2 * Copyright 2008 Jerome Glisse.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22 * DEALINGS IN THE SOFTWARE.
25 * Jerome Glisse <glisse@freedesktop.org>
27 #include <linux/list_sort.h>
29 #include <drm/amdgpu_drm.h>
31 #include "amdgpu_trace.h"
33 #define AMDGPU_CS_MAX_PRIORITY 32u
34 #define AMDGPU_CS_NUM_BUCKETS (AMDGPU_CS_MAX_PRIORITY + 1)
36 /* This is based on the bucket sort with O(n) time complexity.
37 * An item with priority "i" is added to bucket[i]. The lists are then
38 * concatenated in descending order.
40 struct amdgpu_cs_buckets {
41 struct list_head bucket[AMDGPU_CS_NUM_BUCKETS];
44 static void amdgpu_cs_buckets_init(struct amdgpu_cs_buckets *b)
48 for (i = 0; i < AMDGPU_CS_NUM_BUCKETS; i++)
49 INIT_LIST_HEAD(&b->bucket[i]);
52 static void amdgpu_cs_buckets_add(struct amdgpu_cs_buckets *b,
53 struct list_head *item, unsigned priority)
55 /* Since buffers which appear sooner in the relocation list are
56 * likely to be used more often than buffers which appear later
57 * in the list, the sort mustn't change the ordering of buffers
58 * with the same priority, i.e. it must be stable.
60 list_add_tail(item, &b->bucket[min(priority, AMDGPU_CS_MAX_PRIORITY)]);
63 static void amdgpu_cs_buckets_get_list(struct amdgpu_cs_buckets *b,
64 struct list_head *out_list)
68 /* Connect the sorted buckets in the output list. */
69 for (i = 0; i < AMDGPU_CS_NUM_BUCKETS; i++) {
70 list_splice(&b->bucket[i], out_list);
74 int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
75 u32 ip_instance, u32 ring,
76 struct amdgpu_ring **out_ring)
78 /* Right now all IPs have only one instance - multiple rings. */
79 if (ip_instance != 0) {
80 DRM_ERROR("invalid ip instance: %d\n", ip_instance);
86 DRM_ERROR("unknown ip type: %d\n", ip_type);
88 case AMDGPU_HW_IP_GFX:
89 if (ring < adev->gfx.num_gfx_rings) {
90 *out_ring = &adev->gfx.gfx_ring[ring];
92 DRM_ERROR("only %d gfx rings are supported now\n",
93 adev->gfx.num_gfx_rings);
97 case AMDGPU_HW_IP_COMPUTE:
98 if (ring < adev->gfx.num_compute_rings) {
99 *out_ring = &adev->gfx.compute_ring[ring];
101 DRM_ERROR("only %d compute rings are supported now\n",
102 adev->gfx.num_compute_rings);
106 case AMDGPU_HW_IP_DMA:
107 if (ring < adev->sdma.num_instances) {
108 *out_ring = &adev->sdma.instance[ring].ring;
110 DRM_ERROR("only %d SDMA rings are supported\n",
111 adev->sdma.num_instances);
115 case AMDGPU_HW_IP_UVD:
116 *out_ring = &adev->uvd.ring;
118 case AMDGPU_HW_IP_VCE:
120 *out_ring = &adev->vce.ring[ring];
122 DRM_ERROR("only two VCE rings are supported\n");
130 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
132 union drm_amdgpu_cs *cs = data;
133 uint64_t *chunk_array_user;
134 uint64_t *chunk_array;
135 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
140 if (cs->in.num_chunks == 0)
143 chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
147 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
153 p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
156 INIT_LIST_HEAD(&p->validated);
157 chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
158 if (copy_from_user(chunk_array, chunk_array_user,
159 sizeof(uint64_t)*cs->in.num_chunks)) {
164 p->nchunks = cs->in.num_chunks;
165 p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
172 for (i = 0; i < p->nchunks; i++) {
173 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
174 struct drm_amdgpu_cs_chunk user_chunk;
175 uint32_t __user *cdata;
177 chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
178 if (copy_from_user(&user_chunk, chunk_ptr,
