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 i915_gem_retire_requests_ring(engine);
786 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
788 INIT_LIST_HEAD(&ordered_vmas);
789 INIT_LIST_HEAD(&pinned_vmas);
790 while (!list_empty(vmas)) {
791 struct drm_i915_gem_exec_object2 *entry;
792 bool need_fence, need_mappable;
794 vma = list_first_entry(vmas, struct i915_vma, exec_list);
796 entry = vma->exec_entry;
798 if (ctx->flags & CONTEXT_NO_ZEROMAP)
799 entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
801 if (!has_fenced_gpu_access)
802 entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
804 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
805 obj->tiling_mode != I915_TILING_NONE;
806 need_mappable = need_fence || need_reloc_mappable(vma);
808 if (entry->flags & EXEC_OBJECT_PINNED)
809 list_move_tail(&vma->exec_list, &pinned_vmas);
810 else if (need_mappable) {
811 entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
812 list_move(&vma->exec_list, &ordered_vmas);
814 list_move_tail(&vma->exec_list, &ordered_vmas);
816 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
817 obj->base.pending_write_domain = 0;
819 list_splice(&ordered_vmas, vmas);
820 list_splice(&pinned_vmas, vmas);
822 /* Attempt to pin all of the buffers into the GTT.
823 * This is done in 3 phases:
825 * 1a. Unbind all objects that do not match the GTT constraints for
826 * the execbuffer (fenceable, mappable, alignment etc).
827 * 1b. Increment pin count for already bound objects.
828 * 2. Bind new objects.
829 * 3. Decrement pin count.
831 * This avoid unnecessary unbinding of later objects in order to make
832 * room for the earlier objects *unless* we need to defragment.
838 /* Unbind any ill-fitting objects or pin. */
839 list_for_each_entry(vma, vmas, exec_list) {
840 if (!drm_mm_node_allocated(&vma->node))
843 if (eb_vma_misplaced(vma))
844 ret = i915_vma_unbind(vma);
846 ret = i915_gem_execbuffer_reserve_vma(vma,
853 /* Bind fresh objects */
854 list_for_each_entry(vma, vmas, exec_list) {
855 if (drm_mm_node_allocated(&vma->node))
858 ret = i915_gem_execbuffer_reserve_vma(vma, engine,
865 if (ret != -ENOSPC || retry++)
868 /* Decrement pin count for bound objects */
869 list_for_each_entry(vma, vmas, exec_list)
870 i915_gem_execbuffer_unreserve_vma(vma);
872 ret = i915_gem_evict_vm(vm, true);
879 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
880 struct drm_i915_gem_execbuffer2 *args,
881 struct drm_file *file,
882 struct intel_engine_cs *engine,
884 struct drm_i915_gem_exec_object2 *exec,
885 struct i915_gem_context *ctx)
887 struct drm_i915_gem_relocation_entry *reloc;
888 struct i915_address_space *vm;
889 struct i915_vma *vma;
893 unsigned count = args->buffer_count;
895 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
897 /* We may process another execbuffer during the unlock... */
898 while (!list_empty(&eb->vmas)) {
899 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
900 list_del_init(&vma->exec_list);
901 i915_gem_execbuffer_unreserve_vma(vma);
902 i915_gem_object_put(vma->obj);
905 mutex_unlock(&dev->struct_mutex);
908 for (i = 0; i < count; i++)
909 total += exec[i].relocation_count;
911 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
912 reloc = drm_malloc_ab(total, sizeof(*reloc));
913 if (reloc == NULL || reloc_offset == NULL) {
914 drm_free_large(reloc);
915 drm_free_large(reloc_offset);
916 mutex_lock(&dev->struct_mutex);
921 for (i = 0; i < count; i++) {
922 struct drm_i915_gem_relocation_entry __user *user_relocs;
923 u64 invalid_offset = (u64)-1;
926 user_relocs = u64_to_user_ptr(exec[i].relocs_ptr);
928 if (copy_from_user(reloc+total, user_relocs,
929 exec[i].relocation_count * sizeof(*reloc))) {
931 mutex_lock(&dev->struct_mutex);
935 /* As we do not update the known relocation offsets after
936 * relocating (due to the complexities in lock handling),
937 * we need to mark them as invalid now so that we force the
938 * relocation processing next time. Just in case the target
939 * object is evicted and then rebound into its old
940 * presumed_offset before the next execbuffer - if that
941 * happened we would make the mistake of assuming that the
942 * relocations were valid.
