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 DBG_USE_CPU_RELOC 0 /* -1 force GTT relocs; 1 force CPU relocs */
44 #define __EXEC_OBJECT_HAS_PIN (1<<31)
45 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
46 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
47 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
48 #define __EXEC_OBJECT_INTERNAL_FLAGS (0xf<<28) /* all of the above */
50 #define BATCH_OFFSET_BIAS (256*1024)
52 struct i915_execbuffer_params {
53 struct drm_device *dev;
54 struct drm_file *file;
55 struct i915_vma *batch;
57 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_gem_obj_lookup_or_create_vma(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);
284 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
286 if (!i915_gem_object_has_struct_page(obj))
289 if (DBG_USE_CPU_RELOC)
290 return DBG_USE_CPU_RELOC > 0;
292 return (HAS_LLC(obj->base.dev) ||
293 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
294 obj->cache_level != I915_CACHE_NONE);
297 /* Used to convert any address to canonical form.
298 * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
299 * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
300 * addresses to be in a canonical form:
301 * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
302 * canonical form [63:48] == [47]."
304 #define GEN8_HIGH_ADDRESS_BIT 47
305 static inline uint64_t gen8_canonical_addr(uint64_t address)
307 return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
310 static inline uint64_t gen8_noncanonical_addr(uint64_t address)
312 return address & ((1ULL << (GEN8_HIGH_ADDRESS_BIT + 1)) - 1);
315 static inline uint64_t
316 relocation_target(const struct drm_i915_gem_relocation_entry *reloc,
317 uint64_t target_offset)
319 return gen8_canonical_addr((int)reloc->delta + target_offset);
323 struct drm_i915_private *i915;
324 struct drm_mm_node node;
327 bool use_64bit_reloc;
330 static void reloc_cache_init(struct reloc_cache *cache,
331 struct drm_i915_private *i915)
336 cache->use_64bit_reloc = INTEL_GEN(cache->i915) >= 8;
337 cache->node.allocated = false;
340 static inline void *unmask_page(unsigned long p)
342 return (void *)(uintptr_t)(p & PAGE_MASK);
345 static inline unsigned int unmask_flags(unsigned long p)
347 return p & ~PAGE_MASK;
350 #define KMAP 0x4 /* after CLFLUSH_FLAGS */
352 static void reloc_cache_fini(struct reloc_cache *cache)
359 vaddr = unmask_page(cache->vaddr);
360 if (cache->vaddr & KMAP) {
361 if (cache->vaddr & CLFLUSH_AFTER)
364 kunmap_atomic(vaddr);
365 i915_gem_obj_finish_shmem_access((struct drm_i915_gem_object *)cache->node.mm);
368 io_mapping_unmap_atomic((void __iomem *)vaddr);
369 if (cache->node.allocated) {
370 struct i915_ggtt *ggtt = &cache->i915->ggtt;
372 ggtt->base.clear_range(&ggtt->base,
376 drm_mm_remove_node(&cache->node);
378 i915_vma_unpin((struct i915_vma *)cache->node.mm);
383 static void *reloc_kmap(struct drm_i915_gem_object *obj,
384 struct reloc_cache *cache,
390 kunmap_atomic(unmask_page(cache->vaddr));
392 unsigned int flushes;
395 ret = i915_gem_obj_prepare_shmem_write(obj, &flushes);
399 BUILD_BUG_ON(KMAP & CLFLUSH_FLAGS);
400 BUILD_BUG_ON((KMAP | CLFLUSH_FLAGS) & PAGE_MASK);
402 cache->vaddr = flushes | KMAP;
403 cache->node.mm = (void *)obj;
408 vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj, page));
409 cache->vaddr = unmask_flags(cache->vaddr) | (unsigned long)vaddr;
415 static void *reloc_iomap(struct drm_i915_gem_object *obj,
416 struct reloc_cache *cache,
419 struct i915_ggtt *ggtt = &cache->i915->ggtt;
420 unsigned long offset;
423 if (cache->node.allocated) {
425 ggtt->base.insert_page(&ggtt->base,
426 i915_gem_object_get_dma_address(obj, page),
427 cache->node.start, I915_CACHE_NONE, 0);
429 return unmask_page(cache->vaddr);
433 io_mapping_unmap_atomic(unmask_page(cache->vaddr));
435 struct i915_vma *vma;
438 if (use_cpu_reloc(obj))
441 ret = i915_gem_object_set_to_gtt_domain(obj, true);
445 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
446 PIN_MAPPABLE | PIN_NONBLOCK);
448 memset(&cache->node, 0, sizeof(cache->node));
449 ret = drm_mm_insert_node_in_range_generic
450 (&ggtt->base.mm, &cache->node,
452 0, ggtt->mappable_end,
453 DRM_MM_SEARCH_DEFAULT,
454 DRM_MM_CREATE_DEFAULT);
455 if (ret) /* no inactive aperture space, use cpu reloc */
458 ret = i915_vma_put_fence(vma);
464 cache->node.start = vma->node.start;
465 cache->node.mm = (void *)vma;
469 offset = cache->node.start;
470 if (cache->node.allocated) {
471 ggtt->base.insert_page(&ggtt->base,
472 i915_gem_object_get_dma_address(obj, page),
473 offset, I915_CACHE_NONE, 0);
475 offset += page << PAGE_SHIFT;
478 vaddr = io_mapping_map_atomic_wc(&cache->i915->ggtt.mappable, offset);
480 cache->vaddr = (unsigned long)vaddr;
485 static void *reloc_vaddr(struct drm_i915_gem_object *obj,
486 struct reloc_cache *cache,
491 if (cache->page == page) {
492 vaddr = unmask_page(cache->vaddr);
495 if ((cache->vaddr & KMAP) == 0)
496 vaddr = reloc_iomap(obj, cache, page);
498 vaddr = reloc_kmap(obj, cache, page);
504 static void clflush_write32(u32 *addr, u32 value, unsigned int flushes)
506 if (unlikely(flushes & (CLFLUSH_BEFORE | CLFLUSH_AFTER))) {
507 if (flushes & CLFLUSH_BEFORE) {
514 /* Writes to the same cacheline are serialised by the CPU
515 * (including clflush). On the write path, we only require
516 * that it hits memory in an orderly fashion and place
517 * mb barriers at the start and end of the relocation phase
518 * to ensure ordering of clflush wrt to the system.
