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.
28 #include <linux/file.h>
29 #include <linux/pagemap.h>
30 #include <linux/sync_file.h>
31 #include <linux/dma-buf.h>
33 #include <drm/amdgpu_drm.h>
34 #include <drm/drm_syncobj.h>
36 #include "amdgpu_trace.h"
37 #include "amdgpu_gmc.h"
38 #include "amdgpu_gem.h"
39 #include "amdgpu_ras.h"
41 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
42 struct drm_amdgpu_cs_chunk_fence *data,
45 struct drm_gem_object *gobj;
50 gobj = drm_gem_object_lookup(p->filp, data->handle);
54 bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
55 p->uf_entry.priority = 0;
56 p->uf_entry.tv.bo = &bo->tbo;
57 /* One for TTM and one for the CS job */
58 p->uf_entry.tv.num_shared = 2;
60 drm_gem_object_put(gobj);
62 size = amdgpu_bo_size(bo);
63 if (size != PAGE_SIZE || (data->offset + 8) > size) {
68 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
73 *offset = data->offset;
82 static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
83 struct drm_amdgpu_bo_list_in *data)
86 struct drm_amdgpu_bo_list_entry *info = NULL;
88 r = amdgpu_bo_create_list_entry_array(data, &info);
92 r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
106 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
108 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
109 struct amdgpu_vm *vm = &fpriv->vm;
110 uint64_t *chunk_array_user;
111 uint64_t *chunk_array;
112 unsigned size, num_ibs = 0;
113 uint32_t uf_offset = 0;
117 if (cs->in.num_chunks == 0)
120 chunk_array = kvmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
124 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
130 /* skip guilty context job */
131 if (atomic_read(&p->ctx->guilty) == 1) {
137 chunk_array_user = u64_to_user_ptr(cs->in.chunks);
138 if (copy_from_user(chunk_array, chunk_array_user,
139 sizeof(uint64_t)*cs->in.num_chunks)) {
144 p->nchunks = cs->in.num_chunks;
145 p->chunks = kvmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
152 for (i = 0; i < p->nchunks; i++) {
153 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
154 struct drm_amdgpu_cs_chunk user_chunk;
155 uint32_t __user *cdata;
157 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
158 if (copy_from_user(&user_chunk, chunk_ptr,
159 sizeof(struct drm_amdgpu_cs_chunk))) {
162 goto free_partial_kdata;
164 p->chunks[i].chunk_id = user_chunk.chunk_id;
165 p->chunks[i].length_dw = user_chunk.length_dw;
167 size = p->chunks[i].length_dw;
168 cdata = u64_to_user_ptr(user_chunk.chunk_data);
170 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
171 if (p->chunks[i].kdata == NULL) {
174 goto free_partial_kdata;
176 size *= sizeof(uint32_t);
177 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
179 goto free_partial_kdata;
182 switch (p->chunks[i].chunk_id) {
183 case AMDGPU_CHUNK_ID_IB:
187 case AMDGPU_CHUNK_ID_FENCE:
188 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
189 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
191 goto free_partial_kdata;
194 ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
197 goto free_partial_kdata;
201 case AMDGPU_CHUNK_ID_BO_HANDLES:
202 size = sizeof(struct drm_amdgpu_bo_list_in);
203 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
205 goto free_partial_kdata;
208 ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
210 goto free_partial_kdata;
214 case AMDGPU_CHUNK_ID_DEPENDENCIES:
215 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
216 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
217 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
218 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
219 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
224 goto free_partial_kdata;
228 ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
232 if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
237 if (p->uf_entry.tv.bo)
238 p->job->uf_addr = uf_offset;
241 /* Use this opportunity to fill in task info for the vm */
242 amdgpu_vm_set_task_info(vm);
250 kvfree(p->chunks[i].kdata);
260 /* Convert microseconds to bytes. */
261 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
263 if (us <= 0 || !adev->mm_stats.log2_max_MBps)
266 /* Since accum_us is incremented by a million per second, just
267 * multiply it by the number of MB/s to get the number of bytes.