179 sizeof(struct drm_amdgpu_cs_chunk))) {
182 goto free_partial_kdata;
184 p->chunks[i].chunk_id = user_chunk.chunk_id;
185 p->chunks[i].length_dw = user_chunk.length_dw;
187 size = p->chunks[i].length_dw;
188 cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
189 p->chunks[i].user_ptr = cdata;
191 p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
192 if (p->chunks[i].kdata == NULL) {
195 goto free_partial_kdata;
197 size *= sizeof(uint32_t);
198 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
200 goto free_partial_kdata;
203 switch (p->chunks[i].chunk_id) {
204 case AMDGPU_CHUNK_ID_IB:
208 case AMDGPU_CHUNK_ID_FENCE:
209 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
210 if (p->chunks[i].length_dw * sizeof(uint32_t) >= size) {
212 struct drm_gem_object *gobj;
213 struct drm_amdgpu_cs_chunk_fence *fence_data;
215 fence_data = (void *)p->chunks[i].kdata;
216 handle = fence_data->handle;
217 gobj = drm_gem_object_lookup(p->adev->ddev,
221 goto free_partial_kdata;
224 p->uf.bo = gem_to_amdgpu_bo(gobj);
225 amdgpu_bo_ref(p->uf.bo);
226 drm_gem_object_unreference_unlocked(gobj);
227 p->uf.offset = fence_data->offset;
230 goto free_partial_kdata;
234 case AMDGPU_CHUNK_ID_DEPENDENCIES:
239 goto free_partial_kdata;
244 p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
257 drm_free_large(p->chunks[i].kdata);
261 amdgpu_bo_list_put(p->bo_list);
262 amdgpu_ctx_put(p->ctx);
269 /* Returns how many bytes TTM can move per IB.
271 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
273 u64 real_vram_size = adev->mc.real_vram_size;
274 u64 vram_usage = atomic64_read(&adev->vram_usage);
276 /* This function is based on the current VRAM usage.
278 * - If all of VRAM is free, allow relocating the number of bytes that
279 * is equal to 1/4 of the size of VRAM for this IB.
281 * - If more than one half of VRAM is occupied, only allow relocating
282 * 1 MB of data for this IB.
284 * - From 0 to one half of used VRAM, the threshold decreases
299 * Note: It's a threshold, not a limit. The threshold must be crossed
300 * for buffer relocations to stop, so any buffer of an arbitrary size
301 * can be moved as long as the threshold isn't crossed before
302 * the relocation takes place. We don't want to disable buffer
303 * relocations completely.
305 * The idea is that buffers should be placed in VRAM at creation time
306 * and TTM should only do a minimum number of relocations during
307 * command submission. In practice, you need to submit at least
308 * a dozen IBs to move all buffers to VRAM if they are in GTT.
310 * Also, things can get pretty crazy under memory pressure and actual
311 * VRAM usage can change a lot, so playing safe even at 50% does
312 * consistently increase performance.
315 u64 half_vram = real_vram_size >> 1;
316 u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
317 u64 bytes_moved_threshold = half_free_vram >> 1;
318 return max(bytes_moved_threshold, 1024*1024ull);
321 int amdgpu_cs_list_validate(struct amdgpu_device *adev,
322 struct amdgpu_vm *vm,
323 struct list_head *validated)
325 struct amdgpu_bo_list_entry *lobj;
326 struct amdgpu_bo *bo;
327 u64 bytes_moved = 0, initial_bytes_moved;
328 u64 bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(adev);
331 list_for_each_entry(lobj, validated, tv.head) {
333 if (!bo->pin_count) {
334 u32 domain = lobj->prefered_domains;
336 amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
338 /* Check if this buffer will be moved and don't move it
339 * if we have moved too many buffers for this IB already.
