 list_for_each_entry(vma, vmas, exec_list) {
1322 struct drm_i915_gem_object *obj = vma->obj;
1324 obj->base.write_domain = obj->base.pending_write_domain;
1325 if (obj->base.write_domain)
1326 vma->exec_entry->flags |= EXEC_OBJECT_WRITE;
1328 obj->base.pending_read_domains |= obj->base.read_domains;
1329 obj->base.read_domains = obj->base.pending_read_domains;
1331 i915_vma_move_to_active(vma, req, vma->exec_entry->flags);
1332 eb_export_fence(obj, req, vma->exec_entry->flags);
1337 i915_reset_gen7_sol_offsets(struct drm_i915_gem_request *req)
1342 if (!IS_GEN7(req->i915) || req->engine->id != RCS) {
1343 DRM_DEBUG("sol reset is gen7/rcs only\n");
1347 cs = intel_ring_begin(req, 4 * 3);
1351 for (i = 0; i < 4; i++) {
1352 *cs++ = MI_LOAD_REGISTER_IMM(1);
1353 *cs++ = i915_mmio_reg_offset(GEN7_SO_WRITE_OFFSET(i));
1357 intel_ring_advance(req, cs);
1362 static struct i915_vma *
1363 i915_gem_execbuffer_parse(struct intel_engine_cs *engine,
1364 struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1365 struct drm_i915_gem_object *batch_obj,
1367 u32 batch_start_offset,
1371 struct drm_i915_gem_object *shadow_batch_obj;
1372 struct i915_vma *vma;
1375 shadow_batch_obj = i915_gem_batch_pool_get(&engine->batch_pool,
1376 PAGE_ALIGN(batch_len));
1377 if (IS_ERR(shadow_batch_obj))
1378 return ERR_CAST(shadow_batch_obj);
1380 ret = intel_engine_cmd_parser(engine,
1387 if (ret == -EACCES) /* unhandled chained batch */
1394 vma = i915_gem_object_ggtt_pin(shadow_batch_obj, NULL, 0, 0, 0);
1398 memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1400 vma->exec_entry = shadow_exec_entry;
1401 vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1402 i915_gem_object_get(shadow_batch_obj);
1403 list_add_tail(&vma->exec_list, &eb->vmas);
1406 i915_gem_object_unpin_pages(shadow_batch_obj);
1411 add_to_client(struct drm_i915_gem_request *req,
1412 struct drm_file *file)
1414 req->file_priv = file->driver_priv;
1415 list_add_tail(&req->client_link, &req->file_priv->mm.request_list);
1419 execbuf_submit(struct i915_execbuffer_params *params,
1420 struct drm_i915_gem_execbuffer2 *args,
1421 struct list_head *vmas)
1423 u64 exec_start, exec_len;
1426 ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1430 ret = i915_switch_context(params->request);
1434 if (args->flags & I915_EXEC_CONSTANTS_MASK) {
1435 DRM_DEBUG("I915_EXEC_CONSTANTS_* unsupported\n");
1439 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1440 ret = i915_reset_gen7_sol_offsets(params->request);
1445 exec_len = args->batch_len;
1446 exec_start = params->batch->node.start +
1447 params->args_batch_start_offset;
1450 exec_len = params->batch->size - params->args_batch_start_offset;
1452 ret = params->engine->emit_bb_start(params->request,
1453 exec_start, exec_len,
1454 params->dispatch_flags);
1458 i915_gem_execbuffer_move_to_active(vmas, params->request);
1464 * Find one BSD ring to dispatch the corresponding BSD command.
1465 * The engine index is returned.
1468 gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
1469 struct drm_file *file)
1471 struct drm_i915_file_private *file_priv = file->driver_priv;
1473 /* Check whether the file_priv has already selected one ring. */
1474 if ((int)file_priv->bsd_engine < 0)
1475 file_priv->bsd_engine = atomic_fetch_xor(1,
1476 &dev_priv->mm.bsd_engine_dispatch_index);
1478 return file_priv->bsd_engine;
1481 #define I915_USER_RINGS (4)
1483 static const enum intel_engine_id user_ring_map[I915_USER_RINGS + 1] = {
1484 [I915_EXEC_DEFAULT] = RCS,
1485 [I915_EXEC_RENDER] = RCS,
1486 [I915_EXEC_BLT] = BCS,
1487 [I915_EXEC_BSD] = VCS,
1488 [I915_EXEC_VEBOX] = VECS
1491 static struct intel_engine_cs *
1492 eb_select_engine(struct drm_i915_private *dev_priv,
1493 struct drm_file *file,
1494 struct drm_i915_gem_execbuffer2 *args)
1496 unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
1497 struct intel_engine_cs *engine;
1499 if (user_ring_id > I915_USER_RINGS) {
1500 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id);
1504 if ((user_ring_id != I915_EXEC_BSD) &&
1505 ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1506 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1507 "bsd dispatch flags: %d\n", (int)(args->flags));
1511 if (user_ring_id == I915_EXEC_BSD && HAS_BSD2(dev_priv)) {
1512 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
1514 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
1515 bsd_idx = gen8_dispatch_bsd_engine(dev_priv, file);
1516 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
1517 bsd_idx <= I915_EXEC_BSD_RING2) {
1518 bsd_idx >>= I915_EXEC_BSD_SHIFT;