269 return us << adev->mm_stats.log2_max_MBps;
272 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
274 if (!adev->mm_stats.log2_max_MBps)
277 return bytes >> adev->mm_stats.log2_max_MBps;
280 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
281 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
282 * which means it can go over the threshold once. If that happens, the driver
283 * will be in debt and no other buffer migrations can be done until that debt
286 * This approach allows moving a buffer of any size (it's important to allow
289 * The currency is simply time in microseconds and it increases as the clock
290 * ticks. The accumulated microseconds (us) are converted to bytes and
293 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
297 s64 time_us, increment_us;
298 u64 free_vram, total_vram, used_vram;
299 /* Allow a maximum of 200 accumulated ms. This is basically per-IB
302 * It means that in order to get full max MBps, at least 5 IBs per
303 * second must be submitted and not more than 200ms apart from each
306 const s64 us_upper_bound = 200000;
308 if (!adev->mm_stats.log2_max_MBps) {
314 total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
315 used_vram = amdgpu_vram_mgr_usage(&adev->mman.vram_mgr);
316 free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
318 spin_lock(&adev->mm_stats.lock);
320 /* Increase the amount of accumulated us. */
321 time_us = ktime_to_us(ktime_get());
322 increment_us = time_us - adev->mm_stats.last_update_us;
323 adev->mm_stats.last_update_us = time_us;
324 adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
327 /* This prevents the short period of low performance when the VRAM
328 * usage is low and the driver is in debt or doesn't have enough
329 * accumulated us to fill VRAM quickly.
331 * The situation can occur in these cases:
332 * - a lot of VRAM is freed by userspace
333 * - the presence of a big buffer causes a lot of evictions
334 * (solution: split buffers into smaller ones)
336 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
337 * accum_us to a positive number.
339 if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
342 /* Be more aggresive on dGPUs. Try to fill a portion of free
345 if (!(adev->flags & AMD_IS_APU))
346 min_us = bytes_to_us(adev, free_vram / 4);
348 min_us = 0; /* Reset accum_us on APUs. */
350 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
353 /* This is set to 0 if the driver is in debt to disallow (optional)
356 *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
358 /* Do the same for visible VRAM if half of it is free */
359 if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
360 u64 total_vis_vram = adev->gmc.visible_vram_size;
362 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr);
364 if (used_vis_vram < total_vis_vram) {
365 u64 free_vis_vram = total_vis_vram - used_vis_vram;
366 adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
367 increment_us, us_upper_bound);
369 if (free_vis_vram >= total_vis_vram / 2)
370 adev->mm_stats.accum_us_vis =
371 max(bytes_to_us(adev, free_vis_vram / 2),
372 adev->mm_stats.accum_us_vis);
375 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
380 spin_unlock(&adev->mm_stats.lock);
383 /* Report how many bytes have really been moved for the last command
384 * submission. This can result in a debt that can stop buffer migrations
387 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
390 spin_lock(&adev->mm_stats.lock);
391 adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
392 adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
393 spin_unlock(&adev->mm_stats.lock);
396 static int amdgpu_cs_bo_validate(void *param, struct amdgpu_bo *bo)
398 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
399 struct amdgpu_cs_parser *p = param;
400 struct ttm_operation_ctx ctx = {
401 .interruptible = true,
402 .no_wait_gpu = false,
403 .resv = bo->tbo.base.resv
408 if (bo->tbo.pin_count)
411 /* Don't move this buffer if we have depleted our allowance
412 * to move it. Don't move anything if the threshold is zero.
