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/pagemap.h>
29 #include <drm/amdgpu_drm.h>
31 #include "amdgpu_trace.h"
33 int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
34 u32 ip_instance, u32 ring,
35 struct amdgpu_ring **out_ring)
37 /* Right now all IPs have only one instance - multiple rings. */
38 if (ip_instance != 0) {
39 DRM_ERROR("invalid ip instance: %d\n", ip_instance);
45 DRM_ERROR("unknown ip type: %d\n", ip_type);
47 case AMDGPU_HW_IP_GFX:
48 if (ring < adev->gfx.num_gfx_rings) {
49 *out_ring = &adev->gfx.gfx_ring[ring];
51 DRM_ERROR("only %d gfx rings are supported now\n",
52 adev->gfx.num_gfx_rings);
56 case AMDGPU_HW_IP_COMPUTE:
57 if (ring < adev->gfx.num_compute_rings) {
58 *out_ring = &adev->gfx.compute_ring[ring];
60 DRM_ERROR("only %d compute rings are supported now\n",
61 adev->gfx.num_compute_rings);
65 case AMDGPU_HW_IP_DMA:
66 if (ring < adev->sdma.num_instances) {
67 *out_ring = &adev->sdma.instance[ring].ring;
69 DRM_ERROR("only %d SDMA rings are supported\n",
70 adev->sdma.num_instances);
74 case AMDGPU_HW_IP_UVD:
75 *out_ring = &adev->uvd.ring;
77 case AMDGPU_HW_IP_VCE:
79 *out_ring = &adev->vce.ring[ring];
81 DRM_ERROR("only two VCE rings are supported\n");
89 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
90 struct drm_amdgpu_cs_chunk_fence *data,
93 struct drm_gem_object *gobj;
95 gobj = drm_gem_object_lookup(p->adev->ddev, p->filp,
100 p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
101 p->uf_entry.priority = 0;
102 p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
103 p->uf_entry.tv.shared = true;
104 p->uf_entry.user_pages = NULL;
105 *offset = data->offset;
107 drm_gem_object_unreference_unlocked(gobj);
109 if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
110 amdgpu_bo_unref(&p->uf_entry.robj);
117 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
119 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
120 struct amdgpu_vm *vm = &fpriv->vm;
121 union drm_amdgpu_cs *cs = data;
122 uint64_t *chunk_array_user;
123 uint64_t *chunk_array;
124 unsigned size, num_ibs = 0;
125 uint32_t uf_offset = 0;
129 if (cs->in.num_chunks == 0)
132 chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
136 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
143 chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
144 if (copy_from_user(chunk_array, chunk_array_user,
145 sizeof(uint64_t)*cs->in.num_chunks)) {
150 p->nchunks = cs->in.num_chunks;
151 p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
158 for (i = 0; i < p->nchunks; i++) {
159 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
160 struct drm_amdgpu_cs_chunk user_chunk;
161 uint32_t __user *cdata;
163 chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
164 if (copy_from_user(&user_chunk, chunk_ptr,
165 sizeof(struct drm_amdgpu_cs_chunk))) {
168 goto free_partial_kdata;
170 p->chunks[i].chunk_id = user_chunk.chunk_id;
171 p->chunks[i].length_dw = user_chunk.length_dw;
173 size = p->chunks[i].length_dw;
174 cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
176 p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
177 if (p->chunks[i].kdata == NULL) {
180 goto free_partial_kdata;
182 size *= sizeof(uint32_t);
183 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
185 goto free_partial_kdata;
188 switch (p->chunks[i].chunk_id) {
189 case AMDGPU_CHUNK_ID_IB:
193 case AMDGPU_CHUNK_ID_FENCE:
194 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
195 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
197 goto free_partial_kdata;
200 ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
203 goto free_partial_kdata;
207 case AMDGPU_CHUNK_ID_DEPENDENCIES:
212 goto free_partial_kdata;
216 ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
220 if (p->uf_entry.robj) {
221 p->job->uf_bo = amdgpu_bo_ref(p->uf_entry.robj);
222 p->job->uf_offset = uf_offset;
232 drm_free_large(p->chunks[i].kdata);
235 amdgpu_ctx_put(p->ctx);
242 /* Returns how many bytes TTM can move per IB.