520 if (flushes & CLFLUSH_AFTER)
527 relocate_entry(struct drm_i915_gem_object *obj,
528 const struct drm_i915_gem_relocation_entry *reloc,
529 struct reloc_cache *cache,
532 u64 offset = reloc->offset;
533 bool wide = cache->use_64bit_reloc;
536 target_offset = relocation_target(reloc, target_offset);
538 vaddr = reloc_vaddr(obj, cache, offset >> PAGE_SHIFT);
540 return PTR_ERR(vaddr);
542 clflush_write32(vaddr + offset_in_page(offset),
543 lower_32_bits(target_offset),
547 offset += sizeof(u32);
548 target_offset >>= 32;
556 static bool object_is_idle(struct drm_i915_gem_object *obj)
558 unsigned long active = i915_gem_object_get_active(obj);
561 for_each_active(active, idx) {
562 if (!i915_gem_active_is_idle(&obj->last_read[idx],
563 &obj->base.dev->struct_mutex))
571 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
573 struct drm_i915_gem_relocation_entry *reloc,
574 struct reloc_cache *cache)
576 struct drm_device *dev = obj->base.dev;
577 struct drm_gem_object *target_obj;
578 struct drm_i915_gem_object *target_i915_obj;
579 struct i915_vma *target_vma;
580 uint64_t target_offset;
583 /* we've already hold a reference to all valid objects */
584 target_vma = eb_get_vma(eb, reloc->target_handle);
585 if (unlikely(target_vma == NULL))
587 target_i915_obj = target_vma->obj;
588 target_obj = &target_vma->obj->base;
590 target_offset = gen8_canonical_addr(target_vma->node.start);
592 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
593 * pipe_control writes because the gpu doesn't properly redirect them
594 * through the ppgtt for non_secure batchbuffers. */
595 if (unlikely(IS_GEN6(dev) &&
596 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
597 ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
599 if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
603 /* Validate that the target is in a valid r/w GPU domain */
604 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
605 DRM_DEBUG("reloc with multiple write domains: "
606 "obj %p target %d offset %d "
607 "read %08x write %08x",
608 obj, reloc->target_handle,
611 reloc->write_domain);
614 if (unlikely((reloc->write_domain | reloc->read_domains)
615 & ~I915_GEM_GPU_DOMAINS)) {
616 DRM_DEBUG("reloc with read/write non-GPU domains: "
617 "obj %p target %d offset %d "
618 "read %08x write %08x",
619 obj, reloc->target_handle,
622 reloc->write_domain);
626 target_obj->pending_read_domains |= reloc->read_domains;
627 target_obj->pending_write_domain |= reloc->write_domain;
629 /* If the relocation already has the right value in it, no
630 * more work needs to be done.
632 if (target_offset == reloc->presumed_offset)
635 /* Check that the relocation address is valid... */
636 if (unlikely(reloc->offset >
637 obj->base.size - (cache->use_64bit_reloc ? 8 : 4))) {
638 DRM_DEBUG("Relocation beyond object bounds: "
639 "obj %p target %d offset %d size %d.\n",
640 obj, reloc->target_handle,
642 (int) obj->base.size);
645 if (unlikely(reloc->offset & 3)) {
646 DRM_DEBUG("Relocation not 4-byte aligned: "
647 "obj %p target %d offset %d.\n",
648 obj, reloc->target_handle,
649 (int) reloc->offset);
653 /* We can't wait for rendering with pagefaults disabled */
654 if (pagefault_disabled() && !object_is_idle(obj))
657 ret = relocate_entry(obj, reloc, cache, target_offset);
661 /* and update the user's relocation entry */
662 reloc->presumed_offset = target_offset;
667 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
670 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
671 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
672 struct drm_i915_gem_relocation_entry __user *user_relocs;
673 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
674 struct reloc_cache cache;
677 user_relocs = u64_to_user_ptr(entry->relocs_ptr);
678 reloc_cache_init(&cache, eb->i915);
680 remain = entry->relocation_count;
682 struct drm_i915_gem_relocation_entry *r = stack_reloc;
684 if (count > ARRAY_SIZE(stack_reloc))
685 count = ARRAY_SIZE(stack_reloc);
688 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0]))) {
694 u64 offset = r->presumed_offset;
696 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r, &cache);
700 if (r->presumed_offset != offset &&
701 __put_user(r->presumed_offset,
702 &user_relocs->presumed_offset)) {
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 /* This is the fast path and we cannot handle a pagefault whilst
745 * holding the struct mutex lest the user pass in the relocations
746 * contained within a mmaped bo. For in such a case we, the page
747 * fault handler would call i915_gem_fault() and we would try to
748 * acquire the struct mutex again. Obviously this is bad and so
749 * lockdep complains vehemently.
