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");
87 if (!(*out_ring && (*out_ring)->adev)) {
88 DRM_ERROR("Ring %d is not initialized on IP %d\n",
96 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
97 struct drm_amdgpu_cs_chunk_fence *data,
100 struct drm_gem_object *gobj;
103 gobj = drm_gem_object_lookup(p->filp, data->handle);
107 p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
108 p->uf_entry.priority = 0;
109 p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
110 p->uf_entry.tv.shared = true;
111 p->uf_entry.user_pages = NULL;
113 size = amdgpu_bo_size(p->uf_entry.robj);
114 if (size != PAGE_SIZE || (data->offset + 8) > size)
117 *offset = data->offset;
119 drm_gem_object_unreference_unlocked(gobj);
121 if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
122 amdgpu_bo_unref(&p->uf_entry.robj);
129 int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
131 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
132 struct amdgpu_vm *vm = &fpriv->vm;
133 union drm_amdgpu_cs *cs = data;
134 uint64_t *chunk_array_user;
135 uint64_t *chunk_array;
136 unsigned size, num_ibs = 0;
137 uint32_t uf_offset = 0;
141 if (cs->in.num_chunks == 0)
144 chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
148 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
155 chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
156 if (copy_from_user(chunk_array, chunk_array_user,
157 sizeof(uint64_t)*cs->in.num_chunks)) {
162 p->nchunks = cs->in.num_chunks;
163 p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
170 for (i = 0; i < p->nchunks; i++) {
171 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
172 struct drm_amdgpu_cs_chunk user_chunk;
173 uint32_t __user *cdata;
175 chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
176 if (copy_from_user(&user_chunk, chunk_ptr,
177 sizeof(struct drm_amdgpu_cs_chunk))) {
180 goto free_partial_kdata;
182 p->chunks[i].chunk_id = user_chunk.chunk_id;
183 p->chunks[i].length_dw = user_chunk.length_dw;
185 size = p->chunks[i].length_dw;
186 cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
188 p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
189 if (p->chunks[i].kdata == NULL) {
192 goto free_partial_kdata;
194 size *= sizeof(uint32_t);
195 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
197 goto free_partial_kdata;
200 switch (p->chunks[i].chunk_id) {
201 case AMDGPU_CHUNK_ID_IB:
205 case AMDGPU_CHUNK_ID_FENCE:
206 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
207 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
209 goto free_partial_kdata;
212 ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
215 goto free_partial_kdata;
219 case AMDGPU_CHUNK_ID_DEPENDENCIES:
224 goto free_partial_kdata;
228 ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
232 if (p->uf_entry.robj)
233 p->job->uf_addr = uf_offset;
241 drm_free_large(p->chunks[i].kdata);
246 amdgpu_ctx_put(p->ctx);
253 /* Convert microseconds to bytes. */
254 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
256 if (us <= 0 || !adev->mm_stats.log2_max_MBps)
259 /* Since accum_us is incremented by a million per second, just
260 * multiply it by the number of MB/s to get the number of bytes.
262 return us << adev->mm_stats.log2_max_MBps;
265 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
267 if (!adev->mm_stats.log2_max_MBps)
270 return bytes >> adev->mm_stats.log2_max_MBps;
273 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
274 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
275 * which means it can go over the threshold once. If that happens, the driver
276 * will be in debt and no other buffer migrations can be done until that debt
279 * This approach allows moving a buffer of any size (it's important to allow
282 * The currency is simply time in microseconds and it increases as the clock
283 * ticks. The accumulated microseconds (us) are converted to bytes and
286 static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
288 s64 time_us, increment_us;
290 u64 free_vram, total_vram, used_vram;
292 /* Allow a maximum of 200 accumulated ms. This is basically per-IB
295 * It means that in order to get full max MBps, at least 5 IBs per
296 * second must be submitted and not more than 200ms apart from each
299 const s64 us_upper_bound = 200000;
301 if (!adev->mm_stats.log2_max_MBps)
304 total_vram = adev->mc.real_vram_size - adev->vram_pin_size;
305 used_vram = atomic64_read(&adev->vram_usage);
306 free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
308 spin_lock(&adev->mm_stats.lock);
310 /* Increase the amount of accumulated us. */
311 time_us = ktime_to_us(ktime_get());
312 increment_us = time_us - adev->mm_stats.last_update_us;
313 adev->mm_stats.last_update_us = time_us;
314 adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
317 /* This prevents the short period of low performance when the VRAM
318 * usage is low and the driver is in debt or doesn't have enough
319 * accumulated us to fill VRAM quickly.