244 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
246 u64 real_vram_size = adev->mc.real_vram_size;
247 u64 vram_usage = atomic64_read(&adev->vram_usage);
249 /* This function is based on the current VRAM usage.
251 * - If all of VRAM is free, allow relocating the number of bytes that
252 * is equal to 1/4 of the size of VRAM for this IB.
254 * - If more than one half of VRAM is occupied, only allow relocating
255 * 1 MB of data for this IB.
257 * - From 0 to one half of used VRAM, the threshold decreases
272 * Note: It's a threshold, not a limit. The threshold must be crossed
273 * for buffer relocations to stop, so any buffer of an arbitrary size
274 * can be moved as long as the threshold isn't crossed before
275 * the relocation takes place. We don't want to disable buffer
276 * relocations completely.
278 * The idea is that buffers should be placed in VRAM at creation time
279 * and TTM should only do a minimum number of relocations during
280 * command submission. In practice, you need to submit at least
281 * a dozen IBs to move all buffers to VRAM if they are in GTT.
283 * Also, things can get pretty crazy under memory pressure and actual
284 * VRAM usage can change a lot, so playing safe even at 50% does
285 * consistently increase performance.
288 u64 half_vram = real_vram_size >> 1;
289 u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
290 u64 bytes_moved_threshold = half_free_vram >> 1;
291 return max(bytes_moved_threshold, 1024*1024ull);
294 int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
295 struct list_head *validated)
297 struct amdgpu_bo_list_entry *lobj;
298 u64 initial_bytes_moved;
301 list_for_each_entry(lobj, validated, tv.head) {
302 struct amdgpu_bo *bo = lobj->robj;
303 bool binding_userptr = false;
304 struct mm_struct *usermm;
307 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
308 if (usermm && usermm != current->mm)
311 /* Check if we have user pages and nobody bound the BO already */
312 if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
313 size_t size = sizeof(struct page *);
315 size *= bo->tbo.ttm->num_pages;
316 memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
317 binding_userptr = true;
323 /* Avoid moving this one if we have moved too many buffers
324 * for this IB already.
326 * Note that this allows moving at least one buffer of
327 * any size, because it doesn't take the current "bo"
328 * into account. We don't want to disallow buffer moves
331 if (p->bytes_moved <= p->bytes_moved_threshold)
332 domain = bo->prefered_domains;
334 domain = bo->allowed_domains;
337 amdgpu_ttm_placement_from_domain(bo, domain);
338 initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
339 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
340 p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
344 if (r != -ERESTARTSYS && domain != bo->allowed_domains) {
345 domain = bo->allowed_domains;
351 if (binding_userptr) {
352 drm_free_large(lobj->user_pages);
353 lobj->user_pages = NULL;
359 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
360 union drm_amdgpu_cs *cs)
362 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
363 struct amdgpu_bo_list_entry *e;
364 struct list_head duplicates;
365 bool need_mmap_lock = false;
366 unsigned i, tries = 10;
369 INIT_LIST_HEAD(&p->validated);
371 p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
373 need_mmap_lock = p->bo_list->first_userptr !=
374 p->bo_list->num_entries;