944 for (j = 0; j < exec[i].relocation_count; j++) {
945 if (__copy_to_user(&user_relocs[j].presumed_offset,
947 sizeof(invalid_offset))) {
949 mutex_lock(&dev->struct_mutex);
954 reloc_offset[i] = total;
955 total += exec[i].relocation_count;
958 ret = i915_mutex_lock_interruptible(dev);
960 mutex_lock(&dev->struct_mutex);
964 /* reacquire the objects */
966 ret = eb_lookup_vmas(eb, exec, args, vm, file);
970 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
971 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
976 list_for_each_entry(vma, &eb->vmas, exec_list) {
977 int offset = vma->exec_entry - exec;
978 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
979 reloc + reloc_offset[offset]);
984 /* Leave the user relocations as are, this is the painfully slow path,
985 * and we want to avoid the complication of dropping the lock whilst
986 * having buffers reserved in the aperture and so causing spurious
987 * ENOSPC for random operations.
991 drm_free_large(reloc);
992 drm_free_large(reloc_offset);
997 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
998 struct list_head *vmas)
1000 const unsigned other_rings = ~intel_engine_flag(req->engine);
1001 struct i915_vma *vma;
1002 uint32_t flush_domains = 0;
1003 bool flush_chipset = false;
1006 list_for_each_entry(vma, vmas, exec_list) {
1007 struct drm_i915_gem_object *obj = vma->obj;
1009 if (obj->active & other_rings) {
1010 ret = i915_gem_object_sync(obj, req);
1015 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
1016 flush_chipset |= i915_gem_clflush_object(obj, false);
1018 flush_domains |= obj->base.write_domain;
1022 i915_gem_chipset_flush(req->engine->i915);
1024 if (flush_domains & I915_GEM_DOMAIN_GTT)
1027 /* Unconditionally invalidate GPU caches and TLBs. */
1028 return req->engine->emit_flush(req, EMIT_INVALIDATE);
1032 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
1034 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
1037 /* Kernel clipping was a DRI1 misfeature */
1038 if (exec->num_cliprects || exec->cliprects_ptr)
1041 if (exec->DR4 == 0xffffffff) {
1042 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1045 if (exec->DR1 || exec->DR4)
1048 if ((exec->batch_start_offset | exec->batch_len) & 0x7)
1055 validate_exec_list(struct drm_device *dev,
1056 struct drm_i915_gem_exec_object2 *exec,
1059 unsigned relocs_total = 0;
1060 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
1061 unsigned invalid_flags;
1064 /* INTERNAL flags must not overlap with external ones */
1065 BUILD_BUG_ON(__EXEC_OBJECT_INTERNAL_FLAGS & ~__EXEC_OBJECT_UNKNOWN_FLAGS);
1067 invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
1068 if (USES_FULL_PPGTT(dev))
1069 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
1071 for (i = 0; i < count; i++) {
1072 char __user *ptr = u64_to_user_ptr(exec[i].relocs_ptr);
1073 int length; /* limited by fault_in_pages_readable() */
1075 if (exec[i].flags & invalid_flags)
1078 /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
1079 * any non-page-aligned or non-canonical addresses.
1081 if (exec[i].flags & EXEC_OBJECT_PINNED) {
1082 if (exec[i].offset !=
1083 gen8_canonical_addr(exec[i].offset & PAGE_MASK))
1086 /* From drm_mm perspective address space is continuous,
1087 * so from this point we're always using non-canonical
1090 exec[i].offset = gen8_noncanonical_addr(exec[i].offset);
1093 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
1096 /* First check for malicious input causing overflow in
1097 * the worst case where we need to allocate the entire
1098 * relocation tree as a single array.