414 if (p->bytes_moved < p->bytes_moved_threshold &&
415 (!bo->tbo.base.dma_buf ||
416 list_empty(&bo->tbo.base.dma_buf->attachments))) {
417 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
418 (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
419 /* And don't move a CPU_ACCESS_REQUIRED BO to limited
420 * visible VRAM if we've depleted our allowance to do
423 if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
424 domain = bo->preferred_domains;
426 domain = bo->allowed_domains;
428 domain = bo->preferred_domains;
431 domain = bo->allowed_domains;
435 amdgpu_bo_placement_from_domain(bo, domain);
436 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
438 p->bytes_moved += ctx.bytes_moved;
439 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
440 amdgpu_bo_in_cpu_visible_vram(bo))
441 p->bytes_moved_vis += ctx.bytes_moved;
443 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
444 domain = bo->allowed_domains;
451 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
452 struct list_head *validated)
454 struct ttm_operation_ctx ctx = { true, false };
455 struct amdgpu_bo_list_entry *lobj;
458 list_for_each_entry(lobj, validated, tv.head) {
459 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
460 struct mm_struct *usermm;
462 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
463 if (usermm && usermm != current->mm)
466 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
467 lobj->user_invalidated && lobj->user_pages) {
468 amdgpu_bo_placement_from_domain(bo,
469 AMDGPU_GEM_DOMAIN_CPU);
470 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
474 amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
478 r = amdgpu_cs_bo_validate(p, bo);
482 kvfree(lobj->user_pages);
483 lobj->user_pages = NULL;
488 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
489 union drm_amdgpu_cs *cs)
491 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
492 struct amdgpu_vm *vm = &fpriv->vm;
493 struct amdgpu_bo_list_entry *e;
494 struct list_head duplicates;
495 struct amdgpu_bo *gds;
496 struct amdgpu_bo *gws;
497 struct amdgpu_bo *oa;
500 INIT_LIST_HEAD(&p->validated);
502 /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
503 if (cs->in.bo_list_handle) {
507 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
511 } else if (!p->bo_list) {
512 /* Create a empty bo_list when no handle is provided */
513 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
519 /* One for TTM and one for the CS job */
520 amdgpu_bo_list_for_each_entry(e, p->bo_list)
521 e->tv.num_shared = 2;
523 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
525 INIT_LIST_HEAD(&duplicates);
526 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
528 if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
529 list_add(&p->uf_entry.tv.head, &p->validated);
531 /* Get userptr backing pages. If pages are updated after registered
532 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
533 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
535 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
536 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
537 bool userpage_invalidated = false;
540 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
541 sizeof(struct page *),
542 GFP_KERNEL | __GFP_ZERO);
543 if (!e->user_pages) {
544 DRM_ERROR("kvmalloc_array failure\n");
548 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
550 kvfree(e->user_pages);
551 e->user_pages = NULL;
555 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
556 if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
557 userpage_invalidated = true;
561 e->user_invalidated = userpage_invalidated;
564 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
566 if (unlikely(r != 0)) {
567 if (r != -ERESTARTSYS)
568 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
572 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
573 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
575 e->bo_va = amdgpu_vm_bo_find(vm, bo);
577 if (bo->tbo.base.dma_buf && !amdgpu_bo_explicit_sync(bo)) {
578 e->chain = dma_fence_chain_alloc();
586 /* Move fence waiting after getting reservation lock of
587 * PD root. Then there is no need on a ctx mutex lock.
589 r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entity);
590 if (unlikely(r != 0)) {
591 if (r != -ERESTARTSYS)
592 DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
596 amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
597 &p->bytes_moved_vis_threshold);
599 p->bytes_moved_vis = 0;
601 r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
602 amdgpu_cs_bo_validate, p);
604 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
608 r = amdgpu_cs_list_validate(p, &duplicates);
612 r = amdgpu_cs_list_validate(p, &p->validated);
616 amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
619 gds = p->bo_list->gds_obj;
620 gws = p->bo_list->gws_obj;
621 oa = p->bo_list->oa_obj;
624 p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
625 p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
628 p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
629 p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
632 p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
633 p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
636 if (!r && p->uf_entry.tv.bo) {
637 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
639 r = amdgpu_ttm_alloc_gart(&uf->tbo);
640 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
645 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
646 dma_fence_chain_free(e->chain);
649 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
655 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
657 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
658 struct amdgpu_bo_list_entry *e;
661 list_for_each_entry(e, &p->validated, tv.head) {
662 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
663 struct dma_resv *resv = bo->tbo.base.resv;
664 enum amdgpu_sync_mode sync_mode;
666 sync_mode = amdgpu_bo_explicit_sync(bo) ?