752 list_for_each_entry(vma, &eb->vmas, exec_list) {
753 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
762 static bool only_mappable_for_reloc(unsigned int flags)
764 return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
765 __EXEC_OBJECT_NEEDS_MAP;
769 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
770 struct intel_engine_cs *engine,
773 struct drm_i915_gem_object *obj = vma->obj;
774 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
779 if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
782 if (!drm_mm_node_allocated(&vma->node)) {
783 /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
784 * limit address to the first 4GBs for unflagged objects.
786 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
787 flags |= PIN_ZONE_4G;
788 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
789 flags |= PIN_GLOBAL | PIN_MAPPABLE;
790 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
791 flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
792 if (entry->flags & EXEC_OBJECT_PINNED)
793 flags |= entry->offset | PIN_OFFSET_FIXED;
794 if ((flags & PIN_MAPPABLE) == 0)
798 ret = i915_vma_pin(vma,
802 if ((ret == -ENOSPC || ret == -E2BIG) &&
803 only_mappable_for_reloc(entry->flags))
804 ret = i915_vma_pin(vma,
807 flags & ~PIN_MAPPABLE);
811 entry->flags |= __EXEC_OBJECT_HAS_PIN;
813 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
814 ret = i915_vma_get_fence(vma);
818 if (i915_vma_pin_fence(vma))
819 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
822 if (entry->offset != vma->node.start) {
823 entry->offset = vma->node.start;
827 if (entry->flags & EXEC_OBJECT_WRITE) {
828 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
829 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
836 need_reloc_mappable(struct i915_vma *vma)
838 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
840 if (entry->relocation_count == 0)
843 if (!i915_vma_is_ggtt(vma))
846 /* See also use_cpu_reloc() */
847 if (HAS_LLC(vma->obj->base.dev))
850 if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
857 eb_vma_misplaced(struct i915_vma *vma)
859 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
861 WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
862 !i915_vma_is_ggtt(vma));
864 if (entry->alignment &&
865 vma->node.start & (entry->alignment - 1))
868 if (vma->node.size < entry->pad_to_size)
871 if (entry->flags & EXEC_OBJECT_PINNED &&
872 vma->node.start != entry->offset)
875 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
876 vma->node.start < BATCH_OFFSET_BIAS)
879 /* avoid costly ping-pong once a batch bo ended up non-mappable */
880 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
881 !i915_vma_is_map_and_fenceable(vma))
882 return !only_mappable_for_reloc(entry->flags);
884 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
885 (vma->node.start + vma->node.size + 4095) >> 32)
892 i915_gem_execbuffer_reserve(struct intel_engine_cs *engine,
893 struct list_head *vmas,
894 struct i915_gem_context *ctx,
897 struct drm_i915_gem_object *obj;
898 struct i915_vma *vma;
899 struct i915_address_space *vm;
900 struct list_head ordered_vmas;
901 struct list_head pinned_vmas;
902 bool has_fenced_gpu_access = INTEL_GEN(engine->i915) < 4;
905 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
907 INIT_LIST_HEAD(&ordered_vmas);
908 INIT_LIST_HEAD(&pinned_vmas);
909 while (!list_empty(vmas)) {
910 struct drm_i915_gem_exec_object2 *entry;
911 bool need_fence, need_mappable;
913 vma = list_first_entry(vmas, struct i915_vma, exec_list);
915 entry = vma->exec_entry;
917 if (ctx->flags & CONTEXT_NO_ZEROMAP)
918 entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
920 if (!has_fenced_gpu_access)
921 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
923 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
924 i915_gem_object_is_tiled(obj);
925 need_mappable = need_fence || need_reloc_mappable(vma);
927 if (entry->flags & EXEC_OBJECT_PINNED)
928 list_move_tail(&vma->exec_list, &pinned_vmas);
929 else if (need_mappable) {
930 entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
931 list_move(&vma->exec_list, &ordered_vmas);
933 list_move_tail(&vma->exec_list, &ordered_vmas);
935 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
936 obj->base.pending_write_domain = 0;
938 list_splice(&ordered_vmas, vmas);
939 list_splice(&pinned_vmas, vmas);
941 /* Attempt to pin all of the buffers into the GTT.
942 * This is done in 3 phases:
944 * 1a. Unbind all objects that do not match the GTT constraints for
945 * the execbuffer (fenceable, mappable, alignment etc).
946 * 1b. Increment pin count for already bound objects.
947 * 2. Bind new objects.
948 * 3. Decrement pin count.
950 * This avoid unnecessary unbinding of later objects in order to make
951 * room for the earlier objects *unless* we need to defragment.