321 * The situation can occur in these cases:
322 * - a lot of VRAM is freed by userspace
323 * - the presence of a big buffer causes a lot of evictions
324 * (solution: split buffers into smaller ones)
326 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
327 * accum_us to a positive number.
329 if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
332 /* Be more aggresive on dGPUs. Try to fill a portion of free
335 if (!(adev->flags & AMD_IS_APU))
336 min_us = bytes_to_us(adev, free_vram / 4);
338 min_us = 0; /* Reset accum_us on APUs. */
340 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
343 /* This returns 0 if the driver is in debt to disallow (optional)
346 max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
348 spin_unlock(&adev->mm_stats.lock);
352 /* Report how many bytes have really been moved for the last command
353 * submission. This can result in a debt that can stop buffer migrations
356 static void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev,
359 spin_lock(&adev->mm_stats.lock);
360 adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
361 spin_unlock(&adev->mm_stats.lock);
364 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
365 struct amdgpu_bo *bo)
367 u64 initial_bytes_moved;
374 /* Don't move this buffer if we have depleted our allowance
375 * to move it. Don't move anything if the threshold is zero.
377 if (p->bytes_moved < p->bytes_moved_threshold)
378 domain = bo->prefered_domains;
380 domain = bo->allowed_domains;
383 amdgpu_ttm_placement_from_domain(bo, domain);
384 initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
385 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
386 p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
389 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
390 domain = bo->allowed_domains;
397 /* Last resort, try to evict something from the current working set */
398 static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
399 struct amdgpu_bo_list_entry *lobj)
401 uint32_t domain = lobj->robj->allowed_domains;
407 for (;&p->evictable->tv.head != &p->validated;
408 p->evictable = list_prev_entry(p->evictable, tv.head)) {
410 struct amdgpu_bo_list_entry *candidate = p->evictable;
411 struct amdgpu_bo *bo = candidate->robj;
412 u64 initial_bytes_moved;
415 /* If we reached our current BO we can forget it */
416 if (candidate == lobj)
419 /* We can't move pinned BOs here */
423 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
425 /* Check if this BO is in one of the domains we need space for */
426 if (!(other & domain))
429 /* Check if we can move this BO somewhere else */
430 other = bo->allowed_domains & ~domain;
434 /* Good we can try to move this BO somewhere else */
435 amdgpu_ttm_placement_from_domain(bo, other);
436 initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
437 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
438 p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
444 p->evictable = list_prev_entry(p->evictable, tv.head);
445 list_move(&candidate->tv.head, &p->validated);
453 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
454 struct list_head *validated)
456 struct amdgpu_bo_list_entry *lobj;
459 list_for_each_entry(lobj, validated, tv.head) {
460 struct amdgpu_bo *bo = lobj->robj;
461 bool binding_userptr = false;
462 struct mm_struct *usermm;
464 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
465 if (usermm && usermm != current->mm)
468 /* Check if we have user pages and nobody bound the BO already */
469 if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
470 size_t size = sizeof(struct page *);
472 size *= bo->tbo.ttm->num_pages;
473 memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
474 binding_userptr = true;
477 if (p->evictable == lobj)
481 r = amdgpu_cs_bo_validate(p, bo);
482 } while (r == -ENOMEM && amdgpu_cs_try_evict(p, lobj));
487 r = amdgpu_cs_bo_validate(p, bo);
492 if (binding_userptr) {
493 drm_free_large(lobj->user_pages);
494 lobj->user_pages = NULL;
500 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
501 union drm_amdgpu_cs *cs)
503 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
504 struct amdgpu_bo_list_entry *e;
505 struct list_head duplicates;
506 bool need_mmap_lock = false;
507 unsigned i, tries = 10;
