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>
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>
35 #include "amdgpu_cs.h"
37 #include "amdgpu_trace.h"
38 #include "amdgpu_gmc.h"
39 #include "amdgpu_gem.h"
40 #include "amdgpu_ras.h"
42 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p,
43 struct amdgpu_device *adev,
44 struct drm_file *filp,
45 union drm_amdgpu_cs *cs)
47 struct amdgpu_fpriv *fpriv = filp->driver_priv;
49 if (cs->in.num_chunks == 0)
52 memset(p, 0, sizeof(*p));
56 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
60 if (atomic_read(&p->ctx->guilty)) {
61 amdgpu_ctx_put(p->ctx);
67 static int amdgpu_cs_job_idx(struct amdgpu_cs_parser *p,
68 struct drm_amdgpu_cs_chunk_ib *chunk_ib)
70 struct drm_sched_entity *entity;
74 r = amdgpu_ctx_get_entity(p->ctx, chunk_ib->ip_type,
75 chunk_ib->ip_instance,
76 chunk_ib->ring, &entity);
81 * Abort if there is no run queue associated with this entity.
82 * Possibly because of disabled HW IP.
84 if (entity->rq == NULL)
87 /* Check if we can add this IB to some existing job */
88 for (i = 0; i < p->gang_size; ++i)
89 if (p->entities[i] == entity)
92 /* If not increase the gang size if possible */
93 if (i == AMDGPU_CS_GANG_SIZE)
96 p->entities[i] = entity;
101 static int amdgpu_cs_p1_ib(struct amdgpu_cs_parser *p,
102 struct drm_amdgpu_cs_chunk_ib *chunk_ib,
103 unsigned int *num_ibs)
107 r = amdgpu_cs_job_idx(p, chunk_ib);
112 p->gang_leader_idx = r;
116 static int amdgpu_cs_p1_user_fence(struct amdgpu_cs_parser *p,
117 struct drm_amdgpu_cs_chunk_fence *data,
120 struct drm_gem_object *gobj;
121 struct amdgpu_bo *bo;
125 gobj = drm_gem_object_lookup(p->filp, data->handle);
129 bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
130 p->uf_entry.priority = 0;
131 p->uf_entry.tv.bo = &bo->tbo;
132 /* One for TTM and two for the CS job */
133 p->uf_entry.tv.num_shared = 3;
135 drm_gem_object_put(gobj);
137 size = amdgpu_bo_size(bo);
138 if (size != PAGE_SIZE || (data->offset + 8) > size) {
143 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
148 *offset = data->offset;
153 amdgpu_bo_unref(&bo);
157 static int amdgpu_cs_p1_bo_handles(struct amdgpu_cs_parser *p,
158 struct drm_amdgpu_bo_list_in *data)
160 struct drm_amdgpu_bo_list_entry *info;
163 r = amdgpu_bo_create_list_entry_array(data, &info);
167 r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
181 /* Copy the data from userspace and go over it the first time */
182 static int amdgpu_cs_pass1(struct amdgpu_cs_parser *p,
183 union drm_amdgpu_cs *cs)
185 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
186 unsigned int num_ibs[AMDGPU_CS_GANG_SIZE] = { };
187 struct amdgpu_vm *vm = &fpriv->vm;
188 uint64_t *chunk_array_user;
189 uint64_t *chunk_array;
190 uint32_t uf_offset = 0;
195 chunk_array = kvmalloc_array(cs->in.num_chunks, sizeof(uint64_t),
201 chunk_array_user = u64_to_user_ptr(cs->in.chunks);
202 if (copy_from_user(chunk_array, chunk_array_user,
203 sizeof(uint64_t)*cs->in.num_chunks)) {
208 p->nchunks = cs->in.num_chunks;
209 p->chunks = kvmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
216 for (i = 0; i < p->nchunks; i++) {
217 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
218 struct drm_amdgpu_cs_chunk user_chunk;
219 uint32_t __user *cdata;
221 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
222 if (copy_from_user(&user_chunk, chunk_ptr,
223 sizeof(struct drm_amdgpu_cs_chunk))) {
226 goto free_partial_kdata;
228 p->chunks[i].chunk_id = user_chunk.chunk_id;
229 p->chunks[i].length_dw = user_chunk.length_dw;
231 size = p->chunks[i].length_dw;
232 cdata = u64_to_user_ptr(user_chunk.chunk_data);
234 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t),
236 if (p->chunks[i].kdata == NULL) {
239 goto free_partial_kdata;
241 size *= sizeof(uint32_t);
242 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
244 goto free_partial_kdata;
247 /* Assume the worst on the following checks */
249 switch (p->chunks[i].chunk_id) {
250 case AMDGPU_CHUNK_ID_IB:
251 if (size < sizeof(struct drm_amdgpu_cs_chunk_ib))
252 goto free_partial_kdata;
254 ret = amdgpu_cs_p1_ib(p, p->chunks[i].kdata, num_ibs);
256 goto free_partial_kdata;
259 case AMDGPU_CHUNK_ID_FENCE:
260 if (size < sizeof(struct drm_amdgpu_cs_chunk_fence))
261 goto free_partial_kdata;
263 ret = amdgpu_cs_p1_user_fence(p, p->chunks[i].kdata,