375 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
378 INIT_LIST_HEAD(&duplicates);
379 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
381 if (p->uf_entry.robj)
382 list_add(&p->uf_entry.tv.head, &p->validated);
385 down_read(¤t->mm->mmap_sem);
388 struct list_head need_pages;
391 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
393 if (unlikely(r != 0))
394 goto error_free_pages;
396 /* Without a BO list we don't have userptr BOs */
400 INIT_LIST_HEAD(&need_pages);
401 for (i = p->bo_list->first_userptr;
402 i < p->bo_list->num_entries; ++i) {
404 e = &p->bo_list->array[i];
406 if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
407 &e->user_invalidated) && e->user_pages) {
409 /* We acquired a page array, but somebody
410 * invalidated it. Free it an try again
412 release_pages(e->user_pages,
413 e->robj->tbo.ttm->num_pages,
415 drm_free_large(e->user_pages);
416 e->user_pages = NULL;
419 if (e->robj->tbo.ttm->state != tt_bound &&
421 list_del(&e->tv.head);
422 list_add(&e->tv.head, &need_pages);
424 amdgpu_bo_unreserve(e->robj);
428 if (list_empty(&need_pages))
431 /* Unreserve everything again. */
432 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
434 /* We tried to often, just abort */
437 goto error_free_pages;
440 /* Fill the page arrays for all useptrs. */
441 list_for_each_entry(e, &need_pages, tv.head) {
442 struct ttm_tt *ttm = e->robj->tbo.ttm;
444 e->user_pages = drm_calloc_large(ttm->num_pages,
445 sizeof(struct page*));
446 if (!e->user_pages) {
448 goto error_free_pages;
451 r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
453 drm_free_large(e->user_pages);
454 e->user_pages = NULL;
455 goto error_free_pages;
460 list_splice(&need_pages, &p->validated);
463 amdgpu_vm_get_pt_bos(&fpriv->vm, &duplicates);
465 p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
468 r = amdgpu_cs_list_validate(p, &duplicates);
472 r = amdgpu_cs_list_validate(p, &p->validated);
477 struct amdgpu_bo *gds = p->bo_list->gds_obj;
478 struct amdgpu_bo *gws = p->bo_list->gws_obj;
479 struct amdgpu_bo *oa = p->bo_list->oa_obj;
480 struct amdgpu_vm *vm = &fpriv->vm;
483 for (i = 0; i < p->bo_list->num_entries; i++) {
484 struct amdgpu_bo *bo = p->bo_list->array[i].robj;
486 p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
490 p->job->gds_base = amdgpu_bo_gpu_offset(gds);
491 p->job->gds_size = amdgpu_bo_size(gds);
494 p->job->gws_base = amdgpu_bo_gpu_offset(gws);
495 p->job->gws_size = amdgpu_bo_size(gws);
498 p->job->oa_base = amdgpu_bo_gpu_offset(oa);
499 p->job->oa_size = amdgpu_bo_size(oa);
505 amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
506 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
512 up_read(¤t->mm->mmap_sem);
515 for (i = p->bo_list->first_userptr;
516 i < p->bo_list->num_entries; ++i) {
517 e = &p->bo_list->array[i];
522 release_pages(e->user_pages,
523 e->robj->tbo.ttm->num_pages,
525 drm_free_large(e->user_pages);
532 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
534 struct amdgpu_bo_list_entry *e;
537 list_for_each_entry(e, &p->validated, tv.head) {
538 struct reservation_object *resv = e->robj->tbo.resv;
539 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
548 * cs_parser_fini() - clean parser states
549 * @parser: parser structure holding parsing context.
550 * @error: error number
552 * If error is set than unvalidate buffer, otherwise just free memory
553 * used by parsing context.
