2 * Copyright © 2008,2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eric Anholt <eric@anholt.net>
25 * Chris Wilson <chris@chris-wilson.co.uk>
30 #include <drm/i915_drm.h>
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/dma_remapping.h>
35 #include <linux/uaccess.h>
37 #define __EXEC_OBJECT_HAS_PIN (1<<31)
38 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
39 #define __EXEC_OBJECT_NEEDS_MAP (1<<29)
40 #define __EXEC_OBJECT_NEEDS_BIAS (1<<28)
41 #define __EXEC_OBJECT_INTERNAL_FLAGS (0xf<<28) /* all of the above */
43 #define BATCH_OFFSET_BIAS (256*1024)
45 struct i915_execbuffer_params {
46 struct drm_device *dev;
47 struct drm_file *file;
49 u32 args_batch_start_offset;
50 u32 batch_obj_vm_offset;
51 struct intel_engine_cs *engine;
52 struct drm_i915_gem_object *batch_obj;
53 struct i915_gem_context *ctx;
54 struct drm_i915_gem_request *request;
58 struct list_head vmas;
61 struct i915_vma *lut[0];
62 struct hlist_head buckets[0];
66 static struct eb_vmas *
67 eb_create(struct drm_i915_gem_execbuffer2 *args)
69 struct eb_vmas *eb = NULL;
71 if (args->flags & I915_EXEC_HANDLE_LUT) {
72 unsigned size = args->buffer_count;
73 size *= sizeof(struct i915_vma *);
74 size += sizeof(struct eb_vmas);
75 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
79 unsigned size = args->buffer_count;
80 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
81 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
82 while (count > 2*size)
84 eb = kzalloc(count*sizeof(struct hlist_head) +
85 sizeof(struct eb_vmas),
92 eb->and = -args->buffer_count;
94 INIT_LIST_HEAD(&eb->vmas);
99 eb_reset(struct eb_vmas *eb)
102 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
106 eb_lookup_vmas(struct eb_vmas *eb,
107 struct drm_i915_gem_exec_object2 *exec,
108 const struct drm_i915_gem_execbuffer2 *args,
109 struct i915_address_space *vm,
110 struct drm_file *file)
112 struct drm_i915_gem_object *obj;
113 struct list_head objects;
116 INIT_LIST_HEAD(&objects);
117 spin_lock(&file->table_lock);
118 /* Grab a reference to the object and release the lock so we can lookup
119 * or create the VMA without using GFP_ATOMIC */
120 for (i = 0; i < args->buffer_count; i++) {
121 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
123 spin_unlock(&file->table_lock);
124 DRM_DEBUG("Invalid object handle %d at index %d\n",
130 if (!list_empty(&obj->obj_exec_link)) {
131 spin_unlock(&file->table_lock);
132 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
133 obj, exec[i].handle, i);
138 i915_gem_object_get(obj);
139 list_add_tail(&obj->obj_exec_link, &objects);
141 spin_unlock(&file->table_lock);
144 while (!list_empty(&objects)) {
145 struct i915_vma *vma;
147 obj = list_first_entry(&objects,
148 struct drm_i915_gem_object,
152 * NOTE: We can leak any vmas created here when something fails
153 * later on. But that's no issue since vma_unbind can deal with
154 * vmas which are not actually bound. And since only
155 * lookup_or_create exists as an interface to get at the vma
156 * from the (obj, vm) we don't run the risk of creating
157 * duplicated vmas for the same vm.
159 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
161 DRM_DEBUG("Failed to lookup VMA\n");
166 /* Transfer ownership from the objects list to the vmas list. */
167 list_add_tail(&vma->exec_list, &eb->vmas);
168 list_del_init(&obj->obj_exec_link);
170 vma->exec_entry = &exec[i];
174 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
175 vma->exec_handle = handle;
176 hlist_add_head(&vma->exec_node,
177 &eb->buckets[handle & eb->and]);
186 while (!list_empty(&objects)) {
187 obj = list_first_entry(&objects,
188 struct drm_i915_gem_object,
190 list_del_init(&obj->obj_exec_link);
191 i915_gem_object_put(obj);
194 * Objects already transfered to the vmas list will be unreferenced by
201 static inline struct i915_vma *
202 eb_get_batch_vma(struct eb_vmas *eb)
204 /* The batch is always the LAST item in the VMA list */
205 struct i915_vma *vma = list_last_entry(&eb->vmas, typeof(*vma), exec_list);
210 static struct drm_i915_gem_object *
211 eb_get_batch(struct eb_vmas *eb)
213 struct i915_vma *vma = eb_get_batch_vma(eb);
216 * SNA is doing fancy tricks with compressing batch buffers, which leads
217 * to negative relocation deltas. Usually that works out ok since the
218 * relocate address is still positive, except when the batch is placed
219 * very low in the GTT. Ensure this doesn't happen.
221 * Note that actual hangs have only been observed on gen7, but for
222 * paranoia do it everywhere.
224 if ((vma->exec_entry->flags & EXEC_OBJECT_PINNED) == 0)
225 vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
230 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
233 if (handle >= -eb->and)
235 return eb->lut[handle];
237 struct hlist_head *head;
238 struct i915_vma *vma;
240 head = &eb->buckets[handle & eb->and];
241 hlist_for_each_entry(vma, head, exec_node) {
242 if (vma->exec_handle == handle)
250 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
252 struct drm_i915_gem_exec_object2 *entry;
253 struct drm_i915_gem_object *obj = vma->obj;
255 if (!drm_mm_node_allocated(&vma->node))
258 entry = vma->exec_entry;
260 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
261 i915_gem_object_unpin_fence(obj);
263 if (entry->flags & __EXEC_OBJECT_HAS_PIN)
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 i915_gem_object_put(vma->obj);
284 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
286 return (HAS_LLC(obj->base.dev) ||
287 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
288 obj->cache_level != I915_CACHE_NONE);
291 /* Used to convert any address to canonical form.