266 goto free_partial_kdata;
269 case AMDGPU_CHUNK_ID_BO_HANDLES:
270 if (size < sizeof(struct drm_amdgpu_bo_list_in))
271 goto free_partial_kdata;
273 ret = amdgpu_cs_p1_bo_handles(p, p->chunks[i].kdata);
275 goto free_partial_kdata;
278 case AMDGPU_CHUNK_ID_DEPENDENCIES:
279 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
280 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
281 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
282 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
283 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
287 goto free_partial_kdata;
294 for (i = 0; i < p->gang_size; ++i) {
295 ret = amdgpu_job_alloc(p->adev, num_ibs[i], &p->jobs[i], vm);
299 ret = drm_sched_job_init(&p->jobs[i]->base, p->entities[i],
304 p->gang_leader = p->jobs[p->gang_leader_idx];
306 if (p->ctx->vram_lost_counter != p->gang_leader->vram_lost_counter) {
311 if (p->uf_entry.tv.bo)
312 p->gang_leader->uf_addr = uf_offset;
315 /* Use this opportunity to fill in task info for the vm */
316 amdgpu_vm_set_task_info(vm);
324 kvfree(p->chunks[i].kdata);
334 static int amdgpu_cs_p2_ib(struct amdgpu_cs_parser *p,
335 struct amdgpu_cs_chunk *chunk,
336 unsigned int *ce_preempt,
337 unsigned int *de_preempt)
339 struct drm_amdgpu_cs_chunk_ib *chunk_ib = chunk->kdata;
340 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
341 struct amdgpu_vm *vm = &fpriv->vm;
342 struct amdgpu_ring *ring;
343 struct amdgpu_job *job;
344 struct amdgpu_ib *ib;
347 r = amdgpu_cs_job_idx(p, chunk_ib);
352 ring = amdgpu_job_ring(job);
353 ib = &job->ibs[job->num_ibs++];
355 /* MM engine doesn't support user fences */
356 if (p->uf_entry.tv.bo && ring->funcs->no_user_fence)
359 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
360 chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
361 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
366 /* Each GFX command submit allows only 1 IB max
367 * preemptible for CE & DE */
368 if (*ce_preempt > 1 || *de_preempt > 1)
372 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
373 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
375 r = amdgpu_ib_get(p->adev, vm, ring->funcs->parse_cs ?
376 chunk_ib->ib_bytes : 0,
377 AMDGPU_IB_POOL_DELAYED, ib);
379 DRM_ERROR("Failed to get ib !\n");
383 ib->gpu_addr = chunk_ib->va_start;
384 ib->length_dw = chunk_ib->ib_bytes / 4;
385 ib->flags = chunk_ib->flags;
389 static int amdgpu_cs_p2_dependencies(struct amdgpu_cs_parser *p,
390 struct amdgpu_cs_chunk *chunk)
392 struct drm_amdgpu_cs_chunk_dep *deps = chunk->kdata;
393 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
397 num_deps = chunk->length_dw * 4 /
398 sizeof(struct drm_amdgpu_cs_chunk_dep);
400 for (i = 0; i < num_deps; ++i) {
401 struct amdgpu_ctx *ctx;
402 struct drm_sched_entity *entity;
403 struct dma_fence *fence;
405 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
409 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
411 deps[i].ring, &entity);
417 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
421 return PTR_ERR(fence);
425 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
426 struct drm_sched_fence *s_fence;
427 struct dma_fence *old = fence;
429 s_fence = to_drm_sched_fence(fence);
430 fence = dma_fence_get(&s_fence->scheduled);
434 r = amdgpu_sync_fence(&p->gang_leader->sync, fence);
435 dma_fence_put(fence);
442 static int amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser *p,
443 uint32_t handle, u64 point,
446 struct dma_fence *fence;
449 r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
451 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
456 r = amdgpu_sync_fence(&p->gang_leader->sync, fence);
457 dma_fence_put(fence);
462 static int amdgpu_cs_p2_syncobj_in(struct amdgpu_cs_parser *p,
463 struct amdgpu_cs_chunk *chunk)
465 struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
469 num_deps = chunk->length_dw * 4 /
470 sizeof(struct drm_amdgpu_cs_chunk_sem);
471 for (i = 0; i < num_deps; ++i) {
472 r = amdgpu_syncobj_lookup_and_add(p, deps[i].handle, 0, 0);
480 static int amdgpu_cs_p2_syncobj_timeline_wait(struct amdgpu_cs_parser *p,
481 struct amdgpu_cs_chunk *chunk)
483 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
487 num_deps = chunk->length_dw * 4 /
488 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
489 for (i = 0; i < num_deps; ++i) {