341 * Note that this allows moving at least one buffer of
342 * any size, because it doesn't take the current "bo"
343 * into account. We don't want to disallow buffer moves
346 if ((lobj->allowed_domains & current_domain) != 0 &&
347 (domain & current_domain) == 0 && /* will be moved */
348 bytes_moved > bytes_moved_threshold) {
350 domain = current_domain;
354 amdgpu_ttm_placement_from_domain(bo, domain);
355 initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
356 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
357 bytes_moved += atomic64_read(&adev->num_bytes_moved) -
361 if (r != -ERESTARTSYS && domain != lobj->allowed_domains) {
362 domain = lobj->allowed_domains;
368 lobj->bo_va = amdgpu_vm_bo_find(vm, bo);
373 static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p)
375 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
376 struct amdgpu_cs_buckets buckets;
377 struct list_head duplicates;
378 bool need_mmap_lock = false;
382 need_mmap_lock = p->bo_list->has_userptr;
383 amdgpu_cs_buckets_init(&buckets);
384 for (i = 0; i < p->bo_list->num_entries; i++)
385 amdgpu_cs_buckets_add(&buckets, &p->bo_list->array[i].tv.head,
386 p->bo_list->array[i].priority);
388 amdgpu_cs_buckets_get_list(&buckets, &p->validated);
391 p->vm_bos = amdgpu_vm_get_bos(p->adev, &fpriv->vm,
395 down_read(¤t->mm->mmap_sem);
397 INIT_LIST_HEAD(&duplicates);
398 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates);
399 if (unlikely(r != 0))
402 r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &p->validated);
406 r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &duplicates);
410 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
414 up_read(¤t->mm->mmap_sem);
419 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
421 struct amdgpu_bo_list_entry *e;
424 list_for_each_entry(e, &p->validated, tv.head) {
425 struct reservation_object *resv = e->robj->tbo.resv;
426 r = amdgpu_sync_resv(p->adev, &p->ibs[0].sync, resv, p->filp);
434 static int cmp_size_smaller_first(void *priv, struct list_head *a,
437 struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head);
438 struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head);
440 /* Sort A before B if A is smaller. */
441 return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
445 * cs_parser_fini() - clean parser states
446 * @parser: parser structure holding parsing context.
447 * @error: error number
449 * If error is set than unvalidate buffer, otherwise just free memory
450 * used by parsing context.
452 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
457 /* Sort the buffer list from the smallest to largest buffer,
458 * which affects the order of buffers in the LRU list.
459 * This assures that the smallest buffers are added first
460 * to the LRU list, so they are likely to be later evicted
461 * first, instead of large buffers whose eviction is more
464 * This slightly lowers the number of bytes moved by TTM
465 * per frame under memory pressure.
467 list_sort(NULL, &parser->validated, cmp_size_smaller_first);
469 ttm_eu_fence_buffer_objects(&parser->ticket,
472 } else if (backoff) {
473 ttm_eu_backoff_reservation(&parser->ticket,
476 fence_put(parser->fence);
479 amdgpu_ctx_put(parser->ctx);
481 amdgpu_bo_list_put(parser->bo_list);
483 drm_free_large(parser->vm_bos);
484 for (i = 0; i < parser->nchunks; i++)
485 drm_free_large(parser->chunks[i].kdata);
486 kfree(parser->chunks);
488 for (i = 0; i < parser->num_ibs; i++)
489 amdgpu_ib_free(parser->adev, &parser->ibs[i]);
492 amdgpu_bo_unref(&parser->uf.bo);
495 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
496 struct amdgpu_vm *vm)
498 struct amdgpu_device *adev = p->adev;
499 struct amdgpu_bo_va *bo_va;
500 struct amdgpu_bo *bo;
503 r = amdgpu_vm_update_page_directory(adev, vm);
507 r = amdgpu_sync_fence(adev, &p->ibs[0].sync, vm->page_directory_fence);
511 r = amdgpu_vm_clear_freed(adev, vm);
516 for (i = 0; i < p->bo_list->num_entries; i++) {
519 /* ignore duplicates */
520 bo = p->bo_list->array[i].robj;
524 bo_va = p->bo_list->array[i].bo_va;
528 r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
532 f = bo_va->last_pt_update;
533 r = amdgpu_sync_fence(adev, &p->ibs[0].sync, f);