957 /* Unbind any ill-fitting objects or pin. */
958 list_for_each_entry(vma, vmas, exec_list) {
959 if (!drm_mm_node_allocated(&vma->node))
962 if (eb_vma_misplaced(vma))
963 ret = i915_vma_unbind(vma);
965 ret = i915_gem_execbuffer_reserve_vma(vma,
972 /* Bind fresh objects */
973 list_for_each_entry(vma, vmas, exec_list) {
974 if (drm_mm_node_allocated(&vma->node))
977 ret = i915_gem_execbuffer_reserve_vma(vma, engine,
984 if (ret != -ENOSPC || retry++)
987 /* Decrement pin count for bound objects */
988 list_for_each_entry(vma, vmas, exec_list)
989 i915_gem_execbuffer_unreserve_vma(vma);
991 ret = i915_gem_evict_vm(vm, true);
998 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
999 struct drm_i915_gem_execbuffer2 *args,
1000 struct drm_file *file,
1001 struct intel_engine_cs *engine,
1003 struct drm_i915_gem_exec_object2 *exec,
1004 struct i915_gem_context *ctx)
1006 struct drm_i915_gem_relocation_entry *reloc;
1007 struct i915_address_space *vm;
1008 struct i915_vma *vma;
1012 unsigned count = args->buffer_count;
1014 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
1016 /* We may process another execbuffer during the unlock... */
1017 while (!list_empty(&eb->vmas)) {
1018 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
1019 list_del_init(&vma->exec_list);
1020 i915_gem_execbuffer_unreserve_vma(vma);
1024 mutex_unlock(&dev->struct_mutex);
1027 for (i = 0; i < count; i++)
1028 total += exec[i].relocation_count;
1030 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
1031 reloc = drm_malloc_ab(total, sizeof(*reloc));
1032 if (reloc == NULL || reloc_offset == NULL) {
1033 drm_free_large(reloc);
1034 drm_free_large(reloc_offset);
1035 mutex_lock(&dev->struct_mutex);
1040 for (i = 0; i < count; i++) {
1041 struct drm_i915_gem_relocation_entry __user *user_relocs;
1042 u64 invalid_offset = (u64)-1;
1045 user_relocs = u64_to_user_ptr(exec[i].relocs_ptr);
1047 if (copy_from_user(reloc+total, user_relocs,
1048 exec[i].relocation_count * sizeof(*reloc))) {
1050 mutex_lock(&dev->struct_mutex);
1054 /* As we do not update the known relocation offsets after
1055 * relocating (due to the complexities in lock handling),
1056 * we need to mark them as invalid now so that we force the
1057 * relocation processing next time. Just in case the target
1058 * object is evicted and then rebound into its old
1059 * presumed_offset before the next execbuffer - if that
1060 * happened we would make the mistake of assuming that the
1061 * relocations were valid.
1063 for (j = 0; j < exec[i].relocation_count; j++) {
1064 if (__copy_to_user(&user_relocs[j].presumed_offset,
1066 sizeof(invalid_offset))) {
1068 mutex_lock(&dev->struct_mutex);
1073 reloc_offset[i] = total;
1074 total += exec[i].relocation_count;
1077 ret = i915_mutex_lock_interruptible(dev);
1079 mutex_lock(&dev->struct_mutex);
1083 /* reacquire the objects */
1085 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1089 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1090 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1095 list_for_each_entry(vma, &eb->vmas, exec_list) {
1096 int offset = vma->exec_entry - exec;
1097 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
1098 reloc + reloc_offset[offset]);
1103 /* Leave the user relocations as are, this is the painfully slow path,
1104 * and we want to avoid the complication of dropping the lock whilst
1105 * having buffers reserved in the aperture and so causing spurious
1106 * ENOSPC for random operations.
1110 drm_free_large(reloc);
1111 drm_free_large(reloc_offset);
1115 static unsigned int eb_other_engines(struct drm_i915_gem_request *req)
1119 mask = ~intel_engine_flag(req->engine) & I915_BO_ACTIVE_MASK;
1120 mask <<= I915_BO_ACTIVE_SHIFT;
1126 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
1127 struct list_head *vmas)
1129 const unsigned int other_rings = eb_other_engines(req);
1130 struct i915_vma *vma;
1133 list_for_each_entry(vma, vmas, exec_list) {
1134 struct drm_i915_gem_object *obj = vma->obj;
1135 struct reservation_object *resv;
1137 if (obj->flags & other_rings) {
1138 ret = i915_gem_request_await_object
1139 (req, obj, obj->base.pending_write_domain);
1144 resv = i915_gem_object_get_dmabuf_resv(obj);
1146 ret = i915_sw_fence_await_reservation
1147 (&req->submit, resv, &i915_fence_ops,
1148 obj->base.pending_write_domain, 10*HZ,
1149 GFP_KERNEL | __GFP_NOWARN);
1154 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
1155 i915_gem_clflush_object(obj, false);
1158 /* Unconditionally flush any chipset caches (for streaming writes). */
1159 i915_gem_chipset_flush(req->engine->i915);
1161 /* Unconditionally invalidate GPU caches and TLBs. */
1162 return req->engine->emit_flush(req, EMIT_INVALIDATE);
1166 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
1168 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
1171 /* Kernel clipping was a DRI1 misfeature */
1172 if (exec->num_cliprects || exec->cliprects_ptr)
1175 if (exec->DR4 == 0xffffffff) {
1176 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1179 if (exec->DR1 || exec->DR4)
1182 if ((exec->batch_start_offset | exec->batch_len) & 0x7)
1189 validate_exec_list(struct drm_device *dev,
1190 struct drm_i915_gem_exec_object2 *exec,
1193 unsigned relocs_total = 0;
1194 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
1195 unsigned invalid_flags;
1198 /* INTERNAL flags must not overlap with external ones */
1199 BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & ~__EXEC_OBJECT_UNKNOWN_FLAGS);
1201 invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
1202 if (USES_FULL_PPGTT(dev))
1203 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
1205 for (i = 0; i < count; i++) {
1206 char __user *ptr = u64_to_user_ptr(exec[i].relocs_ptr);
1207 int length; /* limited by fault_in_pages_readable() */
1209 if (exec[i].flags & invalid_flags)
1212 /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
1213 * any non-page-aligned or non-canonical addresses.