667 AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
668 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, sync_mode,
677 * amdgpu_cs_parser_fini() - clean parser states
678 * @parser: parser structure holding parsing context.
679 * @error: error number
680 * @backoff: indicator to backoff the reservation
682 * If error is set then unvalidate buffer, otherwise just free memory
683 * used by parsing context.
685 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
690 if (error && backoff) {
691 struct amdgpu_bo_list_entry *e;
693 amdgpu_bo_list_for_each_entry(e, parser->bo_list) {
694 dma_fence_chain_free(e->chain);
698 ttm_eu_backoff_reservation(&parser->ticket,
702 for (i = 0; i < parser->num_post_deps; i++) {
703 drm_syncobj_put(parser->post_deps[i].syncobj);
704 kfree(parser->post_deps[i].chain);
706 kfree(parser->post_deps);
708 dma_fence_put(parser->fence);
711 amdgpu_ctx_put(parser->ctx);
714 amdgpu_bo_list_put(parser->bo_list);
716 for (i = 0; i < parser->nchunks; i++)
717 kvfree(parser->chunks[i].kdata);
718 kvfree(parser->chunks);
720 amdgpu_job_free(parser->job);
721 if (parser->uf_entry.tv.bo) {
722 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
724 amdgpu_bo_unref(&uf);
728 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
730 struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
731 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
732 struct amdgpu_device *adev = p->adev;
733 struct amdgpu_vm *vm = &fpriv->vm;
734 struct amdgpu_bo_list_entry *e;
735 struct amdgpu_bo_va *bo_va;
736 struct amdgpu_bo *bo;
739 /* Only for UVD/VCE VM emulation */
740 if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
743 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
744 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
745 struct amdgpu_bo_va_mapping *m;
746 struct amdgpu_bo *aobj = NULL;
747 struct amdgpu_cs_chunk *chunk;
748 uint64_t offset, va_start;
749 struct amdgpu_ib *ib;
752 chunk = &p->chunks[i];
753 ib = &p->job->ibs[j];
754 chunk_ib = chunk->kdata;
756 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
759 va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
760 r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
762 DRM_ERROR("IB va_start is invalid\n");
766 if ((va_start + chunk_ib->ib_bytes) >
767 (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
768 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
772 /* the IB should be reserved at this point */
773 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
778 offset = m->start * AMDGPU_GPU_PAGE_SIZE;
779 kptr += va_start - offset;
781 if (ring->funcs->parse_cs) {
782 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
783 amdgpu_bo_kunmap(aobj);
785 r = amdgpu_ring_parse_cs(ring, p, j);
789 ib->ptr = (uint32_t *)kptr;
790 r = amdgpu_ring_patch_cs_in_place(ring, p, j);
791 amdgpu_bo_kunmap(aobj);
801 return amdgpu_cs_sync_rings(p);
804 r = amdgpu_vm_clear_freed(adev, vm, NULL);
808 r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false, NULL);
812 r = amdgpu_sync_vm_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
816 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
817 bo_va = fpriv->csa_va;
819 r = amdgpu_vm_bo_update(adev, bo_va, false, NULL);
823 r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
828 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
829 /* ignore duplicates */
830 bo = ttm_to_amdgpu_bo(e->tv.bo);
838 r = amdgpu_vm_bo_update(adev, bo_va, false, NULL);
842 r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
847 r = amdgpu_vm_handle_moved(adev, vm);
851 r = amdgpu_vm_update_pdes(adev, vm, false);
855 r = amdgpu_sync_vm_fence(&p->job->sync, vm->last_update);
859 p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
861 if (amdgpu_vm_debug) {
862 /* Invalidate all BOs to test for userspace bugs */
863 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
864 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
866 /* ignore duplicates */
870 amdgpu_vm_bo_invalidate(adev, bo, false);
874 return amdgpu_cs_sync_rings(p);
877 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
878 struct amdgpu_cs_parser *parser)
880 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