292 * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
293 * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
294 * addresses to be in a canonical form:
295 * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
296 * canonical form [63:48] == [47]."
298 #define GEN8_HIGH_ADDRESS_BIT 47
299 static inline uint64_t gen8_canonical_addr(uint64_t address)
301 return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
304 static inline uint64_t gen8_noncanonical_addr(uint64_t address)
306 return address & ((1ULL << (GEN8_HIGH_ADDRESS_BIT + 1)) - 1);
309 static inline uint64_t
310 relocation_target(struct drm_i915_gem_relocation_entry *reloc,
311 uint64_t target_offset)
313 return gen8_canonical_addr((int)reloc->delta + target_offset);
317 relocate_entry_cpu(struct drm_i915_gem_object *obj,
318 struct drm_i915_gem_relocation_entry *reloc,
319 uint64_t target_offset)
321 struct drm_device *dev = obj->base.dev;
322 uint32_t page_offset = offset_in_page(reloc->offset);
323 uint64_t delta = relocation_target(reloc, target_offset);
327 ret = i915_gem_object_set_to_cpu_domain(obj, true);
331 vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
332 reloc->offset >> PAGE_SHIFT));
333 *(uint32_t *)(vaddr + page_offset) = lower_32_bits(delta);
335 if (INTEL_INFO(dev)->gen >= 8) {
336 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
338 if (page_offset == 0) {
339 kunmap_atomic(vaddr);
340 vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
341 (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
344 *(uint32_t *)(vaddr + page_offset) = upper_32_bits(delta);
347 kunmap_atomic(vaddr);
353 relocate_entry_gtt(struct drm_i915_gem_object *obj,
354 struct drm_i915_gem_relocation_entry *reloc,
355 uint64_t target_offset)
357 struct drm_device *dev = obj->base.dev;
358 struct drm_i915_private *dev_priv = to_i915(dev);
359 struct i915_ggtt *ggtt = &dev_priv->ggtt;
360 uint64_t delta = relocation_target(reloc, target_offset);
362 void __iomem *reloc_page;
365 ret = i915_gem_object_set_to_gtt_domain(obj, true);
369 ret = i915_gem_object_put_fence(obj);
373 /* Map the page containing the relocation we're going to perform. */
374 offset = i915_gem_obj_ggtt_offset(obj);
375 offset += reloc->offset;
376 reloc_page = io_mapping_map_atomic_wc(ggtt->mappable,
378 iowrite32(lower_32_bits(delta), reloc_page + offset_in_page(offset));
380 if (INTEL_INFO(dev)->gen >= 8) {
381 offset += sizeof(uint32_t);
383 if (offset_in_page(offset) == 0) {
384 io_mapping_unmap_atomic(reloc_page);
386 io_mapping_map_atomic_wc(ggtt->mappable,
390 iowrite32(upper_32_bits(delta),
391 reloc_page + offset_in_page(offset));
394 io_mapping_unmap_atomic(reloc_page);
400 clflush_write32(void *addr, uint32_t value)
402 /* This is not a fast path, so KISS. */
403 drm_clflush_virt_range(addr, sizeof(uint32_t));
404 *(uint32_t *)addr = value;
405 drm_clflush_virt_range(addr, sizeof(uint32_t));
409 relocate_entry_clflush(struct drm_i915_gem_object *obj,
410 struct drm_i915_gem_relocation_entry *reloc,
411 uint64_t target_offset)
413 struct drm_device *dev = obj->base.dev;
414 uint32_t page_offset = offset_in_page(reloc->offset);
415 uint64_t delta = relocation_target(reloc, target_offset);
419 ret = i915_gem_object_set_to_gtt_domain(obj, true);
423 vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
424 reloc->offset >> PAGE_SHIFT));
425 clflush_write32(vaddr + page_offset, lower_32_bits(delta));
427 if (INTEL_INFO(dev)->gen >= 8) {
428 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
430 if (page_offset == 0) {
431 kunmap_atomic(vaddr);
432 vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
433 (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
436 clflush_write32(vaddr + page_offset, upper_32_bits(delta));
439 kunmap_atomic(vaddr);
444 static bool object_is_idle(struct drm_i915_gem_object *obj)
446 unsigned long active = obj->active;
449 for_each_active(active, idx) {
450 if (!i915_gem_active_is_idle(&obj->last_read[idx],
451 &obj->base.dev->struct_mutex))
459 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
461 struct drm_i915_gem_relocation_entry *reloc)
463 struct drm_device *dev = obj->base.dev;
464 struct drm_gem_object *target_obj;
465 struct drm_i915_gem_object *target_i915_obj;
466 struct i915_vma *target_vma;
467 uint64_t target_offset;
470 /* we've already hold a reference to all valid objects */
471 target_vma = eb_get_vma(eb, reloc->target_handle);
472 if (unlikely(target_vma == NULL))
474 target_i915_obj = target_vma->obj;
475 target_obj = &target_vma->obj->base;
477 target_offset = gen8_canonical_addr(target_vma->node.start);
479 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
480 * pipe_control writes because the gpu doesn't properly redirect them
481 * through the ppgtt for non_secure batchbuffers. */
482 if (unlikely(IS_GEN6(dev) &&
483 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
484 ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
486 if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
490 /* Validate that the target is in a valid r/w GPU domain */
491 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
492 DRM_DEBUG("reloc with multiple write domains: "
493 "obj %p target %d offset %d "
494 "read %08x write %08x",
495 obj, reloc->target_handle,
498 reloc->write_domain);
501 if (unlikely((reloc->write_domain | reloc->read_domains)
502 & ~I915_GEM_GPU_DOMAINS)) {
503 DRM_DEBUG("reloc with read/write non-GPU domains: "
504 "obj %p target %d offset %d "
505 "read %08x write %08x",
506 obj, reloc->target_handle,
509 reloc->write_domain);
513 target_obj->pending_read_domains |= reloc->read_domains;
514 target_obj->pending_write_domain |= reloc->write_domain;
516 /* If the relocation already has the right value in it, no
517 * more work needs to be done.