1100 if (exec[i].relocation_count > relocs_max - relocs_total)
1102 relocs_total += exec[i].relocation_count;
1104 length = exec[i].relocation_count *
1105 sizeof(struct drm_i915_gem_relocation_entry);
1107 * We must check that the entire relocation array is safe
1108 * to read, but since we may need to update the presumed
1109 * offsets during execution, check for full write access.
1111 if (!access_ok(VERIFY_WRITE, ptr, length))
1114 if (likely(!i915.prefault_disable)) {
1115 if (fault_in_multipages_readable(ptr, length))
1123 static struct i915_gem_context *
1124 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
1125 struct intel_engine_cs *engine, const u32 ctx_id)
1127 struct i915_gem_context *ctx = NULL;
1128 struct i915_ctx_hang_stats *hs;
1130 if (engine->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
1131 return ERR_PTR(-EINVAL);
1133 ctx = i915_gem_context_lookup(file->driver_priv, ctx_id);
1137 hs = &ctx->hang_stats;
1139 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1140 return ERR_PTR(-EIO);
1147 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1148 struct drm_i915_gem_request *req)
1150 struct intel_engine_cs *engine = i915_gem_request_get_engine(req);
1151 struct i915_vma *vma;
1153 list_for_each_entry(vma, vmas, exec_list) {
1154 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
1155 struct drm_i915_gem_object *obj = vma->obj;
1156 u32 old_read = obj->base.read_domains;
1157 u32 old_write = obj->base.write_domain;
1159 obj->dirty = 1; /* be paranoid */
1160 obj->base.write_domain = obj->base.pending_write_domain;
1161 if (obj->base.write_domain == 0)
1162 obj->base.pending_read_domains |= obj->base.read_domains;
1163 obj->base.read_domains = obj->base.pending_read_domains;
1165 i915_vma_move_to_active(vma, req);
1166 if (obj->base.write_domain) {
1167 i915_gem_active_set(&obj->last_write, req);
1169 intel_fb_obj_invalidate(obj, ORIGIN_CS);
1171 /* update for the implicit flush after a batch */
1172 obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1174 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
1175 i915_gem_active_set(&obj->last_fence, req);
1176 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
1177 struct drm_i915_private *dev_priv = engine->i915;
1178 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
1179 &dev_priv->mm.fence_list);
1183 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1188 i915_gem_execbuffer_retire_commands(struct i915_execbuffer_params *params)
1190 /* Add a breadcrumb for the completion of the batch buffer */
1191 __i915_add_request(params->request, params->batch_obj, true);
1195 i915_reset_gen7_sol_offsets(struct drm_i915_gem_request *req)
1197 struct intel_ring *ring = req->ring;
1200 if (!IS_GEN7(req->i915) || req->engine->id != RCS) {
1201 DRM_DEBUG("sol reset is gen7/rcs only\n");
1205 ret = intel_ring_begin(req, 4 * 3);
1209 for (i = 0; i < 4; i++) {
1210 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1211 intel_ring_emit_reg(ring, GEN7_SO_WRITE_OFFSET(i));
1212 intel_ring_emit(ring, 0);
1215 intel_ring_advance(ring);
1220 static struct drm_i915_gem_object*
1221 i915_gem_execbuffer_parse(struct intel_engine_cs *engine,
1222 struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1224 struct drm_i915_gem_object *batch_obj,