555 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
557 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
561 amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
563 ttm_eu_fence_buffer_objects(&parser->ticket,
566 } else if (backoff) {
567 ttm_eu_backoff_reservation(&parser->ticket,
570 fence_put(parser->fence);
573 amdgpu_ctx_put(parser->ctx);
575 amdgpu_bo_list_put(parser->bo_list);
577 for (i = 0; i < parser->nchunks; i++)
578 drm_free_large(parser->chunks[i].kdata);
579 kfree(parser->chunks);
581 amdgpu_job_free(parser->job);
582 amdgpu_bo_unref(&parser->uf_entry.robj);
585 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
586 struct amdgpu_vm *vm)
588 struct amdgpu_device *adev = p->adev;
589 struct amdgpu_bo_va *bo_va;
590 struct amdgpu_bo *bo;
593 r = amdgpu_vm_update_page_directory(adev, vm);
597 r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
601 r = amdgpu_vm_clear_freed(adev, vm);
606 for (i = 0; i < p->bo_list->num_entries; i++) {
609 /* ignore duplicates */
610 bo = p->bo_list->array[i].robj;
614 bo_va = p->bo_list->array[i].bo_va;
618 r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
622 f = bo_va->last_pt_update;
623 r = amdgpu_sync_fence(adev, &p->job->sync, f);
630 r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
632 if (amdgpu_vm_debug && p->bo_list) {
633 /* Invalidate all BOs to test for userspace bugs */
634 for (i = 0; i < p->bo_list->num_entries; i++) {
635 /* ignore duplicates */
636 bo = p->bo_list->array[i].robj;
640 amdgpu_vm_bo_invalidate(adev, bo);
647 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
648 struct amdgpu_cs_parser *p)
650 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
651 struct amdgpu_vm *vm = &fpriv->vm;
652 struct amdgpu_ring *ring = p->job->ring;
655 /* Only for UVD/VCE VM emulation */
656 if (ring->funcs->parse_cs) {
657 for (i = 0; i < p->job->num_ibs; i++) {
658 r = amdgpu_ring_parse_cs(ring, p, i);
664 r = amdgpu_bo_vm_update_pte(p, vm);
666 amdgpu_cs_sync_rings(p);
671 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
674 r = amdgpu_gpu_reset(adev);
681 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
682 struct amdgpu_cs_parser *parser)
684 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
685 struct amdgpu_vm *vm = &fpriv->vm;
689 for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
690 struct amdgpu_cs_chunk *chunk;
691 struct amdgpu_ib *ib;
692 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
693 struct amdgpu_ring *ring;
695 chunk = &parser->chunks[i];
696 ib = &parser->job->ibs[j];
697 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
699 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
702 r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
703 chunk_ib->ip_instance, chunk_ib->ring,
708 if (parser->job->ring && parser->job->ring != ring)
711 parser->job->ring = ring;
713 if (ring->funcs->parse_cs) {
714 struct amdgpu_bo_va_mapping *m;
715 struct amdgpu_bo *aobj = NULL;
719 m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
722 DRM_ERROR("IB va_start is invalid\n");
726 if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
727 (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
728 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
732 /* the IB should be reserved at this point */
733 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
738 offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
739 kptr += chunk_ib->va_start - offset;
741 r = amdgpu_ib_get(adev, NULL, chunk_ib->ib_bytes, ib);
743 DRM_ERROR("Failed to get ib !\n");
747 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
748 amdgpu_bo_kunmap(aobj);
750 r = amdgpu_ib_get(adev, vm, 0, ib);
752 DRM_ERROR("Failed to get ib !\n");
756 ib->gpu_addr = chunk_ib->va_start;
759 ib->length_dw = chunk_ib->ib_bytes / 4;
760 ib->flags = chunk_ib->flags;
764 /* UVD & VCE fw doesn't support user fences */
765 if (parser->job->uf_bo && (
766 parser->job->ring->type == AMDGPU_RING_TYPE_UVD ||
767 parser->job->ring->type == AMDGPU_RING_TYPE_VCE))
773 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
774 struct amdgpu_cs_parser *p)
776 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