881 struct amdgpu_vm *vm = &fpriv->vm;
882 int r, ce_preempt = 0, de_preempt = 0;
883 struct amdgpu_ring *ring;
886 for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
887 struct amdgpu_cs_chunk *chunk;
888 struct amdgpu_ib *ib;
889 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
890 struct drm_sched_entity *entity;
892 chunk = &parser->chunks[i];
893 ib = &parser->job->ibs[j];
894 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
896 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
899 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
900 (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
901 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
902 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
908 /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
909 if (ce_preempt > 1 || de_preempt > 1)
913 r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
914 chunk_ib->ip_instance, chunk_ib->ring,
919 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
920 parser->job->preamble_status |=
921 AMDGPU_PREAMBLE_IB_PRESENT;
923 if (parser->entity && parser->entity != entity)
926 /* Return if there is no run queue associated with this entity.
927 * Possibly because of disabled HW IP*/
928 if (entity->rq == NULL)
931 parser->entity = entity;
933 ring = to_amdgpu_ring(entity->rq->sched);
934 r = amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
935 chunk_ib->ib_bytes : 0,
936 AMDGPU_IB_POOL_DELAYED, ib);
938 DRM_ERROR("Failed to get ib !\n");
942 ib->gpu_addr = chunk_ib->va_start;
943 ib->length_dw = chunk_ib->ib_bytes / 4;
944 ib->flags = chunk_ib->flags;
949 /* MM engine doesn't support user fences */
950 ring = to_amdgpu_ring(parser->entity->rq->sched);
951 if (parser->job->uf_addr && ring->funcs->no_user_fence)
957 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
958 struct amdgpu_cs_chunk *chunk)
960 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
963 struct drm_amdgpu_cs_chunk_dep *deps;
965 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
966 num_deps = chunk->length_dw * 4 /
967 sizeof(struct drm_amdgpu_cs_chunk_dep);
969 for (i = 0; i < num_deps; ++i) {
970 struct amdgpu_ctx *ctx;
971 struct drm_sched_entity *entity;
972 struct dma_fence *fence;
974 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
978 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
980 deps[i].ring, &entity);
986 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
990 return PTR_ERR(fence);
994 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
995 struct drm_sched_fence *s_fence;
996 struct dma_fence *old = fence;
998 s_fence = to_drm_sched_fence(fence);
999 fence = dma_fence_get(&s_fence->scheduled);
1003 r = amdgpu_sync_fence(&p->job->sync, fence);
1004 dma_fence_put(fence);
1011 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1012 uint32_t handle, u64 point,
1015 struct dma_fence *fence;
1018 r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1020 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
1025 r = amdgpu_sync_fence(&p->job->sync, fence);
1026 dma_fence_put(fence);
1031 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1032 struct amdgpu_cs_chunk *chunk)
1034 struct drm_amdgpu_cs_chunk_sem *deps;
1038 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1039 num_deps = chunk->length_dw * 4 /
1040 sizeof(struct drm_amdgpu_cs_chunk_sem);
1041 for (i = 0; i < num_deps; ++i) {
1042 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1052 static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1053 struct amdgpu_cs_chunk *chunk)
1055 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1059 syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1060 num_deps = chunk->length_dw * 4 /
1061 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1062 for (i = 0; i < num_deps; ++i) {
1063 r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1064 syncobj_deps[i].handle,
1065 syncobj_deps[i].point,
1066 syncobj_deps[i].flags);
1074 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1075 struct amdgpu_cs_chunk *chunk)
1077 struct drm_amdgpu_cs_chunk_sem *deps;
1081 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1082 num_deps = chunk->length_dw * 4 /