510 INIT_LIST_HEAD(&p->validated);
512 p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
514 need_mmap_lock = p->bo_list->first_userptr !=
515 p->bo_list->num_entries;
516 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
519 INIT_LIST_HEAD(&duplicates);
520 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
522 if (p->uf_entry.robj && !p->uf_entry.robj->parent)
523 list_add(&p->uf_entry.tv.head, &p->validated);
526 down_read(¤t->mm->mmap_sem);
529 struct list_head need_pages;
532 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
534 if (unlikely(r != 0)) {
535 if (r != -ERESTARTSYS)
536 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
537 goto error_free_pages;
540 /* Without a BO list we don't have userptr BOs */
544 INIT_LIST_HEAD(&need_pages);
545 for (i = p->bo_list->first_userptr;
546 i < p->bo_list->num_entries; ++i) {
548 e = &p->bo_list->array[i];
550 if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
551 &e->user_invalidated) && e->user_pages) {
553 /* We acquired a page array, but somebody
554 * invalidated it. Free it an try again
556 release_pages(e->user_pages,
557 e->robj->tbo.ttm->num_pages,
559 drm_free_large(e->user_pages);
560 e->user_pages = NULL;
563 if (e->robj->tbo.ttm->state != tt_bound &&
565 list_del(&e->tv.head);
566 list_add(&e->tv.head, &need_pages);
568 amdgpu_bo_unreserve(e->robj);
572 if (list_empty(&need_pages))
575 /* Unreserve everything again. */
576 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
578 /* We tried too many times, just abort */
581 DRM_ERROR("deadlock in %s\n", __func__);
582 goto error_free_pages;
585 /* Fill the page arrays for all useptrs. */
586 list_for_each_entry(e, &need_pages, tv.head) {
587 struct ttm_tt *ttm = e->robj->tbo.ttm;
589 e->user_pages = drm_calloc_large(ttm->num_pages,
590 sizeof(struct page*));
591 if (!e->user_pages) {
593 DRM_ERROR("calloc failure in %s\n", __func__);
594 goto error_free_pages;
597 r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
599 DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
600 drm_free_large(e->user_pages);
601 e->user_pages = NULL;
602 goto error_free_pages;
607 list_splice(&need_pages, &p->validated);
610 amdgpu_vm_get_pt_bos(p->adev, &fpriv->vm, &duplicates);
612 p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
614 p->evictable = list_last_entry(&p->validated,
615 struct amdgpu_bo_list_entry,
618 r = amdgpu_cs_list_validate(p, &duplicates);
620 DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
624 r = amdgpu_cs_list_validate(p, &p->validated);
626 DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
630 amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved);
632 fpriv->vm.last_eviction_counter =
633 atomic64_read(&p->adev->num_evictions);
636 struct amdgpu_bo *gds = p->bo_list->gds_obj;
637 struct amdgpu_bo *gws = p->bo_list->gws_obj;
638 struct amdgpu_bo *oa = p->bo_list->oa_obj;
639 struct amdgpu_vm *vm = &fpriv->vm;
642 for (i = 0; i < p->bo_list->num_entries; i++) {
643 struct amdgpu_bo *bo = p->bo_list->array[i].robj;
645 p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
649 p->job->gds_base = amdgpu_bo_gpu_offset(gds);
650 p->job->gds_size = amdgpu_bo_size(gds);
653 p->job->gws_base = amdgpu_bo_gpu_offset(gws);
654 p->job->gws_size = amdgpu_bo_size(gws);
657 p->job->oa_base = amdgpu_bo_gpu_offset(oa);
658 p->job->oa_size = amdgpu_bo_size(oa);
662 if (!r && p->uf_entry.robj) {
663 struct amdgpu_bo *uf = p->uf_entry.robj;
665 r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
666 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
671 amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
672 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
678 up_read(¤t->mm->mmap_sem);
681 for (i = p->bo_list->first_userptr;
682 i < p->bo_list->num_entries; ++i) {
683 e = &p->bo_list->array[i];
688 release_pages(e->user_pages,
689 e->robj->tbo.ttm->num_pages,
691 drm_free_large(e->user_pages);
698 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
700 struct amdgpu_bo_list_entry *e;
703 list_for_each_entry(e, &p->validated, tv.head) {
704 struct reservation_object *resv = e->robj->tbo.resv;
705 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
714 * cs_parser_fini() - clean parser states
715 * @parser: parser structure holding parsing context.