519 if (target_offset == reloc->presumed_offset)
522 /* Check that the relocation address is valid... */
523 if (unlikely(reloc->offset >
524 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
525 DRM_DEBUG("Relocation beyond object bounds: "
526 "obj %p target %d offset %d size %d.\n",
527 obj, reloc->target_handle,
529 (int) obj->base.size);
532 if (unlikely(reloc->offset & 3)) {
533 DRM_DEBUG("Relocation not 4-byte aligned: "
534 "obj %p target %d offset %d.\n",
535 obj, reloc->target_handle,
536 (int) reloc->offset);
540 /* We can't wait for rendering with pagefaults disabled */
541 if (pagefault_disabled() && !object_is_idle(obj))
544 if (use_cpu_reloc(obj))
545 ret = relocate_entry_cpu(obj, reloc, target_offset);
546 else if (obj->map_and_fenceable)
547 ret = relocate_entry_gtt(obj, reloc, target_offset);
548 else if (static_cpu_has(X86_FEATURE_CLFLUSH))
549 ret = relocate_entry_clflush(obj, reloc, target_offset);
551 WARN_ONCE(1, "Impossible case in relocation handling\n");
558 /* and update the user's relocation entry */
559 reloc->presumed_offset = target_offset;
565 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
568 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
569 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
570 struct drm_i915_gem_relocation_entry __user *user_relocs;
571 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
574 user_relocs = u64_to_user_ptr(entry->relocs_ptr);
576 remain = entry->relocation_count;
578 struct drm_i915_gem_relocation_entry *r = stack_reloc;
580 if (count > ARRAY_SIZE(stack_reloc))
581 count = ARRAY_SIZE(stack_reloc);
584 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
588 u64 offset = r->presumed_offset;
590 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r);
594 if (r->presumed_offset != offset &&
595 __put_user(r->presumed_offset, &user_relocs->presumed_offset)) {
609 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
611 struct drm_i915_gem_relocation_entry *relocs)
613 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
616 for (i = 0; i < entry->relocation_count; i++) {
617 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i]);
626 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
628 struct i915_vma *vma;
631 /* This is the fast path and we cannot handle a pagefault whilst
632 * holding the struct mutex lest the user pass in the relocations
633 * contained within a mmaped bo. For in such a case we, the page
634 * fault handler would call i915_gem_fault() and we would try to
635 * acquire the struct mutex again. Obviously this is bad and so
636 * lockdep complains vehemently.
639 list_for_each_entry(vma, &eb->vmas, exec_list) {
640 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
649 static bool only_mappable_for_reloc(unsigned int flags)
651 return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
652 __EXEC_OBJECT_NEEDS_MAP;
656 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
657 struct intel_engine_cs *engine,
660 struct drm_i915_gem_object *obj = vma->obj;
661 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
666 if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
669 if (!drm_mm_node_allocated(&vma->node)) {
670 /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
671 * limit address to the first 4GBs for unflagged objects.
673 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
674 flags |= PIN_ZONE_4G;
675 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
676 flags |= PIN_GLOBAL | PIN_MAPPABLE;
677 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
678 flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
679 if (entry->flags & EXEC_OBJECT_PINNED)
680 flags |= entry->offset | PIN_OFFSET_FIXED;
681 if ((flags & PIN_MAPPABLE) == 0)
685 ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
686 if ((ret == -ENOSPC || ret == -E2BIG) &&
687 only_mappable_for_reloc(entry->flags))
688 ret = i915_gem_object_pin(obj, vma->vm,
690 flags & ~PIN_MAPPABLE);
694 entry->flags |= __EXEC_OBJECT_HAS_PIN;
696 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
697 ret = i915_gem_object_get_fence(obj);
701 if (i915_gem_object_pin_fence(obj))
702 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
705 if (entry->offset != vma->node.start) {
706 entry->offset = vma->node.start;
710 if (entry->flags & EXEC_OBJECT_WRITE) {
711 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
712 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
719 need_reloc_mappable(struct i915_vma *vma)
721 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
723 if (entry->relocation_count == 0)
729 /* See also use_cpu_reloc() */
730 if (HAS_LLC(vma->obj->base.dev))
733 if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
740 eb_vma_misplaced(struct i915_vma *vma)
742 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
743 struct drm_i915_gem_object *obj = vma->obj;
745 WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP && !vma->is_ggtt);
747 if (entry->alignment &&
748 vma->node.start & (entry->alignment - 1))
751 if (entry->flags & EXEC_OBJECT_PINNED &&
752 vma->node.start != entry->offset)
755 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
756 vma->node.start < BATCH_OFFSET_BIAS)
759 /* avoid costly ping-pong once a batch bo ended up non-mappable */
760 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
761 return !only_mappable_for_reloc(entry->flags);
763 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
764 (vma->node.start + vma->node.size - 1) >> 32)
771 i915_gem_execbuffer_reserve(struct intel_engine_cs *engine,
772 struct list_head *vmas,
773 struct i915_gem_context *ctx,
776 struct drm_i915_gem_object *obj;
777 struct i915_vma *vma;
778 struct i915_address_space *vm;
779 struct list_head ordered_vmas;
780 struct list_head pinned_vmas;
781 bool has_fenced_gpu_access = INTEL_GEN(engine->i915) < 4;
784 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
786 INIT_LIST_HEAD(&ordered_vmas);
787 INIT_LIST_HEAD(&pinned_vmas);
788 while (!list_empty(vmas)) {
789 struct drm_i915_gem_exec_object2 *entry;
790 bool need_fence, need_mappable;
792 vma = list_first_entry(vmas, struct i915_vma, exec_list);
794 entry = vma->exec_entry;
796 if (ctx->flags & CONTEXT_NO_ZEROMAP)
797 entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
799 if (!has_fenced_gpu_access)
800 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
802 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
803 obj->tiling_mode != I915_TILING_NONE;
804 need_mappable = need_fence || need_reloc_mappable(vma);
806 if (entry->flags & EXEC_OBJECT_PINNED)
807 list_move_tail(&vma->exec_list, &pinned_vmas);
808 else if (need_mappable) {
809 entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
810 list_move(&vma->exec_list, &ordered_vmas);
812 list_move_tail(&vma->exec_list, &ordered_vmas);
814 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
815 obj->base.pending_write_domain = 0;
817 list_splice(&ordered_vmas, vmas);
818 list_splice(&pinned_vmas, vmas);
820 /* Attempt to pin all of the buffers into the GTT.