1521 DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
1526 engine = dev_priv->engine[_VCS(bsd_idx)];
1528 engine = dev_priv->engine[user_ring_map[user_ring_id]];
1532 DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id);
1540 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1541 struct drm_file *file,
1542 struct drm_i915_gem_execbuffer2 *args,
1543 struct drm_i915_gem_exec_object2 *exec)
1545 struct drm_i915_private *dev_priv = to_i915(dev);
1546 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1548 struct drm_i915_gem_exec_object2 shadow_exec_entry;
1549 struct intel_engine_cs *engine;
1550 struct i915_gem_context *ctx;
1551 struct i915_address_space *vm;
1552 struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1553 struct i915_execbuffer_params *params = ¶ms_master;
1554 const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1556 struct dma_fence *in_fence = NULL;
1557 struct sync_file *out_fence = NULL;
1558 int out_fence_fd = -1;
1562 if (!i915_gem_check_execbuffer(args))
1565 ret = validate_exec_list(dev, exec, args->buffer_count);
1570 if (args->flags & I915_EXEC_SECURE) {
1571 if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
1574 dispatch_flags |= I915_DISPATCH_SECURE;
1576 if (args->flags & I915_EXEC_IS_PINNED)
1577 dispatch_flags |= I915_DISPATCH_PINNED;
1579 engine = eb_select_engine(dev_priv, file, args);
1583 if (args->buffer_count < 1) {
1584 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1588 if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1589 if (!HAS_RESOURCE_STREAMER(dev_priv)) {
1590 DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1593 if (engine->id != RCS) {
1594 DRM_DEBUG("RS is not available on %s\n",
1599 dispatch_flags |= I915_DISPATCH_RS;
1602 if (args->flags & I915_EXEC_FENCE_IN) {
1603 in_fence = sync_file_get_fence(lower_32_bits(args->rsvd2));
1608 if (args->flags & I915_EXEC_FENCE_OUT) {
1609 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
1610 if (out_fence_fd < 0) {
1616 /* Take a local wakeref for preparing to dispatch the execbuf as
1617 * we expect to access the hardware fairly frequently in the
1618 * process. Upon first dispatch, we acquire another prolonged
1619 * wakeref that we hold until the GPU has been idle for at least
1622 intel_runtime_pm_get(dev_priv);
1624 ret = i915_mutex_lock_interruptible(dev);
1628 ctx = i915_gem_validate_context(dev, file, engine, ctx_id);
1630 mutex_unlock(&dev->struct_mutex);
1635 i915_gem_context_get(ctx);
1638 vm = &ctx->ppgtt->base;
1642 memset(¶ms_master, 0x00, sizeof(params_master));
1644 eb = eb_create(dev_priv, args);
1646 i915_gem_context_put(ctx);
1647 mutex_unlock(&dev->struct_mutex);
1652 /* Look up object handles */
1653 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1657 /* take note of the batch buffer before we might reorder the lists */
1658 params->batch = eb_get_batch(eb);
1660 /* Move the objects en-masse into the GTT, evicting if necessary. */
1661 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1662 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1667 /* The objects are in their final locations, apply the relocations. */
1669 ret = i915_gem_execbuffer_relocate(eb);
1671 if (ret == -EFAULT) {
1672 ret = i915_gem_execbuffer_relocate_slow(dev, args, file,
1675 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1681 /* Set the pending read domains for the batch buffer to COMMAND */
1682 if (params->batch->obj->base.pending_write_domain) {
1683 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1687 if (args->batch_start_offset > params->batch->size ||
1688 args->batch_len > params->batch->size - args->batch_start_offset) {
1689 DRM_DEBUG("Attempting to use out-of-bounds batch\n");
1694 params->args_batch_start_offset = args->batch_start_offset;
1695 if (engine->needs_cmd_parser && 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;
1761 ret = i915_gem_request_await_dma_fence(params->request,
1767 if (out_fence_fd != -1) {
1768 out_fence = sync_file_create(¶ms->request->fence);
1775 /* Whilst this request exists, batch_obj will be on the
1776 * active_list, and so will hold the active reference. Only when this
1777 * request is retired will the the batch_obj be moved onto the
1778 * inactive_list and lose its active reference. Hence we do not need
1779 * to explicitly hold another reference here.
1781 params->request->batch = params->batch;
1784 * Save assorted stuff away to pass through to *_submission().
1785 * NB: This data should be 'persistent' and not local as it will
1786 * kept around beyond the duration of the IOCTL once the GPU
1787 * scheduler arrives.