490 r = amdgpu_syncobj_lookup_and_add(p, syncobj_deps[i].handle,
491 syncobj_deps[i].point,
492 syncobj_deps[i].flags);
500 static int amdgpu_cs_p2_syncobj_out(struct amdgpu_cs_parser *p,
501 struct amdgpu_cs_chunk *chunk)
503 struct drm_amdgpu_cs_chunk_sem *deps = chunk->kdata;
507 num_deps = chunk->length_dw * 4 /
508 sizeof(struct drm_amdgpu_cs_chunk_sem);
513 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
515 p->num_post_deps = 0;
521 for (i = 0; i < num_deps; ++i) {
522 p->post_deps[i].syncobj =
523 drm_syncobj_find(p->filp, deps[i].handle);
524 if (!p->post_deps[i].syncobj)
526 p->post_deps[i].chain = NULL;
527 p->post_deps[i].point = 0;
534 static int amdgpu_cs_p2_syncobj_timeline_signal(struct amdgpu_cs_parser *p,
535 struct amdgpu_cs_chunk *chunk)
537 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps = chunk->kdata;
541 num_deps = chunk->length_dw * 4 /
542 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
547 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
549 p->num_post_deps = 0;
554 for (i = 0; i < num_deps; ++i) {
555 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
558 if (syncobj_deps[i].point) {
559 dep->chain = dma_fence_chain_alloc();
564 dep->syncobj = drm_syncobj_find(p->filp,
565 syncobj_deps[i].handle);
567 dma_fence_chain_free(dep->chain);
570 dep->point = syncobj_deps[i].point;
577 static int amdgpu_cs_pass2(struct amdgpu_cs_parser *p)
579 unsigned int ce_preempt = 0, de_preempt = 0;
582 for (i = 0; i < p->nchunks; ++i) {
583 struct amdgpu_cs_chunk *chunk;
585 chunk = &p->chunks[i];
587 switch (chunk->chunk_id) {
588 case AMDGPU_CHUNK_ID_IB:
589 r = amdgpu_cs_p2_ib(p, chunk, &ce_preempt, &de_preempt);
593 case AMDGPU_CHUNK_ID_DEPENDENCIES:
594 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
595 r = amdgpu_cs_p2_dependencies(p, chunk);
599 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
600 r = amdgpu_cs_p2_syncobj_in(p, chunk);
604 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
605 r = amdgpu_cs_p2_syncobj_out(p, chunk);
609 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
610 r = amdgpu_cs_p2_syncobj_timeline_wait(p, chunk);
614 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
615 r = amdgpu_cs_p2_syncobj_timeline_signal(p, chunk);
625 /* Convert microseconds to bytes. */
626 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
628 if (us <= 0 || !adev->mm_stats.log2_max_MBps)
631 /* Since accum_us is incremented by a million per second, just
632 * multiply it by the number of MB/s to get the number of bytes.
634 return us << adev->mm_stats.log2_max_MBps;
637 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
639 if (!adev->mm_stats.log2_max_MBps)
642 return bytes >> adev->mm_stats.log2_max_MBps;
645 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
646 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
647 * which means it can go over the threshold once. If that happens, the driver
648 * will be in debt and no other buffer migrations can be done until that debt
651 * This approach allows moving a buffer of any size (it's important to allow
654 * The currency is simply time in microseconds and it increases as the clock
655 * ticks. The accumulated microseconds (us) are converted to bytes and
658 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
662 s64 time_us, increment_us;
663 u64 free_vram, total_vram, used_vram;
664 /* Allow a maximum of 200 accumulated ms. This is basically per-IB
667 * It means that in order to get full max MBps, at least 5 IBs per
668 * second must be submitted and not more than 200ms apart from each
671 const s64 us_upper_bound = 200000;
673 if (!adev->mm_stats.log2_max_MBps) {
679 total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
680 used_vram = ttm_resource_manager_usage(&adev->mman.vram_mgr.manager);
681 free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
683 spin_lock(&adev->mm_stats.lock);
685 /* Increase the amount of accumulated us. */
686 time_us = ktime_to_us(ktime_get());
687 increment_us = time_us - adev->mm_stats.last_update_us;
688 adev->mm_stats.last_update_us = time_us;
689 adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
692 /* This prevents the short period of low performance when the VRAM
693 * usage is low and the driver is in debt or doesn't have enough
694 * accumulated us to fill VRAM quickly.