821 * This is done in 3 phases:
823 * 1a. Unbind all objects that do not match the GTT constraints for
824 * the execbuffer (fenceable, mappable, alignment etc).
825 * 1b. Increment pin count for already bound objects.
826 * 2. Bind new objects.
827 * 3. Decrement pin count.
829 * This avoid unnecessary unbinding of later objects in order to make
830 * room for the earlier objects *unless* we need to defragment.
836 /* Unbind any ill-fitting objects or pin. */
837 list_for_each_entry(vma, vmas, exec_list) {
838 if (!drm_mm_node_allocated(&vma->node))
841 if (eb_vma_misplaced(vma))
842 ret = i915_vma_unbind(vma);
844 ret = i915_gem_execbuffer_reserve_vma(vma,
851 /* Bind fresh objects */
852 list_for_each_entry(vma, vmas, exec_list) {
853 if (drm_mm_node_allocated(&vma->node))
856 ret = i915_gem_execbuffer_reserve_vma(vma, engine,
863 if (ret != -ENOSPC || retry++)
866 /* Decrement pin count for bound objects */
867 list_for_each_entry(vma, vmas, exec_list)
868 i915_gem_execbuffer_unreserve_vma(vma);
870 ret = i915_gem_evict_vm(vm, true);
877 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
878 struct drm_i915_gem_execbuffer2 *args,
879 struct drm_file *file,
880 struct intel_engine_cs *engine,
882 struct drm_i915_gem_exec_object2 *exec,
883 struct i915_gem_context *ctx)
885 struct drm_i915_gem_relocation_entry *reloc;
886 struct i915_address_space *vm;
887 struct i915_vma *vma;
891 unsigned count = args->buffer_count;
893 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
895 /* We may process another execbuffer during the unlock... */
896 while (!list_empty(&eb->vmas)) {
897 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
898 list_del_init(&vma->exec_list);
899 i915_gem_execbuffer_unreserve_vma(vma);
900 i915_gem_object_put(vma->obj);
903 mutex_unlock(&dev->struct_mutex);
906 for (i = 0; i < count; i++)
907 total += exec[i].relocation_count;
909 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
910 reloc = drm_malloc_ab(total, sizeof(*reloc));
911 if (reloc == NULL || reloc_offset == NULL) {
912 drm_free_large(reloc);
913 drm_free_large(reloc_offset);
914 mutex_lock(&dev->struct_mutex);
919 for (i = 0; i < count; i++) {
920 struct drm_i915_gem_relocation_entry __user *user_relocs;
921 u64 invalid_offset = (u64)-1;
924 user_relocs = u64_to_user_ptr(exec[i].relocs_ptr);
926 if (copy_from_user(reloc+total, user_relocs,
927 exec[i].relocation_count * sizeof(*reloc))) {
929 mutex_lock(&dev->struct_mutex);
933 /* As we do not update the known relocation offsets after
934 * relocating (due to the complexities in lock handling),
935 * we need to mark them as invalid now so that we force the
936 * relocation processing next time. Just in case the target
937 * object is evicted and then rebound into its old
938 * presumed_offset before the next execbuffer - if that
939 * happened we would make the mistake of assuming that the
940 * relocations were valid.
942 for (j = 0; j < exec[i].relocation_count; j++) {
943 if (__copy_to_user(&user_relocs[j].presumed_offset,
945 sizeof(invalid_offset))) {
947 mutex_lock(&dev->struct_mutex);
952 reloc_offset[i] = total;
953 total += exec[i].relocation_count;
956 ret = i915_mutex_lock_interruptible(dev);
958 mutex_lock(&dev->struct_mutex);
962 /* reacquire the objects */
964 ret = eb_lookup_vmas(eb, exec, args, vm, file);
968 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
969 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
974 list_for_each_entry(vma, &eb->vmas, exec_list) {
975 int offset = vma->exec_entry - exec;
976 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
977 reloc + reloc_offset[offset]);
982 /* Leave the user relocations as are, this is the painfully slow path,
983 * and we want to avoid the complication of dropping the lock whilst
984 * having buffers reserved in the aperture and so causing spurious
985 * ENOSPC for random operations.