1225 u32 batch_start_offset,
1229 struct drm_i915_gem_object *shadow_batch_obj;
1230 struct i915_vma *vma;
1233 shadow_batch_obj = i915_gem_batch_pool_get(&engine->batch_pool,
1234 PAGE_ALIGN(batch_len));
1235 if (IS_ERR(shadow_batch_obj))
1236 return shadow_batch_obj;
1238 ret = intel_engine_cmd_parser(engine,
1247 ret = i915_gem_obj_ggtt_pin(shadow_batch_obj, 0, 0);
1251 i915_gem_object_unpin_pages(shadow_batch_obj);
1253 memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1255 vma = i915_gem_obj_to_ggtt(shadow_batch_obj);
1256 vma->exec_entry = shadow_exec_entry;
1257 vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1258 i915_gem_object_get(shadow_batch_obj);
1259 list_add_tail(&vma->exec_list, &eb->vmas);
1261 shadow_batch_obj->base.pending_read_domains = I915_GEM_DOMAIN_COMMAND;
1263 return shadow_batch_obj;
1266 i915_gem_object_unpin_pages(shadow_batch_obj);
1267 if (ret == -EACCES) /* unhandled chained batch */
1270 return ERR_PTR(ret);
1274 execbuf_submit(struct i915_execbuffer_params *params,
1275 struct drm_i915_gem_execbuffer2 *args,
1276 struct list_head *vmas)
1278 struct drm_i915_private *dev_priv = params->request->i915;
1279 u64 exec_start, exec_len;
1284 ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1288 ret = i915_switch_context(params->request);
1292 instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1293 instp_mask = I915_EXEC_CONSTANTS_MASK;
1294 switch (instp_mode) {
1295 case I915_EXEC_CONSTANTS_REL_GENERAL:
1296 case I915_EXEC_CONSTANTS_ABSOLUTE:
1297 case I915_EXEC_CONSTANTS_REL_SURFACE:
1298 if (instp_mode != 0 && params->engine->id != RCS) {
1299 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1303 if (instp_mode != dev_priv->relative_constants_mode) {
1304 if (INTEL_INFO(dev_priv)->gen < 4) {
1305 DRM_DEBUG("no rel constants on pre-gen4\n");
1309 if (INTEL_INFO(dev_priv)->gen > 5 &&
1310 instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1311 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1315 /* The HW changed the meaning on this bit on gen6 */
1316 if (INTEL_INFO(dev_priv)->gen >= 6)
1317 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1321 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1325 if (params->engine->id == RCS &&
1326 instp_mode != dev_priv->relative_constants_mode) {
1327 struct intel_ring *ring = params->request->ring;
1329 ret = intel_ring_begin(params->request, 4);
1333 intel_ring_emit(ring, MI_NOOP);
1334 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1335 intel_ring_emit_reg(ring, INSTPM);
1336 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1337 intel_ring_advance(ring);
1339 dev_priv->relative_constants_mode = instp_mode;
1342 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1343 ret = i915_reset_gen7_sol_offsets(params->request);
1348 exec_len = args->batch_len;
1349 exec_start = params->batch_obj_vm_offset +
1350 params->args_batch_start_offset;
1353 exec_len = params->batch_obj->base.size;
1355 ret = params->engine->emit_bb_start(params->request,
1356 exec_start, exec_len,
1357 params->dispatch_flags);
1361 trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1363 i915_gem_execbuffer_move_to_active(vmas, params->request);
1369 * Find one BSD ring to dispatch the corresponding BSD command.
1370 * The engine index is returned.