540 r = amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
542 if (amdgpu_vm_debug && p->bo_list) {
543 /* Invalidate all BOs to test for userspace bugs */
544 for (i = 0; i < p->bo_list->num_entries; i++) {
545 /* ignore duplicates */
546 bo = p->bo_list->array[i].robj;
550 amdgpu_vm_bo_invalidate(adev, bo);
557 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
558 struct amdgpu_cs_parser *parser)
560 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
561 struct amdgpu_vm *vm = &fpriv->vm;
562 struct amdgpu_ring *ring;
565 if (parser->num_ibs == 0)
568 /* Only for UVD/VCE VM emulation */
569 for (i = 0; i < parser->num_ibs; i++) {
570 ring = parser->ibs[i].ring;
571 if (ring->funcs->parse_cs) {
572 r = amdgpu_ring_parse_cs(ring, parser, i);
578 r = amdgpu_bo_vm_update_pte(parser, vm);
580 amdgpu_cs_sync_rings(parser);
585 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
588 r = amdgpu_gpu_reset(adev);
595 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
596 struct amdgpu_cs_parser *parser)
598 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
599 struct amdgpu_vm *vm = &fpriv->vm;
603 for (i = 0, j = 0; i < parser->nchunks && j < parser->num_ibs; i++) {
604 struct amdgpu_cs_chunk *chunk;
605 struct amdgpu_ib *ib;
606 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
607 struct amdgpu_ring *ring;
609 chunk = &parser->chunks[i];
610 ib = &parser->ibs[j];
611 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
613 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
616 r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
617 chunk_ib->ip_instance, chunk_ib->ring,
622 if (ring->funcs->parse_cs) {
623 struct amdgpu_bo_va_mapping *m;
624 struct amdgpu_bo *aobj = NULL;
628 m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
631 DRM_ERROR("IB va_start is invalid\n");
635 if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
636 (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
637 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
641 /* the IB should be reserved at this point */
642 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
647 offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
648 kptr += chunk_ib->va_start - offset;
650 r = amdgpu_ib_get(ring, NULL, chunk_ib->ib_bytes, ib);
652 DRM_ERROR("Failed to get ib !\n");
656 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
657 amdgpu_bo_kunmap(aobj);
659 r = amdgpu_ib_get(ring, vm, 0, ib);
661 DRM_ERROR("Failed to get ib !\n");
665 ib->gpu_addr = chunk_ib->va_start;
668 ib->length_dw = chunk_ib->ib_bytes / 4;
669 ib->flags = chunk_ib->flags;
670 ib->ctx = parser->ctx;
674 if (!parser->num_ibs)
677 /* add GDS resources to first IB */
678 if (parser->bo_list) {
679 struct amdgpu_bo *gds = parser->bo_list->gds_obj;
680 struct amdgpu_bo *gws = parser->bo_list->gws_obj;
681 struct amdgpu_bo *oa = parser->bo_list->oa_obj;
682 struct amdgpu_ib *ib = &parser->ibs[0];
685 ib->gds_base = amdgpu_bo_gpu_offset(gds);
686 ib->gds_size = amdgpu_bo_size(gds);
689 ib->gws_base = amdgpu_bo_gpu_offset(gws);
690 ib->gws_size = amdgpu_bo_size(gws);
693 ib->oa_base = amdgpu_bo_gpu_offset(oa);
694 ib->oa_size = amdgpu_bo_size(oa);
697 /* wrap the last IB with user fence */
699 struct amdgpu_ib *ib = &parser->ibs[parser->num_ibs - 1];
701 /* UVD & VCE fw doesn't support user fences */
702 if (ib->ring->type == AMDGPU_RING_TYPE_UVD ||
703 ib->ring->type == AMDGPU_RING_TYPE_VCE)
706 ib->user = &parser->uf;
712 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
713 struct amdgpu_cs_parser *p)
715 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
716 struct amdgpu_ib *ib;
722 /* Add dependencies to first IB */
724 for (i = 0; i < p->nchunks; ++i) {
725 struct drm_amdgpu_cs_chunk_dep *deps;
726 struct amdgpu_cs_chunk *chunk;
729 chunk = &p->chunks[i];
731 if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
734 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
735 num_deps = chunk->length_dw * 4 /
736 sizeof(struct drm_amdgpu_cs_chunk_dep);
738 for (j = 0; j < num_deps; ++j) {
739 struct amdgpu_ring *ring;
740 struct amdgpu_ctx *ctx;
743 r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
745 deps[j].ring, &ring);
749 ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