989 drm_free_large(reloc);
990 drm_free_large(reloc_offset);
995 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
996 struct list_head *vmas)
998 const unsigned other_rings = ~intel_engine_flag(req->engine);
999 struct i915_vma *vma;
1000 uint32_t flush_domains = 0;
1001 bool flush_chipset = false;
1004 list_for_each_entry(vma, vmas, exec_list) {
1005 struct drm_i915_gem_object *obj = vma->obj;
1007 if (obj->active & other_rings) {
1008 ret = i915_gem_object_sync(obj, req);
1013 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
1014 flush_chipset |= i915_gem_clflush_object(obj, false);
1016 flush_domains |= obj->base.write_domain;
1020 i915_gem_chipset_flush(req->engine->i915);
1022 if (flush_domains & I915_GEM_DOMAIN_GTT)
1025 /* Unconditionally invalidate GPU caches and TLBs. */
1026 return req->engine->emit_flush(req, EMIT_INVALIDATE);
1030 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
1032 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
1035 /* Kernel clipping was a DRI1 misfeature */
1036 if (exec->num_cliprects || exec->cliprects_ptr)
1039 if (exec->DR4 == 0xffffffff) {
1040 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1043 if (exec->DR1 || exec->DR4)
1046 if ((exec->batch_start_offset | exec->batch_len) & 0x7)
1053 validate_exec_list(struct drm_device *dev,
1054 struct drm_i915_gem_exec_object2 *exec,
1057 unsigned relocs_total = 0;
1058 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
1059 unsigned invalid_flags;
1062 /* INTERNAL flags must not overlap with external ones */
1063 BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & ~__EXEC_OBJECT_UNKNOWN_FLAGS);
1065 invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
1066 if (USES_FULL_PPGTT(dev))
1067 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
1069 for (i = 0; i < count; i++) {
1070 char __user *ptr = u64_to_user_ptr(exec[i].relocs_ptr);
1071 int length; /* limited by fault_in_pages_readable() */
1073 if (exec[i].flags & invalid_flags)
1076 /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
1077 * any non-page-aligned or non-canonical addresses.
1079 if (exec[i].flags & EXEC_OBJECT_PINNED) {
1080 if (exec[i].offset !=
1081 gen8_canonical_addr(exec[i].offset & PAGE_MASK))
1084 /* From drm_mm perspective address space is continuous,
1085 * so from this point we're always using non-canonical
1088 exec[i].offset = gen8_noncanonical_addr(exec[i].offset);
1091 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
1094 /* First check for malicious input causing overflow in
1095 * the worst case where we need to allocate the entire
1096 * relocation tree as a single array.
1098 if (exec[i].relocation_count > relocs_max - relocs_total)
1100 relocs_total += exec[i].relocation_count;
1102 length = exec[i].relocation_count *
1103 sizeof(struct drm_i915_gem_relocation_entry);
1105 * We must check that the entire relocation array is safe
1106 * to read, but since we may need to update the presumed
1107 * offsets during execution, check for full write access.
1109 if (!access_ok(VERIFY_WRITE, ptr, length))
1112 if (likely(!i915.prefault_disable)) {
1113 if (fault_in_multipages_readable(ptr, length))
1121 static struct i915_gem_context *
1122 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
1123 struct intel_engine_cs *engine, const u32 ctx_id)
1125 struct i915_gem_context *ctx = NULL;
1126 struct i915_ctx_hang_stats *hs;
1128 if (engine->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
1129 return ERR_PTR(-EINVAL);
1131 ctx = i915_gem_context_lookup(file->driver_priv, ctx_id);
1135 hs = &ctx->hang_stats;
1137 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1138 return ERR_PTR(-EIO);
1144 void i915_vma_move_to_active(struct i915_vma *vma,
1145 struct drm_i915_gem_request *req,
1148 struct drm_i915_gem_object *obj = vma->obj;
1149 const unsigned int idx = req->engine->id;
1151 GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
1153 obj->dirty = 1; /* be paranoid */
1155 /* Add a reference if we're newly entering the active list.
1156 * The order in which we add operations to the retirement queue is
1157 * vital here: mark_active adds to the start of the callback list,
1158 * such that subsequent callbacks are called first. Therefore we
1159 * add the active reference first and queue for it to be dropped
1162 if (obj->active == 0)
1163 i915_gem_object_get(obj);
1164 obj->active |= 1 << idx;
1165 i915_gem_active_set(&obj->last_read[idx], req);
1167 if (flags & EXEC_OBJECT_WRITE) {
1168 i915_gem_active_set(&obj->last_write, req);
1170 intel_fb_obj_invalidate(obj, ORIGIN_CS);
1172 /* update for the implicit flush after a batch */
1173 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1176 if (flags & EXEC_OBJECT_NEEDS_FENCE) {
1177 i915_gem_active_set(&obj->last_fence, req);
1178 if (flags & __EXEC_OBJECT_HAS_FENCE) {
1179 struct drm_i915_private *dev_priv = req->i915;