696 * The situation can occur in these cases:
697 * - a lot of VRAM is freed by userspace
698 * - the presence of a big buffer causes a lot of evictions
699 * (solution: split buffers into smaller ones)
701 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
702 * accum_us to a positive number.
704 if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
707 /* Be more aggressive on dGPUs. Try to fill a portion of free
710 if (!(adev->flags & AMD_IS_APU))
711 min_us = bytes_to_us(adev, free_vram / 4);
713 min_us = 0; /* Reset accum_us on APUs. */
715 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
718 /* This is set to 0 if the driver is in debt to disallow (optional)
721 *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
723 /* Do the same for visible VRAM if half of it is free */
724 if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
725 u64 total_vis_vram = adev->gmc.visible_vram_size;
727 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr);
729 if (used_vis_vram < total_vis_vram) {
730 u64 free_vis_vram = total_vis_vram - used_vis_vram;
731 adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
732 increment_us, us_upper_bound);
734 if (free_vis_vram >= total_vis_vram / 2)
735 adev->mm_stats.accum_us_vis =
736 max(bytes_to_us(adev, free_vis_vram / 2),
737 adev->mm_stats.accum_us_vis);
740 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
745 spin_unlock(&adev->mm_stats.lock);
748 /* Report how many bytes have really been moved for the last command
749 * submission. This can result in a debt that can stop buffer migrations
752 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
755 spin_lock(&adev->mm_stats.lock);
756 adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
757 adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
758 spin_unlock(&adev->mm_stats.lock);
761 static int amdgpu_cs_bo_validate(void *param, struct amdgpu_bo *bo)
763 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
764 struct amdgpu_cs_parser *p = param;
765 struct ttm_operation_ctx ctx = {
766 .interruptible = true,
767 .no_wait_gpu = false,
768 .resv = bo->tbo.base.resv
773 if (bo->tbo.pin_count)
776 /* Don't move this buffer if we have depleted our allowance
777 * to move it. Don't move anything if the threshold is zero.
779 if (p->bytes_moved < p->bytes_moved_threshold &&
780 (!bo->tbo.base.dma_buf ||
781 list_empty(&bo->tbo.base.dma_buf->attachments))) {
782 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
783 (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
784 /* And don't move a CPU_ACCESS_REQUIRED BO to limited
785 * visible VRAM if we've depleted our allowance to do
788 if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
789 domain = bo->preferred_domains;
791 domain = bo->allowed_domains;
793 domain = bo->preferred_domains;
796 domain = bo->allowed_domains;
800 amdgpu_bo_placement_from_domain(bo, domain);
801 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
803 p->bytes_moved += ctx.bytes_moved;
804 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
805 amdgpu_bo_in_cpu_visible_vram(bo))
806 p->bytes_moved_vis += ctx.bytes_moved;
808 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
809 domain = bo->allowed_domains;
816 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
817 struct list_head *validated)
819 struct ttm_operation_ctx ctx = { true, false };
820 struct amdgpu_bo_list_entry *lobj;
823 list_for_each_entry(lobj, validated, tv.head) {
824 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
825 struct mm_struct *usermm;
827 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
828 if (usermm && usermm != current->mm)
831 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
832 lobj->user_invalidated && lobj->user_pages) {
833 amdgpu_bo_placement_from_domain(bo,
834 AMDGPU_GEM_DOMAIN_CPU);
835 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
839 amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
843 r = amdgpu_cs_bo_validate(p, bo);
847 kvfree(lobj->user_pages);
848 lobj->user_pages = NULL;
853 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
854 union drm_amdgpu_cs *cs)