1373 gen8_dispatch_bsd_engine(struct drm_i915_private *dev_priv,
1374 struct drm_file *file)
1376 struct drm_i915_file_private *file_priv = file->driver_priv;
1378 /* Check whether the file_priv has already selected one ring. */
1379 if ((int)file_priv->bsd_engine < 0) {
1380 /* If not, use the ping-pong mechanism to select one. */
1381 mutex_lock(&dev_priv->drm.struct_mutex);
1382 file_priv->bsd_engine = dev_priv->mm.bsd_engine_dispatch_index;
1383 dev_priv->mm.bsd_engine_dispatch_index ^= 1;
1384 mutex_unlock(&dev_priv->drm.struct_mutex);
1387 return file_priv->bsd_engine;
1390 #define I915_USER_RINGS (4)
1392 static const enum intel_engine_id user_ring_map[I915_USER_RINGS + 1] = {
1393 [I915_EXEC_DEFAULT] = RCS,
1394 [I915_EXEC_RENDER] = RCS,
1395 [I915_EXEC_BLT] = BCS,
1396 [I915_EXEC_BSD] = VCS,
1397 [I915_EXEC_VEBOX] = VECS
1400 static struct intel_engine_cs *
1401 eb_select_engine(struct drm_i915_private *dev_priv,
1402 struct drm_file *file,
1403 struct drm_i915_gem_execbuffer2 *args)
1405 unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
1406 struct intel_engine_cs *engine;
1408 if (user_ring_id > I915_USER_RINGS) {
1409 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id);
1413 if ((user_ring_id != I915_EXEC_BSD) &&
1414 ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1415 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1416 "bsd dispatch flags: %d\n", (int)(args->flags));
1420 if (user_ring_id == I915_EXEC_BSD && HAS_BSD2(dev_priv)) {
1421 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
1423 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
1424 bsd_idx = gen8_dispatch_bsd_engine(dev_priv, file);
1425 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
1426 bsd_idx <= I915_EXEC_BSD_RING2) {
1427 bsd_idx >>= I915_EXEC_BSD_SHIFT;
1430 DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
1435 engine = &dev_priv->engine[_VCS(bsd_idx)];
1437 engine = &dev_priv->engine[user_ring_map[user_ring_id]];
1440 if (!intel_engine_initialized(engine)) {
1441 DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id);
1449 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1450 struct drm_file *file,
1451 struct drm_i915_gem_execbuffer2 *args,
1452 struct drm_i915_gem_exec_object2 *exec)
1454 struct drm_i915_private *dev_priv = to_i915(dev);
1455 struct i915_ggtt *ggtt = &dev_priv->ggtt;
1457 struct drm_i915_gem_object *batch_obj;
1458 struct drm_i915_gem_exec_object2 shadow_exec_entry;
1459 struct intel_engine_cs *engine;
1460 struct i915_gem_context *ctx;
1461 struct i915_address_space *vm;
1462 struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1463 struct i915_execbuffer_params *params = ¶ms_master;
1464 const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1469 if (!i915_gem_check_execbuffer(args))
1472 ret = validate_exec_list(dev, exec, args->buffer_count);
1477 if (args->flags & I915_EXEC_SECURE) {
1478 if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
1481 dispatch_flags |= I915_DISPATCH_SECURE;
1483 if (args->flags & I915_EXEC_IS_PINNED)
1484 dispatch_flags |= I915_DISPATCH_PINNED;
1486 engine = eb_select_engine(dev_priv, file, args);
1490 if (args->buffer_count < 1) {
1491 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1495 if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1496 if (!HAS_RESOURCE_STREAMER(dev)) {
1497 DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1500 if (engine->id != RCS) {
1501 DRM_DEBUG("RS is not available on %s\n",
1506 dispatch_flags |= I915_DISPATCH_RS;
1509 /* Take a local wakeref for preparing to dispatch the execbuf as
1510 * we expect to access the hardware fairly frequently in the
1511 * process. Upon first dispatch, we acquire another prolonged
1512 * wakeref that we hold until the GPU has been idle for at least
1515 intel_runtime_pm_get(dev_priv);
1517 ret = i915_mutex_lock_interruptible(dev);
1521 ctx = i915_gem_validate_context(dev, file, engine, ctx_id);
1523 mutex_unlock(&dev->struct_mutex);
1528 i915_gem_context_get(ctx);
1531 vm = &ctx->ppgtt->base;
1535 memset(¶ms_master, 0x00, sizeof(params_master));
1537 eb = eb_create(args);
1539 i915_gem_context_put(ctx);
1540 mutex_unlock(&dev->struct_mutex);
1545 /* Look up object handles */
1546 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1550 /* take note of the batch buffer before we might reorder the lists */
1551 batch_obj = eb_get_batch(eb);
1553 /* Move the objects en-masse into the GTT, evicting if necessary. */
1554 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1555 ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1560 /* The objects are in their final locations, apply the relocations. */
1562 ret = i915_gem_execbuffer_relocate(eb);
1564 if (ret == -EFAULT) {
1565 ret = i915_gem_execbuffer_relocate_slow(dev, args, file,
1568 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1574 /* Set the pending read domains for the batch buffer to COMMAND */
1575 if (batch_obj->base.pending_write_domain) {
1576 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1581 params->args_batch_start_offset = args->batch_start_offset;
1582 if (intel_engine_needs_cmd_parser(engine) && args->batch_len) {
1583 struct drm_i915_gem_object *parsed_batch_obj;
1585 parsed_batch_obj = i915_gem_execbuffer_parse(engine,
1589 args->batch_start_offset,
1591 drm_is_current_master(file));
1592 if (IS_ERR(parsed_batch_obj)) {
1593 ret = PTR_ERR(parsed_batch_obj);
1598 * parsed_batch_obj == batch_obj means batch not fully parsed:
1599 * Accept, but don't promote to secure.