1181 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
1182 &dev_priv->mm.fence_list);
1186 i915_vma_set_active(vma, idx);
1187 i915_gem_active_set(&vma->last_read[idx], req);
1188 list_move_tail(&vma->vm_link, &vma->vm->active_list);
1192 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1193 struct drm_i915_gem_request *req)
1195 struct i915_vma *vma;
1197 list_for_each_entry(vma, vmas, exec_list) {
1198 struct drm_i915_gem_object *obj = vma->obj;
1199 u32 old_read = obj->base.read_domains;
1200 u32 old_write = obj->base.write_domain;
1202 obj->base.write_domain = obj->base.pending_write_domain;
1203 if (obj->base.write_domain)
1204 vma->exec_entry->flags |= EXEC_OBJECT_WRITE;
1206 obj->base.pending_read_domains |= obj->base.read_domains;
1207 obj->base.read_domains = obj->base.pending_read_domains;
1209 i915_vma_move_to_active(vma, req, vma->exec_entry->flags);
1210 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1215 i915_reset_gen7_sol_offsets(struct drm_i915_gem_request *req)
1217 struct intel_ring *ring = req->ring;
1220 if (!IS_GEN7(req->i915) || req->engine->id != RCS) {
1221 DRM_DEBUG("sol reset is gen7/rcs only\n");
1225 ret = intel_ring_begin(req, 4 * 3);
1229 for (i = 0; i < 4; i++) {
1230 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1231 intel_ring_emit_reg(ring, GEN7_SO_WRITE_OFFSET(i));
1232 intel_ring_emit(ring, 0);
1235 intel_ring_advance(ring);
1240 static struct drm_i915_gem_object*
1241 i915_gem_execbuffer_parse(struct intel_engine_cs *engine,
1242 struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1244 struct drm_i915_gem_object *batch_obj,
1245 u32 batch_start_offset,
1249 struct drm_i915_gem_object *shadow_batch_obj;
1250 struct i915_vma *vma;
1253 shadow_batch_obj = i915_gem_batch_pool_get(&engine->batch_pool,
1254 PAGE_ALIGN(batch_len));
1255 if (IS_ERR(shadow_batch_obj))
1256 return shadow_batch_obj;
1258 ret = intel_engine_cmd_parser(engine,
1267 ret = i915_gem_obj_ggtt_pin(shadow_batch_obj, 0, 0);
1271 i915_gem_object_unpin_pages(shadow_batch_obj);
1273 memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1275 vma = i915_gem_obj_to_ggtt(shadow_batch_obj);
1276 vma->exec_entry = shadow_exec_entry;
1277 vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1278 i915_gem_object_get(shadow_batch_obj);
1279 list_add_tail(&vma->exec_list, &eb->vmas);
1281 shadow_batch_obj->base.pending_read_domains = I915_GEM_DOMAIN_COMMAND;
1283 return shadow_batch_obj;
1286 i915_gem_object_unpin_pages(shadow_batch_obj);
1287 if (ret == -EACCES) /* unhandled chained batch */
1290 return ERR_PTR(ret);
1294 execbuf_submit(struct i915_execbuffer_params *params,
1295 struct drm_i915_gem_execbuffer2 *args,
1296 struct list_head *vmas)
1298 struct drm_i915_private *dev_priv = params->request->i915;
1299 u64 exec_start, exec_len;
1304 ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1308 ret = i915_switch_context(params->request);
1312 instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1313 instp_mask = I915_EXEC_CONSTANTS_MASK;
1314 switch (instp_mode) {
1315 case I915_EXEC_CONSTANTS_REL_GENERAL:
1316 case I915_EXEC_CONSTANTS_ABSOLUTE:
1317 case I915_EXEC_CONSTANTS_REL_SURFACE:
1318 if (instp_mode != 0 && params->engine->id != RCS) {
1319 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1323 if (instp_mode != dev_priv->relative_constants_mode) {
1324 if (INTEL_INFO(dev_priv)->gen < 4) {
1325 DRM_DEBUG("no rel constants on pre-gen4\n");
1329 if (INTEL_INFO(dev_priv)->gen > 5 &&
1330 instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1331 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1335 /* The HW changed the meaning on this bit on gen6 */
1336 if (INTEL_INFO(dev_priv)->gen >= 6)
1337 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1341 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1345 if (params->engine->id == RCS &&
1346 instp_mode != dev_priv->relative_constants_mode) {
1347 struct intel_ring *ring = params->request->ring;
1349 ret = intel_ring_begin(params->request, 4);
1353 intel_ring_emit(ring, MI_NOOP);
1354 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1355 intel_ring_emit_reg(ring, INSTPM);
1356 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1357 intel_ring_advance(ring);
1359 dev_priv->relative_constants_mode = instp_mode;
1362 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1363 ret = i915_reset_gen7_sol_offsets(params->request);
1368 exec_len = args->batch_len;
1369 exec_start = params->batch_obj_vm_offset +
1370 params->args_batch_start_offset;
1373 exec_len = params->batch_obj->base.size;
1375 ret = params->engine->emit_bb_start(params->request,
1376 exec_start, exec_len,
1377 params->dispatch_flags);
1381 trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1383 i915_gem_execbuffer_move_to_active(vmas, params->request);
1389 * Find one BSD ring to dispatch the corresponding BSD command.
1390 * The engine index is returned.