779 for (i = 0; i < p->nchunks; ++i) {
780 struct drm_amdgpu_cs_chunk_dep *deps;
781 struct amdgpu_cs_chunk *chunk;
784 chunk = &p->chunks[i];
786 if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
789 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
790 num_deps = chunk->length_dw * 4 /
791 sizeof(struct drm_amdgpu_cs_chunk_dep);
793 for (j = 0; j < num_deps; ++j) {
794 struct amdgpu_ring *ring;
795 struct amdgpu_ctx *ctx;
798 r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
800 deps[j].ring, &ring);
804 ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
808 fence = amdgpu_ctx_get_fence(ctx, ring,
816 r = amdgpu_sync_fence(adev, &p->job->sync,
829 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
830 union drm_amdgpu_cs *cs)
832 struct amdgpu_ring *ring = p->job->ring;
833 struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
835 struct amdgpu_job *job;
841 r = amd_sched_job_init(&job->base, &ring->sched,
842 entity, amdgpu_job_timeout_func,
843 amdgpu_job_free_func,
846 amdgpu_job_free(job);
850 job->owner = p->filp;
851 job->ctx = entity->fence_context;
852 p->fence = fence_get(fence);
853 cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, fence);
854 job->uf_sequence = cs->out.handle;
856 trace_amdgpu_cs_ioctl(job);
857 amd_sched_entity_push_job(&job->base);
862 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
864 struct amdgpu_device *adev = dev->dev_private;
865 union drm_amdgpu_cs *cs = data;
866 struct amdgpu_cs_parser parser = {};
867 bool reserved_buffers = false;
870 if (!adev->accel_working)
876 r = amdgpu_cs_parser_init(&parser, data);
878 DRM_ERROR("Failed to initialize parser !\n");
879 amdgpu_cs_parser_fini(&parser, r, false);
880 r = amdgpu_cs_handle_lockup(adev, r);
883 r = amdgpu_cs_parser_bos(&parser, data);
885 DRM_ERROR("Not enough memory for command submission!\n");
886 else if (r && r != -ERESTARTSYS)
887 DRM_ERROR("Failed to process the buffer list %d!\n", r);
889 reserved_buffers = true;
890 r = amdgpu_cs_ib_fill(adev, &parser);
894 r = amdgpu_cs_dependencies(adev, &parser);
896 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
902 for (i = 0; i < parser.job->num_ibs; i++)
903 trace_amdgpu_cs(&parser, i);
905 r = amdgpu_cs_ib_vm_chunk(adev, &parser);
909 r = amdgpu_cs_submit(&parser, cs);
912 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
913 r = amdgpu_cs_handle_lockup(adev, r);
918 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
921 * @data: data from userspace
922 * @filp: file private
924 * Wait for the command submission identified by handle to finish.
926 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
927 struct drm_file *filp)
929 union drm_amdgpu_wait_cs *wait = data;
930 struct amdgpu_device *adev = dev->dev_private;
931 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
932 struct amdgpu_ring *ring = NULL;
933 struct amdgpu_ctx *ctx;
937 r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
938 wait->in.ring, &ring);
942 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
946 fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
950 r = fence_wait_timeout(fence, true, timeout);
959 memset(wait, 0, sizeof(*wait));
960 wait->out.status = (r == 0);
966 * amdgpu_cs_find_bo_va - find bo_va for VM address
968 * @parser: command submission parser context
970 * @bo: resulting BO of the mapping found
972 * Search the buffer objects in the command submission context for a certain
973 * virtual memory address. Returns allocation structure when found, NULL
976 struct amdgpu_bo_va_mapping *
977 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
978 uint64_t addr, struct amdgpu_bo **bo)
980 struct amdgpu_bo_va_mapping *mapping;
983 if (!parser->bo_list)
986 addr /= AMDGPU_GPU_PAGE_SIZE;
988 for (i = 0; i < parser->bo_list->num_entries; i++) {
989 struct amdgpu_bo_list_entry *lobj;
991 lobj = &parser->bo_list->array[i];
995 list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
996 if (mapping->it.start > addr ||
997 addr > mapping->it.last)
1000 *bo = lobj->bo_va->bo;
1004 list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
1005 if (mapping->it.start > addr ||
1006 addr > mapping->it.last)
1009 *bo = lobj->bo_va->bo;