1790 params->file = file;
1791 params->engine = engine;
1792 params->dispatch_flags = dispatch_flags;
1795 trace_i915_gem_request_queue(params->request, dispatch_flags);
1797 ret = execbuf_submit(params, args, &eb->vmas);
1799 __i915_add_request(params->request, ret == 0);
1800 add_to_client(params->request, file);
1804 fd_install(out_fence_fd, out_fence->file);
1805 args->rsvd2 &= GENMASK_ULL(0, 31); /* keep in-fence */
1806 args->rsvd2 |= (u64)out_fence_fd << 32;
1809 fput(out_fence->file);
1815 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1816 * batch vma for correctness. For less ugly and less fragility this
1817 * needs to be adjusted to also track the ggtt batch vma properly as
1820 if (dispatch_flags & I915_DISPATCH_SECURE)
1821 i915_vma_unpin(params->batch);
1823 /* the request owns the ref now */
1824 i915_gem_context_put(ctx);
1827 mutex_unlock(&dev->struct_mutex);
1830 /* intel_gpu_busy should also get a ref, so it will free when the device
1831 * is really idle. */
1832 intel_runtime_pm_put(dev_priv);
1833 if (out_fence_fd != -1)
1834 put_unused_fd(out_fence_fd);
1836 dma_fence_put(in_fence);
1841 * Legacy execbuffer just creates an exec2 list from the original exec object
1842 * list array and passes it to the real function.
1845 i915_gem_execbuffer(struct drm_device *dev, void *data,
1846 struct drm_file *file)
1848 struct drm_i915_gem_execbuffer *args = data;
1849 struct drm_i915_gem_execbuffer2 exec2;
1850 struct drm_i915_gem_exec_object *exec_list = NULL;
1851 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1854 if (args->buffer_count < 1) {
1855 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1859 /* Copy in the exec list from userland */
1860 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1861 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1862 if (exec_list == NULL || exec2_list == NULL) {
1863 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1864 args->buffer_count);
1865 drm_free_large(exec_list);
1866 drm_free_large(exec2_list);
1869 ret = copy_from_user(exec_list,
1870 u64_to_user_ptr(args->buffers_ptr),
1871 sizeof(*exec_list) * args->buffer_count);
1873 DRM_DEBUG("copy %d exec entries failed %d\n",
1874 args->buffer_count, ret);
1875 drm_free_large(exec_list);
1876 drm_free_large(exec2_list);
1880 for (i = 0; i < args->buffer_count; i++) {
1881 exec2_list[i].handle = exec_list[i].handle;
1882 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1883 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1884 exec2_list[i].alignment = exec_list[i].alignment;
1885 exec2_list[i].offset = exec_list[i].offset;
1886 if (INTEL_GEN(to_i915(dev)) < 4)
1887 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1889 exec2_list[i].flags = 0;
1892 exec2.buffers_ptr = args->buffers_ptr;
1893 exec2.buffer_count = args->buffer_count;
1894 exec2.batch_start_offset = args->batch_start_offset;
1895 exec2.batch_len = args->batch_len;
1896 exec2.DR1 = args->DR1;
1897 exec2.DR4 = args->DR4;
1898 exec2.num_cliprects = args->num_cliprects;
1899 exec2.cliprects_ptr = args->cliprects_ptr;
1900 exec2.flags = I915_EXEC_RENDER;
1901 i915_execbuffer2_set_context_id(exec2, 0);
1903 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1905 struct drm_i915_gem_exec_object __user *user_exec_list =
1906 u64_to_user_ptr(args->buffers_ptr);
1908 /* Copy the new buffer offsets back to the user's exec list. */
1909 for (i = 0; i < args->buffer_count; i++) {
1910 exec2_list[i].offset =
1911 gen8_canonical_addr(exec2_list[i].offset);
1912 ret = __copy_to_user(&user_exec_list[i].offset,
1913 &exec2_list[i].offset,
1914 sizeof(user_exec_list[i].offset));
1917 DRM_DEBUG("failed to copy %d exec entries "
1918 "back to user (%d)\n",
1919 args->buffer_count, ret);
1925 drm_free_large(exec_list);
1926 drm_free_large(exec2_list);
1931 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1932 struct drm_file *file)
1934 struct drm_i915_gem_execbuffer2 *args = data;
1935 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1938 if (args->buffer_count < 1 ||
1939 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1940 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1944 exec2_list = drm_malloc_gfp(args->buffer_count,
1945 sizeof(*exec2_list),
1947 if (exec2_list == NULL) {
1948 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1949 args->buffer_count);
1952 ret = copy_from_user(exec2_list,
1953 u64_to_user_ptr(args->buffers_ptr),
1954 sizeof(*exec2_list) * args->buffer_count);
1956 DRM_DEBUG("copy %d exec entries failed %d\n",
1957 args->buffer_count, ret);
1958 drm_free_large(exec2_list);
1962 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1964 /* Copy the new buffer offsets back to the user's exec list. */
1965 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1966 u64_to_user_ptr(args->buffers_ptr);
1969 for (i = 0; i < args->buffer_count; i++) {
1970 exec2_list[i].offset =
1971 gen8_canonical_addr(exec2_list[i].offset);
1972 ret = __copy_to_user(&user_exec_list[i].offset,
1973 &exec2_list[i].offset,
1974 sizeof(user_exec_list[i].offset));
1977 DRM_DEBUG("failed to copy %d exec entries "
1979 args->buffer_count);
1985 drm_free_large(exec2_list);