1083 sizeof(struct drm_amdgpu_cs_chunk_sem);
1088 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1090 p->num_post_deps = 0;
1096 for (i = 0; i < num_deps; ++i) {
1097 p->post_deps[i].syncobj =
1098 drm_syncobj_find(p->filp, deps[i].handle);
1099 if (!p->post_deps[i].syncobj)
1101 p->post_deps[i].chain = NULL;
1102 p->post_deps[i].point = 0;
1110 static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1111 struct amdgpu_cs_chunk *chunk)
1113 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1117 syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1118 num_deps = chunk->length_dw * 4 /
1119 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1124 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1126 p->num_post_deps = 0;
1131 for (i = 0; i < num_deps; ++i) {
1132 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1135 if (syncobj_deps[i].point) {
1136 dep->chain = dma_fence_chain_alloc();
1141 dep->syncobj = drm_syncobj_find(p->filp,
1142 syncobj_deps[i].handle);
1143 if (!dep->syncobj) {
1144 dma_fence_chain_free(dep->chain);
1147 dep->point = syncobj_deps[i].point;
1154 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1155 struct amdgpu_cs_parser *p)
1159 for (i = 0; i < p->nchunks; ++i) {
1160 struct amdgpu_cs_chunk *chunk;
1162 chunk = &p->chunks[i];
1164 switch (chunk->chunk_id) {
1165 case AMDGPU_CHUNK_ID_DEPENDENCIES:
1166 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1167 r = amdgpu_cs_process_fence_dep(p, chunk);
1171 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1172 r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1176 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1177 r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1181 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1182 r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1186 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1187 r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1197 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1201 for (i = 0; i < p->num_post_deps; ++i) {
1202 if (p->post_deps[i].chain && p->post_deps[i].point) {
1203 drm_syncobj_add_point(p->post_deps[i].syncobj,
1204 p->post_deps[i].chain,
1205 p->fence, p->post_deps[i].point);
1206 p->post_deps[i].chain = NULL;
1208 drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1214 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1215 union drm_amdgpu_cs *cs)
1217 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1218 struct drm_sched_entity *entity = p->entity;
1219 struct amdgpu_bo_list_entry *e;
1220 struct amdgpu_job *job;
1227 r = drm_sched_job_init(&job->base, entity, &fpriv->vm);
1231 drm_sched_job_arm(&job->base);
1233 /* No memory allocation is allowed while holding the notifier lock.
1234 * The lock is held until amdgpu_cs_submit is finished and fence is
1237 mutex_lock(&p->adev->notifier_lock);
1239 /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1240 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1242 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1243 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1245 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1252 p->fence = dma_fence_get(&job->base.s_fence->finished);
1254 amdgpu_ctx_add_fence(p->ctx, entity, p->fence, &seq);
1255 amdgpu_cs_post_dependencies(p);
1257 if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1258 !p->ctx->preamble_presented) {
1259 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1260 p->ctx->preamble_presented = true;
1263 cs->out.handle = seq;
1264 job->uf_sequence = seq;
1266 amdgpu_job_free_resources(job);
1268 trace_amdgpu_cs_ioctl(job);
1269 amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1270 drm_sched_entity_push_job(&job->base);
1272 amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1274 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1275 struct dma_resv *resv = e->tv.bo->base.resv;
1276 struct dma_fence_chain *chain = e->chain;
1282 * Work around dma_resv shortcommings by wrapping up the
1283 * submission in a dma_fence_chain and add it as exclusive
1286 dma_fence_chain_init(chain, dma_resv_excl_fence(resv),
1287 dma_fence_get(p->fence), 1);