1602 if (parsed_batch_obj != batch_obj) {
1604 * Batch parsed and accepted:
1606 * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1607 * bit from MI_BATCH_BUFFER_START commands issued in
1608 * the dispatch_execbuffer implementations. We
1609 * specifically don't want that set on batches the
1610 * command parser has accepted.
1612 dispatch_flags |= I915_DISPATCH_SECURE;
1613 params->args_batch_start_offset = 0;
1614 batch_obj = parsed_batch_obj;
1618 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1620 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1621 * batch" bit. Hence we need to pin secure batches into the global gtt.
1622 * hsw should have this fixed, but bdw mucks it up again. */
1623 if (dispatch_flags & I915_DISPATCH_SECURE) {
1625 * So on first glance it looks freaky that we pin the batch here
1626 * outside of the reservation loop. But:
1627 * - The batch is already pinned into the relevant ppgtt, so we
1628 * already have the backing storage fully allocated.
1629 * - No other BO uses the global gtt (well contexts, but meh),
1630 * so we don't really have issues with multiple objects not
1631 * fitting due to fragmentation.
1632 * So this is actually safe.
1634 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1638 params->batch_obj_vm_offset = i915_gem_obj_ggtt_offset(batch_obj);
1640 params->batch_obj_vm_offset = i915_gem_obj_offset(batch_obj, vm);
1642 /* Allocate a request for this batch buffer nice and early. */
1643 params->request = i915_gem_request_alloc(engine, ctx);
1644 if (IS_ERR(params->request)) {
1645 ret = PTR_ERR(params->request);
1646 goto err_batch_unpin;
1649 ret = i915_gem_request_add_to_client(params->request, file);
1654 * Save assorted stuff away to pass through to *_submission().
1655 * NB: This data should be 'persistent' and not local as it will
1656 * kept around beyond the duration of the IOCTL once the GPU
1657 * scheduler arrives.
1660 params->file = file;
1661 params->engine = engine;
1662 params->dispatch_flags = dispatch_flags;
1663 params->batch_obj = batch_obj;
1666 ret = execbuf_submit(params, args, &eb->vmas);
1668 i915_gem_execbuffer_retire_commands(params);
1672 * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1673 * batch vma for correctness. For less ugly and less fragility this
1674 * needs to be adjusted to also track the ggtt batch vma properly as
1677 if (dispatch_flags & I915_DISPATCH_SECURE)
1678 i915_gem_object_ggtt_unpin(batch_obj);
1681 /* the request owns the ref now */
1682 i915_gem_context_put(ctx);
1685 mutex_unlock(&dev->struct_mutex);
1688 /* intel_gpu_busy should also get a ref, so it will free when the device
1689 * is really idle. */
1690 intel_runtime_pm_put(dev_priv);
1695 * Legacy execbuffer just creates an exec2 list from the original exec object
1696 * list array and passes it to the real function.