856 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
857 struct amdgpu_vm *vm = &fpriv->vm;
858 struct amdgpu_bo_list_entry *e;
859 struct list_head duplicates;
863 INIT_LIST_HEAD(&p->validated);
865 /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
866 if (cs->in.bo_list_handle) {
870 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
874 } else if (!p->bo_list) {
875 /* Create a empty bo_list when no handle is provided */
876 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
882 mutex_lock(&p->bo_list->bo_list_mutex);
884 /* One for TTM and one for the CS job */
885 amdgpu_bo_list_for_each_entry(e, p->bo_list)
886 e->tv.num_shared = 2;
888 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
890 INIT_LIST_HEAD(&duplicates);
891 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
893 if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
894 list_add(&p->uf_entry.tv.head, &p->validated);
896 /* Get userptr backing pages. If pages are updated after registered
897 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
898 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
900 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
901 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
902 bool userpage_invalidated = false;
905 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
906 sizeof(struct page *),
907 GFP_KERNEL | __GFP_ZERO);
908 if (!e->user_pages) {
909 DRM_ERROR("kvmalloc_array failure\n");
911 goto out_free_user_pages;
914 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
916 kvfree(e->user_pages);
917 e->user_pages = NULL;
918 goto out_free_user_pages;
921 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
922 if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
923 userpage_invalidated = true;
927 e->user_invalidated = userpage_invalidated;
930 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
932 if (unlikely(r != 0)) {
933 if (r != -ERESTARTSYS)
934 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
935 goto out_free_user_pages;
938 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
939 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
941 e->bo_va = amdgpu_vm_bo_find(vm, bo);
944 amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
945 &p->bytes_moved_vis_threshold);
947 p->bytes_moved_vis = 0;
949 r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
950 amdgpu_cs_bo_validate, p);
952 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
956 r = amdgpu_cs_list_validate(p, &duplicates);
960 r = amdgpu_cs_list_validate(p, &p->validated);
964 if (p->uf_entry.tv.bo) {
965 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
967 r = amdgpu_ttm_alloc_gart(&uf->tbo);
971 p->gang_leader->uf_addr += amdgpu_bo_gpu_offset(uf);
974 amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
977 for (i = 0; i < p->gang_size; ++i)
978 amdgpu_job_set_resources(p->jobs[i], p->bo_list->gds_obj,
984 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
987 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
988 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
992 amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
993 kvfree(e->user_pages);
994 e->user_pages = NULL;
996 mutex_unlock(&p->bo_list->bo_list_mutex);
1000 static void trace_amdgpu_cs_ibs(struct amdgpu_cs_parser *p)
1004 if (!trace_amdgpu_cs_enabled())
1007 for (i = 0; i < p->gang_size; ++i) {
1008 struct amdgpu_job *job = p->jobs[i];
1010 for (j = 0; j < job->num_ibs; ++j)
1011 trace_amdgpu_cs(p, job, &job->ibs[j]);
1015 static int amdgpu_cs_patch_ibs(struct amdgpu_cs_parser *p,
1016 struct amdgpu_job *job)
1018 struct amdgpu_ring *ring = amdgpu_job_ring(job);
1022 /* Only for UVD/VCE VM emulation */
1023 if (!ring->funcs->parse_cs && !ring->funcs->patch_cs_in_place)
1026 for (i = 0; i < job->num_ibs; ++i) {
1027 struct amdgpu_ib *ib = &job->ibs[i];
1028 struct amdgpu_bo_va_mapping *m;
1029 struct amdgpu_bo *aobj;
1033 va_start = ib->gpu_addr & AMDGPU_GMC_HOLE_MASK;
1034 r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
1036 DRM_ERROR("IB va_start is invalid\n");
1040 if ((va_start + ib->length_dw * 4) >