1289 rcu_assign_pointer(resv->fence_excl, &chain->base);
1293 ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1294 mutex_unlock(&p->adev->notifier_lock);
1299 drm_sched_job_cleanup(&job->base);
1300 mutex_unlock(&p->adev->notifier_lock);
1303 amdgpu_job_free(job);
1307 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *parser)
1311 if (!trace_amdgpu_cs_enabled())
1314 for (i = 0; i < parser->job->num_ibs; i++)
1315 trace_amdgpu_cs(parser, i);
1318 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1320 struct amdgpu_device *adev = drm_to_adev(dev);
1321 union drm_amdgpu_cs *cs = data;
1322 struct amdgpu_cs_parser parser = {};
1323 bool reserved_buffers = false;
1326 if (amdgpu_ras_intr_triggered())
1329 if (!adev->accel_working)
1335 r = amdgpu_cs_parser_init(&parser, data);
1337 if (printk_ratelimit())
1338 DRM_ERROR("Failed to initialize parser %d!\n", r);
1342 r = amdgpu_cs_ib_fill(adev, &parser);
1346 r = amdgpu_cs_dependencies(adev, &parser);
1348 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1352 r = amdgpu_cs_parser_bos(&parser, data);
1355 DRM_ERROR("Not enough memory for command submission!\n");
1356 else if (r != -ERESTARTSYS && r != -EAGAIN)
1357 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1361 reserved_buffers = true;
1363 trace_amdgpu_cs_ibs(&parser);
1365 r = amdgpu_cs_vm_handling(&parser);
1369 r = amdgpu_cs_submit(&parser, cs);
1371 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1377 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1380 * @data: data from userspace
1381 * @filp: file private
1383 * Wait for the command submission identified by handle to finish.
1385 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1386 struct drm_file *filp)
1388 union drm_amdgpu_wait_cs *wait = data;
1389 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1390 struct drm_sched_entity *entity;
1391 struct amdgpu_ctx *ctx;
1392 struct dma_fence *fence;
1395 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1399 r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1400 wait->in.ring, &entity);
1402 amdgpu_ctx_put(ctx);
1406 fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1410 r = dma_fence_wait_timeout(fence, true, timeout);
1411 if (r > 0 && fence->error)
1413 dma_fence_put(fence);
1417 amdgpu_ctx_put(ctx);
1421 memset(wait, 0, sizeof(*wait));
1422 wait->out.status = (r == 0);
1428 * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1430 * @adev: amdgpu device
1431 * @filp: file private
1432 * @user: drm_amdgpu_fence copied from user space
1434 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1435 struct drm_file *filp,
1436 struct drm_amdgpu_fence *user)
1438 struct drm_sched_entity *entity;
1439 struct amdgpu_ctx *ctx;
1440 struct dma_fence *fence;
1443 ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1445 return ERR_PTR(-EINVAL);
1447 r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1448 user->ring, &entity);
1450 amdgpu_ctx_put(ctx);
1454 fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1455 amdgpu_ctx_put(ctx);
1460 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1461 struct drm_file *filp)
1463 struct amdgpu_device *adev = drm_to_adev(dev);
1464 union drm_amdgpu_fence_to_handle *info = data;
1465 struct dma_fence *fence;
1466 struct drm_syncobj *syncobj;
1467 struct sync_file *sync_file;
1470 fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1472 return PTR_ERR(fence);
1475 fence = dma_fence_get_stub();
1477 switch (info->in.what) {
1478 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1479 r = drm_syncobj_create(&syncobj, 0, fence);
1480 dma_fence_put(fence);
1483 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1484 drm_syncobj_put(syncobj);
1487 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1488 r = drm_syncobj_create(&syncobj, 0, fence);
1489 dma_fence_put(fence);
1492 r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1493 drm_syncobj_put(syncobj);
1496 case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1497 fd = get_unused_fd_flags(O_CLOEXEC);
1499 dma_fence_put(fence);
1503 sync_file = sync_file_create(fence);
1504 dma_fence_put(fence);
1510 fd_install(fd, sync_file->file);
1511 info->out.handle = fd;
1515 dma_fence_put(fence);