1215 if (exec[i].flags & EXEC_OBJECT_PINNED) {
1216 if (exec[i].offset !=
1217 gen8_canonical_addr(exec[i].offset & PAGE_MASK))
1221 /* From drm_mm perspective address space is continuous,
1222 * so from this point we're always using non-canonical
1225 exec[i].offset = gen8_noncanonical_addr(exec[i].offset);
1227 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
1230 /* pad_to_size was once a reserved field, so sanitize it */
1231 if (exec[i].flags & EXEC_OBJECT_PAD_TO_SIZE) {
1232 if (offset_in_page(exec[i].pad_to_size))
1235 exec[i].pad_to_size = 0;
1238 /* First check for malicious input causing overflow in
1239 * the worst case where we need to allocate the entire
1240 * relocation tree as a single array.
1242 if (exec[i].relocation_count > relocs_max - relocs_total)
1244 relocs_total += exec[i].relocation_count;
1246 length = exec[i].relocation_count *
1247 sizeof(struct drm_i915_gem_relocation_entry);
1249 * We must check that the entire relocation array is safe
1250 * to read, but since we may need to update the presumed
1251 * offsets during execution, check for full write access.
1253 if (!access_ok(VERIFY_WRITE, ptr, length))
1256 if (likely(!i915.prefault_disable)) {
1257 if (fault_in_pages_readable(ptr, length))
1265 static struct i915_gem_context *
1266 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
1267 struct intel_engine_cs *engine, const u32 ctx_id)
1269 struct i915_gem_context *ctx;
1270 struct i915_ctx_hang_stats *hs;
1272 ctx = i915_gem_context_lookup(file->driver_priv, ctx_id);
1276 hs = &ctx->hang_stats;
1278 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1279 return ERR_PTR(-EIO);
1285 static bool gpu_write_needs_clflush(struct drm_i915_gem_object *obj)
1287 return !(obj->cache_level == I915_CACHE_NONE ||
1288 obj->cache_level == I915_CACHE_WT);
1291 void i915_vma_move_to_active(struct i915_vma *vma,
1292 struct drm_i915_gem_request *req,
1295 struct drm_i915_gem_object *obj = vma->obj;
1296 const unsigned int idx = req->engine->id;
1298 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1300 obj->dirty = 1; /* be paranoid */
1302 /* Add a reference if we're newly entering the active list.
1303 * The order in which we add operations to the retirement queue is
1304 * vital here: mark_active adds to the start of the callback list,
1305 * such that subsequent callbacks are called first. Therefore we
1306 * add the active reference first and queue for it to be dropped
1309 if (!i915_gem_object_is_active(obj))
1310 i915_gem_object_get(obj);
1311 i915_gem_object_set_active(obj, idx);
1312 i915_gem_active_set(&obj->last_read[idx], req);
1314 if (flags & EXEC_OBJECT_WRITE) {
1315 i915_gem_active_set(&obj->last_write, req);
1317 intel_fb_obj_invalidate(obj, ORIGIN_CS);
1319 /* update for the implicit flush after a batch */
1320 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1321 if (!obj->cache_dirty && gpu_write_needs_clflush(obj))
1322 obj->cache_dirty = true;
1325 if (flags & EXEC_OBJECT_NEEDS_FENCE)
1326 i915_gem_active_set(&vma->last_fence, req);
1328 i915_vma_set_active(vma, idx);
1329 i915_gem_active_set(&vma->last_read[idx], req);
1330 list_move_tail(&vma->vm_link, &vma->vm->active_list);
1333 static void eb_export_fence(struct drm_i915_gem_object *obj,
1334 struct drm_i915_gem_request *req,
1337 struct reservation_object *resv;
1339 resv = i915_gem_object_get_dmabuf_resv(obj);
1343 /* Ignore errors from failing to allocate the new fence, we can't
1344 * handle an error right now. Worst case should be missed
1345 * synchronisation leading to rendering corruption.