1041 (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
1042 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
1046 /* the IB should be reserved at this point */
1047 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
1052 kptr += va_start - (m->start * AMDGPU_GPU_PAGE_SIZE);
1054 if (ring->funcs->parse_cs) {
1055 memcpy(ib->ptr, kptr, ib->length_dw * 4);
1056 amdgpu_bo_kunmap(aobj);
1058 r = amdgpu_ring_parse_cs(ring, p, job, ib);
1062 ib->ptr = (uint32_t *)kptr;
1063 r = amdgpu_ring_patch_cs_in_place(ring, p, job, ib);
1064 amdgpu_bo_kunmap(aobj);
1073 static int amdgpu_cs_patch_jobs(struct amdgpu_cs_parser *p)
1078 for (i = 0; i < p->gang_size; ++i) {
1079 r = amdgpu_cs_patch_ibs(p, p->jobs[i]);
1086 static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
1088 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1089 struct amdgpu_job *job = p->gang_leader;
1090 struct amdgpu_device *adev = p->adev;
1091 struct amdgpu_vm *vm = &fpriv->vm;
1092 struct amdgpu_bo_list_entry *e;
1093 struct amdgpu_bo_va *bo_va;
1094 struct amdgpu_bo *bo;
1098 r = amdgpu_vm_clear_freed(adev, vm, NULL);
1102 r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
1106 r = amdgpu_sync_fence(&job->sync, fpriv->prt_va->last_pt_update);
1110 if (fpriv->csa_va) {
1111 bo_va = fpriv->csa_va;
1113 r = amdgpu_vm_bo_update(adev, bo_va, false);
1117 r = amdgpu_sync_fence(&job->sync, bo_va->last_pt_update);
1122 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1123 /* ignore duplicates */
1124 bo = ttm_to_amdgpu_bo(e->tv.bo);
1132 r = amdgpu_vm_bo_update(adev, bo_va, false);
1136 r = amdgpu_sync_fence(&job->sync, bo_va->last_pt_update);
1141 r = amdgpu_vm_handle_moved(adev, vm);
1145 r = amdgpu_vm_update_pdes(adev, vm, false);
1149 r = amdgpu_sync_fence(&job->sync, vm->last_update);
1153 for (i = 0; i < p->gang_size; ++i) {
1159 job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.bo);
1162 if (amdgpu_vm_debug) {
1163 /* Invalidate all BOs to test for userspace bugs */
1164 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
1165 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1167 /* ignore duplicates */
1171 amdgpu_vm_bo_invalidate(adev, bo, false);
1178 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
1180 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1181 struct amdgpu_job *leader = p->gang_leader;
1182 struct amdgpu_bo_list_entry *e;
1186 list_for_each_entry(e, &p->validated, tv.head) {
1187 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1188 struct dma_resv *resv = bo->tbo.base.resv;
1189 enum amdgpu_sync_mode sync_mode;
1191 sync_mode = amdgpu_bo_explicit_sync(bo) ?
1192 AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
1193 r = amdgpu_sync_resv(p->adev, &leader->sync, resv, sync_mode,
1199 for (i = 0; i < p->gang_size; ++i) {
1200 if (p->jobs[i] == leader)
1203 r = amdgpu_sync_clone(&leader->sync, &p->jobs[i]->sync);
1208 r = amdgpu_ctx_wait_prev_fence(p->ctx, p->entities[p->gang_leader_idx]);
1209 if (r && r != -ERESTARTSYS)
1210 DRM_ERROR("amdgpu_ctx_wait_prev_fence failed.\n");
1214 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1218 for (i = 0; i < p->num_post_deps; ++i) {
1219 if (p->post_deps[i].chain && p->post_deps[i].point) {
1220 drm_syncobj_add_point(p->post_deps[i].syncobj,
1221 p->post_deps[i].chain,
1222 p->fence, p->post_deps[i].point);
1223 p->post_deps[i].chain = NULL;
1225 drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1231 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1232 union drm_amdgpu_cs *cs)
1234 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1235 struct amdgpu_job *leader = p->gang_leader;
1236 struct amdgpu_bo_list_entry *e;
1241 for (i = 0; i < p->gang_size; ++i)
1242 drm_sched_job_arm(&p->jobs[i]->base);
1244 for (i = 0; i < p->gang_size; ++i) {
1245 struct dma_fence *fence;
1247 if (p->jobs[i] == leader)
1250 fence = &p->jobs[i]->base.s_fence->scheduled;
1251 r = amdgpu_sync_fence(&leader->sync, fence);
1256 if (p->gang_size > 1) {
1257 for (i = 0; i < p->gang_size; ++i)
1258 amdgpu_job_set_gang_leader(p->jobs[i], leader);
1261 /* No memory allocation is allowed while holding the notifier lock.