1521 * amdgpu_cs_wait_all_fences - wait on all fences to signal
1523 * @adev: amdgpu device
1524 * @filp: file private
1525 * @wait: wait parameters
1526 * @fences: array of drm_amdgpu_fence
1528 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1529 struct drm_file *filp,
1530 union drm_amdgpu_wait_fences *wait,
1531 struct drm_amdgpu_fence *fences)
1533 uint32_t fence_count = wait->in.fence_count;
1537 for (i = 0; i < fence_count; i++) {
1538 struct dma_fence *fence;
1539 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1541 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1543 return PTR_ERR(fence);
1547 r = dma_fence_wait_timeout(fence, true, timeout);
1548 dma_fence_put(fence);
1556 return fence->error;
1559 memset(wait, 0, sizeof(*wait));
1560 wait->out.status = (r > 0);
1566 * amdgpu_cs_wait_any_fence - wait on any fence to signal
1568 * @adev: amdgpu device
1569 * @filp: file private
1570 * @wait: wait parameters
1571 * @fences: array of drm_amdgpu_fence
1573 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1574 struct drm_file *filp,
1575 union drm_amdgpu_wait_fences *wait,
1576 struct drm_amdgpu_fence *fences)
1578 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1579 uint32_t fence_count = wait->in.fence_count;
1580 uint32_t first = ~0;
1581 struct dma_fence **array;
1585 /* Prepare the fence array */
1586 array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1591 for (i = 0; i < fence_count; i++) {
1592 struct dma_fence *fence;
1594 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1595 if (IS_ERR(fence)) {
1597 goto err_free_fence_array;
1600 } else { /* NULL, the fence has been already signaled */
1607 r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1610 goto err_free_fence_array;
1613 memset(wait, 0, sizeof(*wait));
1614 wait->out.status = (r > 0);
1615 wait->out.first_signaled = first;
1617 if (first < fence_count && array[first])
1618 r = array[first]->error;
1622 err_free_fence_array:
1623 for (i = 0; i < fence_count; i++)
1624 dma_fence_put(array[i]);
1631 * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1634 * @data: data from userspace
1635 * @filp: file private
1637 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1638 struct drm_file *filp)
1640 struct amdgpu_device *adev = drm_to_adev(dev);
1641 union drm_amdgpu_wait_fences *wait = data;
1642 uint32_t fence_count = wait->in.fence_count;
1643 struct drm_amdgpu_fence *fences_user;
1644 struct drm_amdgpu_fence *fences;
1647 /* Get the fences from userspace */
1648 fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1653 fences_user = u64_to_user_ptr(wait->in.fences);
1654 if (copy_from_user(fences, fences_user,
1655 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1657 goto err_free_fences;
1660 if (wait->in.wait_all)
1661 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1663 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1672 * amdgpu_cs_find_mapping - find bo_va for VM address
1674 * @parser: command submission parser context
1676 * @bo: resulting BO of the mapping found
1677 * @map: Placeholder to return found BO mapping
1679 * Search the buffer objects in the command submission context for a certain
1680 * virtual memory address. Returns allocation structure when found, NULL
1683 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1684 uint64_t addr, struct amdgpu_bo **bo,
1685 struct amdgpu_bo_va_mapping **map)
1687 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1688 struct ttm_operation_ctx ctx = { false, false };
1689 struct amdgpu_vm *vm = &fpriv->vm;
1690 struct amdgpu_bo_va_mapping *mapping;
1693 addr /= AMDGPU_GPU_PAGE_SIZE;
1695 mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1696 if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1699 *bo = mapping->bo_va->base.bo;
1702 /* Double check that the BO is reserved by this CS */
1703 if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1706 if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1707 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1708 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1709 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1714 return amdgpu_ttm_alloc_gart(&(*bo)->tbo);