1347 ww_mutex_lock(&resv->lock, NULL);
1348 if (flags & EXEC_OBJECT_WRITE)
1349 reservation_object_add_excl_fence(resv, &req->fence);
1350 else if (reservation_object_reserve_shared(resv) == 0)
1351 reservation_object_add_shared_fence(resv, &req->fence);
1352 ww_mutex_unlock(&resv->lock);
1356 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1357 struct drm_i915_gem_request *req)
1359 struct i915_vma *vma;
1361 list_for_each_entry(vma, vmas, exec_list) {
1362 struct drm_i915_gem_object *obj = vma->obj;
1363 u32 old_read = obj->base.read_domains;
1364 u32 old_write = obj->base.write_domain;
1366 obj->base.write_domain = obj->base.pending_write_domain;
1367 if (obj->base.write_domain)
1368 vma->exec_entry->flags |= EXEC_OBJECT_WRITE;
1370 obj->base.pending_read_domains |= obj->base.read_domains;
1371 obj->base.read_domains = obj->base.pending_read_domains;
1373 i915_vma_move_to_active(vma, req, vma->exec_entry->flags);
1374 eb_export_fence(obj, req, vma->exec_entry->flags);
1375 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1380 i915_reset_gen7_sol_offsets(struct drm_i915_gem_request *req)
1382 struct intel_ring *ring = req->ring;
1385 if (!IS_GEN7(req->i915) || req->engine->id != RCS) {
1386 DRM_DEBUG("sol reset is gen7/rcs only\n");
1390 ret = intel_ring_begin(req, 4 * 3);
1394 for (i = 0; i < 4; i++) {
1395 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1396 intel_ring_emit_reg(ring, GEN7_SO_WRITE_OFFSET(i));
1397 intel_ring_emit(ring, 0);
1400 intel_ring_advance(ring);
1405 static struct i915_vma*
1406 shadow_batch_pin(struct drm_i915_gem_object *obj, struct i915_address_space *vm)
1408 struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
1412 * PPGTT backed shadow buffers must be mapped RO, to prevent
1413 * post-scan tampering
1415 if (CMDPARSER_USES_GGTT(dev_priv)) {
1417 vm = &dev_priv->ggtt.base;
1418 } else if (vm->has_read_only) {
1420 i915_gem_object_set_readonly(obj);
1422 DRM_DEBUG("Cannot prevent post-scan tampering without RO capable vm\n");
1423 return ERR_PTR(-EINVAL);
1426 return i915_gem_object_pin(obj, vm, NULL, 0, 0, flags);
1429 static struct i915_vma *
1430 i915_gem_execbuffer_parse(struct intel_engine_cs *engine,
1431 struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1432 struct i915_execbuffer_params *params,
1434 struct i915_address_space *vm)
1436 struct drm_i915_gem_object *batch_obj = params->batch->obj;
1437 struct drm_i915_gem_object *shadow_batch_obj;
1438 struct i915_vma *vma;
1440 u32 batch_start_offset = params->args_batch_start_offset;
1441 u32 batch_len = params->args_batch_len;
1442 u64 shadow_batch_start;
1446 shadow_batch_obj = i915_gem_batch_pool_get(&engine->batch_pool,
1447 PAGE_ALIGN(batch_len));
1448 if (IS_ERR(shadow_batch_obj))
1449 return ERR_CAST(shadow_batch_obj);
1451 vma = shadow_batch_pin(shadow_batch_obj, vm);
1455 batch_start = gen8_canonical_addr(params->batch->node.start) +
1457 shadow_batch_start = gen8_canonical_addr(vma->node.start);
1459 ret = intel_engine_cmd_parser(params->ctx,
1466 shadow_batch_start);
1468 i915_vma_unpin(vma);
1471 * Unsafe GGTT-backed buffers can still be submitted safely
1473 * For PPGTT backing however, we have no choice but to forcibly
1474 * reject unsafe buffers
1476 if (CMDPARSER_USES_GGTT(eb->i915) && (ret == -EACCES))
1477 /* Execute original buffer non-secure */
1485 memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1487 vma->exec_entry = shadow_exec_entry;
1488 vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1489 i915_gem_object_get(shadow_batch_obj);
1490 list_add_tail(&vma->exec_list, &eb->vmas);
1493 i915_gem_object_unpin_pages(shadow_batch_obj);
1498 execbuf_submit(struct i915_execbuffer_params *params,
1499 struct drm_i915_gem_execbuffer2 *args,
1500 struct list_head *vmas)
1502 u64 exec_start, exec_len;
1505 ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1509 ret = i915_switch_context(params->request);
1513 if (args->flags & I915_EXEC_CONSTANTS_MASK) {
1514 DRM_DEBUG("I915_EXEC_CONSTANTS_* unsupported\n");
1518 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1519 ret = i915_reset_gen7_sol_offsets(params->request);
1524 exec_len = params->args_batch_len;
1525 exec_start = params->batch->node.start +
1526 params->args_batch_start_offset;
1528 ret = params->engine->emit_bb_start(params->request,
1529 exec_start, exec_len,
1530 params->dispatch_flags);
1534 trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1536 i915_gem_execbuffer_move_to_active(vmas, params->request);
1542 * Find one BSD ring to dispatch the corresponding BSD command.
1543 * The engine index is returned.
1546 gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
1547 struct drm_file *file)
1549 struct drm_i915_file_private *file_priv = file->driver_priv;
1551 /* Check whether the file_priv has already selected one ring. */
1552 if ((int)file_priv->bsd_engine < 0)
1553 file_priv->bsd_engine = atomic_fetch_xor(1,
1554 &dev_priv->mm.bsd_engine_dispatch_index);
1556 return file_priv->bsd_engine;
1559 #define I915_USER_RINGS (4)
1561 static const enum intel_engine_id user_ring_map[I915_USER_RINGS + 1] = {
1562 [I915_EXEC_DEFAULT] = RCS,
1563 [I915_EXEC_RENDER] = RCS,
1564 [I915_EXEC_BLT] = BCS,
1565 [I915_EXEC_BSD] = VCS,
1566 [I915_EXEC_VEBOX] = VECS
1569 static struct intel_engine_cs *
1570 eb_select_engine(struct drm_i915_private *dev_priv,
1571 struct drm_file *file,
1572 struct drm_i915_gem_execbuffer2 *args)
1574 unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
1575 struct intel_engine_cs *engine;
1577 if (user_ring_id > I915_USER_RINGS) {
1578 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id);
1582 if ((user_ring_id != I915_EXEC_BSD) &&
1583 ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1584 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1585 "bsd dispatch flags: %d\n", (int)(args->flags));
1589 if (user_ring_id == I915_EXEC_BSD && HAS_BSD2(dev_priv)) {