1699 i915_gem_execbuffer(struct drm_device *dev, void *data,
1700 struct drm_file *file)
1702 struct drm_i915_gem_execbuffer *args = data;
1703 struct drm_i915_gem_execbuffer2 exec2;
1704 struct drm_i915_gem_exec_object *exec_list = NULL;
1705 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1708 if (args->buffer_count < 1) {
1709 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1713 /* Copy in the exec list from userland */
1714 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1715 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1716 if (exec_list == NULL || exec2_list == NULL) {
1717 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1718 args->buffer_count);
1719 drm_free_large(exec_list);
1720 drm_free_large(exec2_list);
1723 ret = copy_from_user(exec_list,
1724 u64_to_user_ptr(args->buffers_ptr),
1725 sizeof(*exec_list) * args->buffer_count);
1727 DRM_DEBUG("copy %d exec entries failed %d\n",
1728 args->buffer_count, ret);
1729 drm_free_large(exec_list);
1730 drm_free_large(exec2_list);
1734 for (i = 0; i < args->buffer_count; i++) {
1735 exec2_list[i].handle = exec_list[i].handle;
1736 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1737 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1738 exec2_list[i].alignment = exec_list[i].alignment;
1739 exec2_list[i].offset = exec_list[i].offset;
1740 if (INTEL_INFO(dev)->gen < 4)
1741 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1743 exec2_list[i].flags = 0;
1746 exec2.buffers_ptr = args->buffers_ptr;
1747 exec2.buffer_count = args->buffer_count;
1748 exec2.batch_start_offset = args->batch_start_offset;
1749 exec2.batch_len = args->batch_len;
1750 exec2.DR1 = args->DR1;
1751 exec2.DR4 = args->DR4;
1752 exec2.num_cliprects = args->num_cliprects;
1753 exec2.cliprects_ptr = args->cliprects_ptr;
1754 exec2.flags = I915_EXEC_RENDER;
1755 i915_execbuffer2_set_context_id(exec2, 0);
1757 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1759 struct drm_i915_gem_exec_object __user *user_exec_list =
1760 u64_to_user_ptr(args->buffers_ptr);
1762 /* Copy the new buffer offsets back to the user's exec list. */
1763 for (i = 0; i < args->buffer_count; i++) {
1764 exec2_list[i].offset =
1765 gen8_canonical_addr(exec2_list[i].offset);
1766 ret = __copy_to_user(&user_exec_list[i].offset,
1767 &exec2_list[i].offset,
1768 sizeof(user_exec_list[i].offset));
1771 DRM_DEBUG("failed to copy %d exec entries "
1772 "back to user (%d)\n",
1773 args->buffer_count, ret);
1779 drm_free_large(exec_list);
1780 drm_free_large(exec2_list);
1785 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1786 struct drm_file *file)
1788 struct drm_i915_gem_execbuffer2 *args = data;
1789 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1792 if (args->buffer_count < 1 ||
1793 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1794 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1798 if (args->rsvd2 != 0) {
1799 DRM_DEBUG("dirty rvsd2 field\n");
1803 exec2_list = drm_malloc_gfp(args->buffer_count,
1804 sizeof(*exec2_list),
1806 if (exec2_list == NULL) {
1807 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1808 args->buffer_count);
1811 ret = copy_from_user(exec2_list,
1812 u64_to_user_ptr(args->buffers_ptr),
1813 sizeof(*exec2_list) * args->buffer_count);
1815 DRM_DEBUG("copy %d exec entries failed %d\n",
1816 args->buffer_count, ret);
1817 drm_free_large(exec2_list);
1821 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1823 /* Copy the new buffer offsets back to the user's exec list. */
1824 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1825 u64_to_user_ptr(args->buffers_ptr);
1828 for (i = 0; i < args->buffer_count; i++) {
1829 exec2_list[i].offset =
1830 gen8_canonical_addr(exec2_list[i].offset);
1831 ret = __copy_to_user(&user_exec_list[i].offset,
1832 &exec2_list[i].offset,
1833 sizeof(user_exec_list[i].offset));
1836 DRM_DEBUG("failed to copy %d exec entries "
1838 args->buffer_count);
1844 drm_free_large(exec2_list);