1393 gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
1394 struct drm_file *file)
1396 struct drm_i915_file_private *file_priv = file->driver_priv;
1398 /* Check whether the file_priv has already selected one ring. */
1399 if ((int)file_priv->bsd_engine < 0) {
1400 /* If not, use the ping-pong mechanism to select one. */
1401 mutex_lock(&dev_priv->drm.struct_mutex);
1402 file_priv->bsd_engine = dev_priv->mm.bsd_engine_dispatch_index;
1403 dev_priv->mm.bsd_engine_dispatch_index ^= 1;
1404 mutex_unlock(&dev_priv->drm.struct_mutex);
1407 return file_priv->bsd_engine;
1410 #define I915_USER_RINGS (4)
1412 static const enum intel_engine_id user_ring_map[I915_USER_RINGS + 1] = {
1413 [I915_EXEC_DEFAULT] = RCS,
1414 [I915_EXEC_RENDER] = RCS,
1415 [I915_EXEC_BLT] = BCS,
1416 [I915_EXEC_BSD] = VCS,
1417 [I915_EXEC_VEBOX] = VECS
1420 static struct intel_engine_cs *
1421 eb_select_engine(struct drm_i915_private *dev_priv,
1422 struct drm_file *file,
1423 struct drm_i915_gem_execbuffer2 *args)
1425 unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
1426 struct intel_engine_cs *engine;
1428 if (user_ring_id > I915_USER_RINGS) {
1429 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id);
1433 if ((user_ring_id != I915_EXEC_BSD) &&
1434 ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1435 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1436 "bsd dispatch flags: %d\n", (int)(args->flags));
1440 if (user_ring_id == I915_EXEC_BSD && HAS_BSD2(dev_priv)) {
1441 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
1443 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
1444 bsd_idx = gen8_dispatch_bsd_engine(dev_priv, file);
1445 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
1446 bsd_idx <= I915_EXEC_BSD_RING2) {
1447 bsd_idx >>= I915_EXEC_BSD_SHIFT;
1450 DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
1455 engine = &dev_priv->engine[_VCS(bsd_idx)];
1457 engine = &dev_priv->engine[user_ring_map[user_ring_id]];
1460 if (!intel_engine_initialized(engine)) {
1461 DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id);
1469 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1470 struct drm_file *file,
1471 struct drm_i915_gem_execbuffer2 *args,
1472 struct drm_i915_gem_exec_object2 *exec)
1474 struct drm_i915_private *dev_priv = to_i915(dev);
1475 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1477 struct drm_i915_gem_object *batch_obj;
1478 struct drm_i915_gem_exec_object2 shadow_exec_entry;
1479 struct intel_engine_cs *engine;
1480 struct i915_gem_context *ctx;
1481 struct i915_address_space *vm;
1482 struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1483 struct i915_execbuffer_params *params = ¶ms_master;
1484 const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1489 if (!i915_gem_check_execbuffer(args))
1492 ret = validate_exec_list(dev, exec, args->buffer_count);
1497 if (args->flags & I915_EXEC_SECURE) {
1498 if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
1501 dispatch_flags |= I915_DISPATCH_SECURE;
1503 if (args->flags & I915_EXEC_IS_PINNED)
1504 dispatch_flags |= I915_DISPATCH_PINNED;
1506 engine = eb_select_engine(dev_priv, file, args);
1510 if (args->buffer_count < 1) {
1511 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1515 if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1516 if (!HAS_RESOURCE_STREAMER(dev)) {
1517 DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1520 if (engine->id != RCS) {
1521 DRM_DEBUG("RS is not available on %s\n",
1526 dispatch_flags |= I915_DISPATCH_RS;
1529 /* Take a local wakeref for preparing to dispatch the execbuf as
1530 * we expect to access the hardware fairly frequently in the
1531 * process. Upon first dispatch, we acquire another prolonged
1532 * wakeref that we hold until the GPU has been idle for at least
1535 intel_runtime_pm_get(dev_priv);
1537 ret = i915_mutex_lock_interruptible(dev);
1541 ctx = i915_gem_validate_context(dev, file, engine, ctx_id);
1543 mutex_unlock(&dev->struct_mutex);
1548 i915_gem_context_get(ctx);
1551 vm = &ctx->ppgtt->base;
1555 memset(¶ms_master, 0x00, sizeof(params_master));
1557 eb = eb_create(args);
1559 i915_gem_context_put(ctx);
1560 mutex_unlock(&dev->struct_mutex);
1565 /* Look up object handles */
1566 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1570 /* take note of the batch buffer before we might reorder the lists */
1571 batch_obj = eb_get_batch(eb);
1573 /* Move the objects en-masse into the GTT, evicting if necessary. */
1574 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1575 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1580 /* The objects are in their final locations, apply the relocations. */
1582 ret = i915_gem_execbuffer_relocate(eb);
1584 if (ret == -EFAULT) {
1585 ret = i915_gem_execbuffer_relocate_slow(dev, args, file,
1588 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1594 /* Set the pending read domains for the batch buffer to COMMAND */
1595 if (batch_obj->base.pending_write_domain) {
1596 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1601 params->args_batch_start_offset = args->batch_start_offset;
1602 if (intel_engine_needs_cmd_parser(engine) && args->batch_len) {
1603 struct drm_i915_gem_object *parsed_batch_obj;
1605 parsed_batch_obj = i915_gem_execbuffer_parse(engine,
1609 args->batch_start_offset,
1611 drm_is_current_master(file));
1612 if (IS_ERR(parsed_batch_obj)) {
1613 ret = PTR_ERR(parsed_batch_obj);
1618 * parsed_batch_obj == batch_obj means batch not fully parsed:
1619 * Accept, but don't promote to secure.
1622 if (parsed_batch_obj != batch_obj) {
1624 * Batch parsed and accepted:
1626 * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1627 * bit from MI_BATCH_BUFFER_START commands issued in
1628 * the dispatch_execbuffer implementations. We
1629 * specifically don't want that set on batches the
1630 * command parser has accepted.
1632 dispatch_flags |= I915_DISPATCH_SECURE;
1633 params->args_batch_start_offset = 0;
1634 batch_obj = parsed_batch_obj;
1638 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1640 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1641 * batch" bit. Hence we need to pin secure batches into the global gtt.
1642 * hsw should have this fixed, but bdw mucks it up again. */
1643 if (dispatch_flags & I915_DISPATCH_SECURE) {
1645 * So on first glance it looks freaky that we pin the batch here
1646 * outside of the reservation loop. But:
1647 * - The batch is already pinned into the relevant ppgtt, so we
1648 * already have the backing storage fully allocated.
1649 * - No other BO uses the global gtt (well contexts, but meh),
1650 * so we don't really have issues with multiple objects not
1651 * fitting due to fragmentation.
1652 * So this is actually safe.