753 fence = amdgpu_ctx_get_fence(ctx, ring,
761 r = amdgpu_sync_fence(adev, &ib->sync, fence);
773 static int amdgpu_cs_free_job(struct amdgpu_job *job)
777 for (i = 0; i < job->num_ibs; i++)
778 amdgpu_ib_free(job->adev, &job->ibs[i]);
781 amdgpu_bo_unref(&job->uf.bo);
785 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
787 struct amdgpu_device *adev = dev->dev_private;
788 union drm_amdgpu_cs *cs = data;
789 struct amdgpu_cs_parser parser = {};
790 bool reserved_buffers = false;
793 if (!adev->accel_working)
799 r = amdgpu_cs_parser_init(&parser, data);
801 DRM_ERROR("Failed to initialize parser !\n");
802 amdgpu_cs_parser_fini(&parser, r, false);
803 r = amdgpu_cs_handle_lockup(adev, r);
806 r = amdgpu_cs_parser_relocs(&parser);
808 DRM_ERROR("Not enough memory for command submission!\n");
809 else if (r && r != -ERESTARTSYS)
810 DRM_ERROR("Failed to process the buffer list %d!\n", r);
812 reserved_buffers = true;
813 r = amdgpu_cs_ib_fill(adev, &parser);
817 r = amdgpu_cs_dependencies(adev, &parser);
819 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
825 for (i = 0; i < parser.num_ibs; i++)
826 trace_amdgpu_cs(&parser, i);
828 r = amdgpu_cs_ib_vm_chunk(adev, &parser);
832 if (amdgpu_enable_scheduler && parser.num_ibs) {
833 struct amdgpu_ring * ring = parser.ibs->ring;
834 struct amd_sched_fence *fence;
835 struct amdgpu_job *job;
837 job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL);
843 job->base.sched = &ring->sched;
844 job->base.s_entity = &parser.ctx->rings[ring->idx].entity;
845 job->adev = parser.adev;
846 job->owner = parser.filp;
847 job->free_job = amdgpu_cs_free_job;
849 job->ibs = parser.ibs;
850 job->num_ibs = parser.num_ibs;
854 if (job->ibs[job->num_ibs - 1].user) {
856 job->ibs[job->num_ibs - 1].user = &job->uf;
860 fence = amd_sched_fence_create(job->base.s_entity,
864 amdgpu_cs_free_job(job);
868 job->base.s_fence = fence;
869 parser.fence = fence_get(&fence->base);
871 cs->out.handle = amdgpu_ctx_add_fence(parser.ctx, ring,
873 job->ibs[job->num_ibs - 1].sequence = cs->out.handle;
875 trace_amdgpu_cs_ioctl(job);
876 amd_sched_entity_push_job(&job->base);
879 struct amdgpu_fence *fence;
881 r = amdgpu_ib_schedule(adev, parser.num_ibs, parser.ibs,
883 fence = parser.ibs[parser.num_ibs - 1].fence;
884 parser.fence = fence_get(&fence->base);
885 cs->out.handle = parser.ibs[parser.num_ibs - 1].sequence;
889 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
890 r = amdgpu_cs_handle_lockup(adev, r);
895 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
898 * @data: data from userspace
899 * @filp: file private
901 * Wait for the command submission identified by handle to finish.
903 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
904 struct drm_file *filp)
906 union drm_amdgpu_wait_cs *wait = data;
907 struct amdgpu_device *adev = dev->dev_private;
908 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
909 struct amdgpu_ring *ring = NULL;
910 struct amdgpu_ctx *ctx;
914 r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
915 wait->in.ring, &ring);
919 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
923 fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
927 r = fence_wait_timeout(fence, true, timeout);
936 memset(wait, 0, sizeof(*wait));
937 wait->out.status = (r == 0);
943 * amdgpu_cs_find_bo_va - find bo_va for VM address
945 * @parser: command submission parser context
947 * @bo: resulting BO of the mapping found
949 * Search the buffer objects in the command submission context for a certain
950 * virtual memory address. Returns allocation structure when found, NULL
953 struct amdgpu_bo_va_mapping *
954 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
955 uint64_t addr, struct amdgpu_bo **bo)
957 struct amdgpu_bo_list_entry *reloc;
958 struct amdgpu_bo_va_mapping *mapping;
960 addr /= AMDGPU_GPU_PAGE_SIZE;
962 list_for_each_entry(reloc, &parser->validated, tv.head) {
966 list_for_each_entry(mapping, &reloc->bo_va->valids, list) {
967 if (mapping->it.start > addr ||
968 addr > mapping->it.last)
971 *bo = reloc->bo_va->bo;
975 list_for_each_entry(mapping, &reloc->bo_va->invalids, list) {
976 if (mapping->it.start > addr ||
977 addr > mapping->it.last)
980 *bo = reloc->bo_va->bo;