1262 * The lock is held until amdgpu_cs_submit is finished and fence is
1265 mutex_lock(&p->adev->notifier_lock);
1267 /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1268 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1271 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1272 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1274 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1281 p->fence = dma_fence_get(&leader->base.s_fence->finished);
1282 list_for_each_entry(e, &p->validated, tv.head) {
1284 /* Everybody except for the gang leader uses READ */
1285 for (i = 0; i < p->gang_size; ++i) {
1286 if (p->jobs[i] == leader)
1289 dma_resv_add_fence(e->tv.bo->base.resv,
1290 &p->jobs[i]->base.s_fence->finished,
1291 DMA_RESV_USAGE_READ);
1294 /* The gang leader is remembered as writer */
1295 e->tv.num_shared = 0;
1298 seq = amdgpu_ctx_add_fence(p->ctx, p->entities[p->gang_leader_idx],
1300 amdgpu_cs_post_dependencies(p);
1302 if ((leader->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1303 !p->ctx->preamble_presented) {
1304 leader->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1305 p->ctx->preamble_presented = true;
1308 cs->out.handle = seq;
1309 leader->uf_sequence = seq;
1311 amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1312 for (i = 0; i < p->gang_size; ++i) {
1313 amdgpu_job_free_resources(p->jobs[i]);
1314 trace_amdgpu_cs_ioctl(p->jobs[i]);
1315 drm_sched_entity_push_job(&p->jobs[i]->base);
1319 amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1320 ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1322 mutex_unlock(&p->adev->notifier_lock);
1323 mutex_unlock(&p->bo_list->bo_list_mutex);
1327 mutex_unlock(&p->adev->notifier_lock);
1330 for (i = 0; i < p->gang_size; ++i)
1331 drm_sched_job_cleanup(&p->jobs[i]->base);
1335 /* Cleanup the parser structure */
1336 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser)
1340 for (i = 0; i < parser->num_post_deps; i++) {
1341 drm_syncobj_put(parser->post_deps[i].syncobj);
1342 kfree(parser->post_deps[i].chain);
1344 kfree(parser->post_deps);
1346 dma_fence_put(parser->fence);
1349 amdgpu_ctx_put(parser->ctx);
1350 if (parser->bo_list)
1351 amdgpu_bo_list_put(parser->bo_list);
1353 for (i = 0; i < parser->nchunks; i++)
1354 kvfree(parser->chunks[i].kdata);
1355 kvfree(parser->chunks);
1356 for (i = 0; i < parser->gang_size; ++i) {
1357 if (parser->jobs[i])
1358 amdgpu_job_free(parser->jobs[i]);
1360 if (parser->uf_entry.tv.bo) {
1361 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
1363 amdgpu_bo_unref(&uf);
1367 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1369 struct amdgpu_device *adev = drm_to_adev(dev);
1370 struct amdgpu_cs_parser parser;
1373 if (amdgpu_ras_intr_triggered())
1376 if (!adev->accel_working)
1379 r = amdgpu_cs_parser_init(&parser, adev, filp, data);
1381 if (printk_ratelimit())
1382 DRM_ERROR("Failed to initialize parser %d!\n", r);
1386 r = amdgpu_cs_pass1(&parser, data);
1390 r = amdgpu_cs_pass2(&parser);
1394 r = amdgpu_cs_parser_bos(&parser, data);
1397 DRM_ERROR("Not enough memory for command submission!\n");
1398 else if (r != -ERESTARTSYS && r != -EAGAIN)
1399 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1403 r = amdgpu_cs_patch_jobs(&parser);
1407 r = amdgpu_cs_vm_handling(&parser);
1411 r = amdgpu_cs_sync_rings(&parser);
1415 trace_amdgpu_cs_ibs(&parser);
1417 r = amdgpu_cs_submit(&parser, data);
1421 amdgpu_cs_parser_fini(&parser);
1425 ttm_eu_backoff_reservation(&parser.ticket, &parser.validated);
1426 mutex_unlock(&parser.bo_list->bo_list_mutex);
1429 amdgpu_cs_parser_fini(&parser);
1434 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1437 * @data: data from userspace
1438 * @filp: file private
1440 * Wait for the command submission identified by handle to finish.
1442 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1443 struct drm_file *filp)
1445 union drm_amdgpu_wait_cs *wait = data;
1446 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1447 struct drm_sched_entity *entity;
1448 struct amdgpu_ctx *ctx;
1449 struct dma_fence *fence;
1452 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1456 r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1457 wait->in.ring, &entity);
1459 amdgpu_ctx_put(ctx);
1463 fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1467 r = dma_fence_wait_timeout(fence, true, timeout);
1468 if (r > 0 && fence->error)
1470 dma_fence_put(fence);
1474 amdgpu_ctx_put(ctx);
1478 memset(wait, 0, sizeof(*wait));
1479 wait->out.status = (r == 0);
1485 * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1487 * @adev: amdgpu device
1488 * @filp: file private
1489 * @user: drm_amdgpu_fence copied from user space
1491 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1492 struct drm_file *filp,
1493 struct drm_amdgpu_fence *user)
1495 struct drm_sched_entity *entity;
1496 struct amdgpu_ctx *ctx;
1497 struct dma_fence *fence;
1500 ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1502 return ERR_PTR(-EINVAL);
1504 r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1505 user->ring, &entity);
1507 amdgpu_ctx_put(ctx);
1511 fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1512 amdgpu_ctx_put(ctx);
1517 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1518 struct drm_file *filp)