1590 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
1592 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
1593 bsd_idx = gen8_dispatch_bsd_engine(dev_priv, file);
1594 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
1595 bsd_idx <= I915_EXEC_BSD_RING2) {
1596 bsd_idx >>= I915_EXEC_BSD_SHIFT;
1599 DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
1604 engine = &dev_priv->engine[_VCS(bsd_idx)];
1606 engine = &dev_priv->engine[user_ring_map[user_ring_id]];
1609 if (!intel_engine_initialized(engine)) {
1610 DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id);
1618 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1619 struct drm_file *file,
1620 struct drm_i915_gem_execbuffer2 *args,
1621 struct drm_i915_gem_exec_object2 *exec)
1623 struct drm_i915_private *dev_priv = to_i915(dev);
1624 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1626 struct drm_i915_gem_exec_object2 shadow_exec_entry;
1627 struct intel_engine_cs *engine;
1628 struct i915_gem_context *ctx;
1629 struct i915_address_space *vm;
1630 struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1631 struct i915_execbuffer_params *params = ¶ms_master;
1632 const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1637 if (!i915_gem_check_execbuffer(args))
1640 ret = validate_exec_list(dev, exec, args->buffer_count);
1645 if (args->flags & I915_EXEC_SECURE) {
1646 if (INTEL_GEN(dev_priv) >= 11)
1649 /* Return -EPERM to trigger fallback code on old binaries. */
1650 if (!HAS_SECURE_BATCHES(dev_priv))
1653 if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
1656 dispatch_flags |= I915_DISPATCH_SECURE;
1658 if (args->flags & I915_EXEC_IS_PINNED)
1659 dispatch_flags |= I915_DISPATCH_PINNED;
1661 engine = eb_select_engine(dev_priv, file, args);
1665 if (args->buffer_count < 1) {
1666 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1670 if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1671 if (!HAS_RESOURCE_STREAMER(dev)) {
1672 DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1675 if (engine->id != RCS) {
1676 DRM_DEBUG("RS is not available on %s\n",
1681 dispatch_flags |= I915_DISPATCH_RS;
1684 /* Take a local wakeref for preparing to dispatch the execbuf as
1685 * we expect to access the hardware fairly frequently in the
1686 * process. Upon first dispatch, we acquire another prolonged
1687 * wakeref that we hold until the GPU has been idle for at least
1690 intel_runtime_pm_get(dev_priv);
1692 ret = i915_mutex_lock_interruptible(dev);
1696 ctx = i915_gem_validate_context(dev, file, engine, ctx_id);
1698 mutex_unlock(&dev->struct_mutex);
1703 i915_gem_context_get(ctx);
1706 vm = &ctx->ppgtt->base;
1710 memset(¶ms_master, 0x00, sizeof(params_master));
1712 eb = eb_create(dev_priv, args);
1714 i915_gem_context_put(ctx);
1715 mutex_unlock(&dev->struct_mutex);
1720 /* Look up object handles */
1721 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1725 /* take note of the batch buffer before we might reorder the lists */
1726 params->batch = eb_get_batch(eb);
1728 /* Move the objects en-masse into the GTT, evicting if necessary. */
1729 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1730 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1735 /* The objects are in their final locations, apply the relocations. */
1737 ret = i915_gem_execbuffer_relocate(eb);
1739 if (ret == -EFAULT) {
1740 ret = i915_gem_execbuffer_relocate_slow(dev, args, file,
1743 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1749 /* Set the pending read domains for the batch buffer to COMMAND */
1750 if (params->batch->obj->base.pending_write_domain) {
1751 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1755 if (args->batch_start_offset > params->batch->size ||
1756 args->batch_len > params->batch->size - args->batch_start_offset) {
1757 DRM_DEBUG("Attempting to use out-of-bounds batch\n");
1763 params->args_batch_start_offset = args->batch_start_offset;
1764 params->args_batch_len = args->batch_len;
1765 if (args->batch_len == 0)
1766 params->args_batch_len = params->batch->size - params->args_batch_start_offset;
1768 if (intel_engine_requires_cmd_parser(engine) ||
1769 (intel_engine_using_cmd_parser(engine) && args->batch_len)) {
1770 struct i915_vma *vma;
1772 vma = i915_gem_execbuffer_parse(engine, &shadow_exec_entry,
1780 if (CMDPARSER_USES_GGTT(dev_priv))
1781 dispatch_flags |= I915_DISPATCH_SECURE;
1782 params->args_batch_start_offset = 0;
1783 params->batch = vma;
1787 params->batch->obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1789 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1790 * batch" bit. Hence we need to pin secure batches into the global gtt.
1791 * hsw should have this fixed, but bdw mucks it up again. */
1792 if (dispatch_flags & I915_DISPATCH_SECURE) {
1793 struct drm_i915_gem_object *obj = params->batch->obj;
1794 struct i915_vma *vma;
1797 * So on first glance it looks freaky that we pin the batch here
1798 * outside of the reservation loop. But:
1799 * - The batch is already pinned into the relevant ppgtt, so we
1800 * already have the backing storage fully allocated.
1801 * - No other BO uses the global gtt (well contexts, but meh),
1802 * so we don't really have issues with multiple objects not
1803 * fitting due to fragmentation.
1804 * So this is actually safe.
1806 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, 0);
1812 params->batch = vma;
1815 /* Allocate a request for this batch buffer nice and early. */
1816 params->request = i915_gem_request_alloc(engine, ctx);
1817 if (IS_ERR(params->request)) {
1818 ret = PTR_ERR(params->request);
1819 goto err_batch_unpin;
1822 /* Whilst this request exists, batch_obj will be on the
1823 * active_list, and so will hold the active reference. Only when this
1824 * request is retired will the the batch_obj be moved onto the
1825 * inactive_list and lose its active reference. Hence we do not need
1826 * to explicitly hold another reference here.