1654 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1658 params->batch_obj_vm_offset = i915_gem_obj_ggtt_offset(batch_obj);
1660 params->batch_obj_vm_offset = i915_gem_obj_offset(batch_obj, vm);
1662 /* Allocate a request for this batch buffer nice and early. */
1663 params->request = i915_gem_request_alloc(engine, ctx);
1664 if (IS_ERR(params->request)) {
1665 ret = PTR_ERR(params->request);
1666 goto err_batch_unpin;
1669 ret = i915_gem_request_add_to_client(params->request, file);
1674 * Save assorted stuff away to pass through to *_submission().
1675 * NB: This data should be 'persistent' and not local as it will
1676 * kept around beyond the duration of the IOCTL once the GPU
1677 * scheduler arrives.
1680 params->file = file;
1681 params->engine = engine;
1682 params->dispatch_flags = dispatch_flags;
1683 params->batch_obj = batch_obj;
1686 ret = execbuf_submit(params, args, &eb->vmas);
1688 __i915_add_request(params->request, params->batch_obj, ret == 0);
1692 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1693 * batch vma for correctness. For less ugly and less fragility this
1694 * needs to be adjusted to also track the ggtt batch vma properly as
1697 if (dispatch_flags & I915_DISPATCH_SECURE)
1698 i915_gem_object_ggtt_unpin(batch_obj);
1701 /* the request owns the ref now */
1702 i915_gem_context_put(ctx);
1705 mutex_unlock(&dev->struct_mutex);
1708 /* intel_gpu_busy should also get a ref, so it will free when the device
1709 * is really idle. */
1710 intel_runtime_pm_put(dev_priv);
1715 * Legacy execbuffer just creates an exec2 list from the original exec object
1716 * list array and passes it to the real function.
1719 i915_gem_execbuffer(struct drm_device *dev, void *data,
1720 struct drm_file *file)
1722 struct drm_i915_gem_execbuffer *args = data;
1723 struct drm_i915_gem_execbuffer2 exec2;
1724 struct drm_i915_gem_exec_object *exec_list = NULL;
1725 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1728 if (args->buffer_count < 1) {
1729 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1733 /* Copy in the exec list from userland */
1734 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1735 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1736 if (exec_list == NULL || exec2_list == NULL) {
1737 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1738 args->buffer_count);
1739 drm_free_large(exec_list);
1740 drm_free_large(exec2_list);
1743 ret = copy_from_user(exec_list,
1744 u64_to_user_ptr(args->buffers_ptr),
1745 sizeof(*exec_list) * args->buffer_count);
1747 DRM_DEBUG("copy %d exec entries failed %d\n",
1748 args->buffer_count, ret);
1749 drm_free_large(exec_list);
1750 drm_free_large(exec2_list);
1754 for (i = 0; i < args->buffer_count; i++) {
1755 exec2_list[i].handle = exec_list[i].handle;
1756 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1757 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1758 exec2_list[i].alignment = exec_list[i].alignment;
1759 exec2_list[i].offset = exec_list[i].offset;
1760 if (INTEL_INFO(dev)->gen < 4)
1761 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1763 exec2_list[i].flags = 0;
1766 exec2.buffers_ptr = args->buffers_ptr;
1767 exec2.buffer_count = args->buffer_count;
1768 exec2.batch_start_offset = args->batch_start_offset;
1769 exec2.batch_len = args->batch_len;
1770 exec2.DR1 = args->DR1;
1771 exec2.DR4 = args->DR4;
1772 exec2.num_cliprects = args->num_cliprects;
1773 exec2.cliprects_ptr = args->cliprects_ptr;
1774 exec2.flags = I915_EXEC_RENDER;
1775 i915_execbuffer2_set_context_id(exec2, 0);
1777 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1779 struct drm_i915_gem_exec_object __user *user_exec_list =
1780 u64_to_user_ptr(args->buffers_ptr);
1782 /* Copy the new buffer offsets back to the user's exec list. */
1783 for (i = 0; i < args->buffer_count; i++) {
1784 exec2_list[i].offset =
1785 gen8_canonical_addr(exec2_list[i].offset);
1786 ret = __copy_to_user(&user_exec_list[i].offset,
1787 &exec2_list[i].offset,
1788 sizeof(user_exec_list[i].offset));
1791 DRM_DEBUG("failed to copy %d exec entries "
1792 "back to user (%d)\n",
1793 args->buffer_count, ret);
1799 drm_free_large(exec_list);
1800 drm_free_large(exec2_list);
1805 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1806 struct drm_file *file)
1808 struct drm_i915_gem_execbuffer2 *args = data;
1809 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1812 if (args->buffer_count < 1 ||
1813 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1814 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1818 if (args->rsvd2 != 0) {
1819 DRM_DEBUG("dirty rvsd2 field\n");
1823 exec2_list = drm_malloc_gfp(args->buffer_count,
1824 sizeof(*exec2_list),
1826 if (exec2_list == NULL) {
1827 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1828 args->buffer_count);
1831 ret = copy_from_user(exec2_list,
1832 u64_to_user_ptr(args->buffers_ptr),
1833 sizeof(*exec2_list) * args->buffer_count);
1835 DRM_DEBUG("copy %d exec entries failed %d\n",
1836 args->buffer_count, ret);
1837 drm_free_large(exec2_list);
1841 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1843 /* Copy the new buffer offsets back to the user's exec list. */
1844 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1845 u64_to_user_ptr(args->buffers_ptr);
1848 for (i = 0; i < args->buffer_count; i++) {
1849 exec2_list[i].offset =
1850 gen8_canonical_addr(exec2_list[i].offset);
1851 ret = __copy_to_user(&user_exec_list[i].offset,
1852 &exec2_list[i].offset,
1853 sizeof(user_exec_list[i].offset));
1856 DRM_DEBUG("failed to copy %d exec entries "
1858 args->buffer_count);
1864 drm_free_large(exec2_list);