1520 struct amdgpu_device *adev = drm_to_adev(dev);
1521 union drm_amdgpu_fence_to_handle *info = data;
1522 struct dma_fence *fence;
1523 struct drm_syncobj *syncobj;
1524 struct sync_file *sync_file;
1527 fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1529 return PTR_ERR(fence);
1532 fence = dma_fence_get_stub();
1534 switch (info->in.what) {
1535 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1536 r = drm_syncobj_create(&syncobj, 0, fence);
1537 dma_fence_put(fence);
1540 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1541 drm_syncobj_put(syncobj);
1544 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1545 r = drm_syncobj_create(&syncobj, 0, fence);
1546 dma_fence_put(fence);
1549 r = drm_syncobj_get_fd(syncobj, (int *)&info->out.handle);
1550 drm_syncobj_put(syncobj);
1553 case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1554 fd = get_unused_fd_flags(O_CLOEXEC);
1556 dma_fence_put(fence);
1560 sync_file = sync_file_create(fence);
1561 dma_fence_put(fence);
1567 fd_install(fd, sync_file->file);
1568 info->out.handle = fd;
1572 dma_fence_put(fence);
1578 * amdgpu_cs_wait_all_fences - wait on all fences to signal
1580 * @adev: amdgpu device
1581 * @filp: file private
1582 * @wait: wait parameters
1583 * @fences: array of drm_amdgpu_fence
1585 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1586 struct drm_file *filp,
1587 union drm_amdgpu_wait_fences *wait,
1588 struct drm_amdgpu_fence *fences)
1590 uint32_t fence_count = wait->in.fence_count;
1594 for (i = 0; i < fence_count; i++) {
1595 struct dma_fence *fence;
1596 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1598 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1600 return PTR_ERR(fence);
1604 r = dma_fence_wait_timeout(fence, true, timeout);
1605 dma_fence_put(fence);
1613 return fence->error;
1616 memset(wait, 0, sizeof(*wait));
1617 wait->out.status = (r > 0);
1623 * amdgpu_cs_wait_any_fence - wait on any fence to signal
1625 * @adev: amdgpu device
1626 * @filp: file private
1627 * @wait: wait parameters
1628 * @fences: array of drm_amdgpu_fence
1630 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1631 struct drm_file *filp,
1632 union drm_amdgpu_wait_fences *wait,
1633 struct drm_amdgpu_fence *fences)
1635 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1636 uint32_t fence_count = wait->in.fence_count;
1637 uint32_t first = ~0;
1638 struct dma_fence **array;
1642 /* Prepare the fence array */
1643 array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1648 for (i = 0; i < fence_count; i++) {
1649 struct dma_fence *fence;
1651 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1652 if (IS_ERR(fence)) {
1654 goto err_free_fence_array;
1657 } else { /* NULL, the fence has been already signaled */
1664 r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1667 goto err_free_fence_array;
1670 memset(wait, 0, sizeof(*wait));
1671 wait->out.status = (r > 0);
1672 wait->out.first_signaled = first;
1674 if (first < fence_count && array[first])
1675 r = array[first]->error;
1679 err_free_fence_array:
1680 for (i = 0; i < fence_count; i++)
1681 dma_fence_put(array[i]);
1688 * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1691 * @data: data from userspace
1692 * @filp: file private
1694 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1695 struct drm_file *filp)
1697 struct amdgpu_device *adev = drm_to_adev(dev);
1698 union drm_amdgpu_wait_fences *wait = data;
1699 uint32_t fence_count = wait->in.fence_count;
1700 struct drm_amdgpu_fence *fences_user;
1701 struct drm_amdgpu_fence *fences;
1704 /* Get the fences from userspace */
1705 fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1710 fences_user = u64_to_user_ptr(wait->in.fences);
1711 if (copy_from_user(fences, fences_user,
1712 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1714 goto err_free_fences;
1717 if (wait->in.wait_all)
1718 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1720 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1729 * amdgpu_cs_find_mapping - find bo_va for VM address
1731 * @parser: command submission parser context
1733 * @bo: resulting BO of the mapping found
1734 * @map: Placeholder to return found BO mapping
1736 * Search the buffer objects in the command submission context for a certain
1737 * virtual memory address. Returns allocation structure when found, NULL
1740 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1741 uint64_t addr, struct amdgpu_bo **bo,
1742 struct amdgpu_bo_va_mapping **map)
1744 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1745 struct ttm_operation_ctx ctx = { false, false };
1746 struct amdgpu_vm *vm = &fpriv->vm;
1747 struct amdgpu_bo_va_mapping *mapping;
1750 addr /= AMDGPU_GPU_PAGE_SIZE;
1752 mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1753 if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1756 *bo = mapping->bo_va->base.bo;
1759 /* Double check that the BO is reserved by this CS */
1760 if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1763 if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1764 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1765 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1766 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1771 return amdgpu_ttm_alloc_gart(&(*bo)->tbo);