1828 params->request->batch = params->batch;
1830 ret = i915_gem_request_add_to_client(params->request, file);
1835 * Save assorted stuff away to pass through to *_submission().
1836 * NB: This data should be 'persistent' and not local as it will
1837 * kept around beyond the duration of the IOCTL once the GPU
1838 * scheduler arrives.
1841 params->file = file;
1842 params->engine = engine;
1843 params->dispatch_flags = dispatch_flags;
1845 ret = execbuf_submit(params, args, &eb->vmas);
1847 __i915_add_request(params->request, ret == 0);
1851 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1852 * batch vma for correctness. For less ugly and less fragility this
1853 * needs to be adjusted to also track the ggtt batch vma properly as
1856 if (dispatch_flags & I915_DISPATCH_SECURE)
1857 i915_vma_unpin(params->batch);
1859 /* the request owns the ref now */
1860 i915_gem_context_put(ctx);
1863 mutex_unlock(&dev->struct_mutex);
1866 /* intel_gpu_busy should also get a ref, so it will free when the device
1867 * is really idle. */
1868 intel_runtime_pm_put(dev_priv);
1873 * Legacy execbuffer just creates an exec2 list from the original exec object
1874 * list array and passes it to the real function.
1877 i915_gem_execbuffer(struct drm_device *dev, void *data,
1878 struct drm_file *file)
1880 struct drm_i915_gem_execbuffer *args = data;
1881 struct drm_i915_gem_execbuffer2 exec2;
1882 struct drm_i915_gem_exec_object *exec_list = NULL;
1883 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1886 if (args->buffer_count < 1) {
1887 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1891 /* Copy in the exec list from userland */
1892 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1893 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1894 if (exec_list == NULL || exec2_list == NULL) {
1895 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1896 args->buffer_count);
1897 drm_free_large(exec_list);
1898 drm_free_large(exec2_list);
1901 ret = copy_from_user(exec_list,
1902 u64_to_user_ptr(args->buffers_ptr),
1903 sizeof(*exec_list) * args->buffer_count);
1905 DRM_DEBUG("copy %d exec entries failed %d\n",
1906 args->buffer_count, ret);
1907 drm_free_large(exec_list);
1908 drm_free_large(exec2_list);
1912 for (i = 0; i < args->buffer_count; i++) {
1913 exec2_list[i].handle = exec_list[i].handle;
1914 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1915 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1916 exec2_list[i].alignment = exec_list[i].alignment;
1917 exec2_list[i].offset = exec_list[i].offset;
1918 if (INTEL_INFO(dev)->gen < 4)
1919 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1921 exec2_list[i].flags = 0;
1924 exec2.buffers_ptr = args->buffers_ptr;
1925 exec2.buffer_count = args->buffer_count;
1926 exec2.batch_start_offset = args->batch_start_offset;
1927 exec2.batch_len = args->batch_len;
1928 exec2.DR1 = args->DR1;
1929 exec2.DR4 = args->DR4;
1930 exec2.num_cliprects = args->num_cliprects;
1931 exec2.cliprects_ptr = args->cliprects_ptr;
1932 exec2.flags = I915_EXEC_RENDER;
1933 i915_execbuffer2_set_context_id(exec2, 0);
1935 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1937 struct drm_i915_gem_exec_object __user *user_exec_list =
1938 u64_to_user_ptr(args->buffers_ptr);
1940 /* Copy the new buffer offsets back to the user's exec list. */
1941 for (i = 0; i < args->buffer_count; i++) {
1942 exec2_list[i].offset =
1943 gen8_canonical_addr(exec2_list[i].offset);
1944 ret = __copy_to_user(&user_exec_list[i].offset,
1945 &exec2_list[i].offset,
1946 sizeof(user_exec_list[i].offset));
1949 DRM_DEBUG("failed to copy %d exec entries "
1950 "back to user (%d)\n",
1951 args->buffer_count, ret);
1957 drm_free_large(exec_list);
1958 drm_free_large(exec2_list);
1963 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1964 struct drm_file *file)
1966 struct drm_i915_gem_execbuffer2 *args = data;
1967 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1970 if (args->buffer_count < 1 ||
1971 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1972 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1976 if (args->rsvd2 != 0) {
1977 DRM_DEBUG("dirty rvsd2 field\n");
1981 exec2_list = drm_malloc_gfp(args->buffer_count,
1982 sizeof(*exec2_list),
1984 if (exec2_list == NULL) {
1985 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1986 args->buffer_count);
1989 ret = copy_from_user(exec2_list,
1990 u64_to_user_ptr(args->buffers_ptr),
1991 sizeof(*exec2_list) * args->buffer_count);
1993 DRM_DEBUG("copy %d exec entries failed %d\n",
1994 args->buffer_count, ret);
1995 drm_free_large(exec2_list);
1999 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
2001 /* Copy the new buffer offsets back to the user's exec list. */
2002 struct drm_i915_gem_exec_object2 __user *user_exec_list =
2003 u64_to_user_ptr(args->buffers_ptr);
2006 for (i = 0; i < args->buffer_count; i++) {
2007 exec2_list[i].offset =
2008 gen8_canonical_addr(exec2_list[i].offset);
2009 ret = __copy_to_user(&user_exec_list[i].offset,
2010 &exec2_list[i].offset,
2011 sizeof(user_exec_list[i].offset));
2014 DRM_DEBUG("failed to copy %d exec entries "
2016 args->buffer_count);
2022 drm_free_large(exec2_list);