716 * @error: error number
718 * If error is set than unvalidate buffer, otherwise just free memory
719 * used by parsing context.
721 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
723 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
727 amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
729 ttm_eu_fence_buffer_objects(&parser->ticket,
732 } else if (backoff) {
733 ttm_eu_backoff_reservation(&parser->ticket,
736 fence_put(parser->fence);
739 amdgpu_ctx_put(parser->ctx);
741 amdgpu_bo_list_put(parser->bo_list);
743 for (i = 0; i < parser->nchunks; i++)
744 drm_free_large(parser->chunks[i].kdata);
745 kfree(parser->chunks);
747 amdgpu_job_free(parser->job);
748 amdgpu_bo_unref(&parser->uf_entry.robj);
751 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
752 struct amdgpu_vm *vm)
754 struct amdgpu_device *adev = p->adev;
755 struct amdgpu_bo_va *bo_va;
756 struct amdgpu_bo *bo;
759 r = amdgpu_vm_update_page_directory(adev, vm);
763 r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
767 r = amdgpu_vm_clear_freed(adev, vm);
772 for (i = 0; i < p->bo_list->num_entries; i++) {
775 /* ignore duplicates */
776 bo = p->bo_list->array[i].robj;
780 bo_va = p->bo_list->array[i].bo_va;
784 r = amdgpu_vm_bo_update(adev, bo_va, false);
788 f = bo_va->last_pt_update;
789 r = amdgpu_sync_fence(adev, &p->job->sync, f);
796 r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
798 if (amdgpu_vm_debug && p->bo_list) {
799 /* Invalidate all BOs to test for userspace bugs */
800 for (i = 0; i < p->bo_list->num_entries; i++) {
801 /* ignore duplicates */
802 bo = p->bo_list->array[i].robj;
806 amdgpu_vm_bo_invalidate(adev, bo);
813 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
814 struct amdgpu_cs_parser *p)
816 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
817 struct amdgpu_vm *vm = &fpriv->vm;
818 struct amdgpu_ring *ring = p->job->ring;
821 /* Only for UVD/VCE VM emulation */
822 if (ring->funcs->parse_cs) {
824 for (i = 0; i < p->job->num_ibs; i++) {
825 r = amdgpu_ring_parse_cs(ring, p, i);
830 p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
832 r = amdgpu_bo_vm_update_pte(p, vm);
837 return amdgpu_cs_sync_rings(p);
840 static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
843 r = amdgpu_gpu_reset(adev);
850 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
851 struct amdgpu_cs_parser *parser)
853 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
854 struct amdgpu_vm *vm = &fpriv->vm;
858 for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
859 struct amdgpu_cs_chunk *chunk;
860 struct amdgpu_ib *ib;
861 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
862 struct amdgpu_ring *ring;
864 chunk = &parser->chunks[i];
865 ib = &parser->job->ibs[j];
866 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
868 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
871 r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
872 chunk_ib->ip_instance, chunk_ib->ring,
877 if (ib->flags & AMDGPU_IB_FLAG_PREAMBLE) {
878 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
879 if (!parser->ctx->preamble_presented) {
880 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
881 parser->ctx->preamble_presented = true;
885 if (parser->job->ring && parser->job->ring != ring)
888 parser->job->ring = ring;
890 if (ring->funcs->parse_cs) {
891 struct amdgpu_bo_va_mapping *m;
892 struct amdgpu_bo *aobj = NULL;
896 m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
899 DRM_ERROR("IB va_start is invalid\n");
903 if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
904 (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
905 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
909 /* the IB should be reserved at this point */
910 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
915 offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
916 kptr += chunk_ib->va_start - offset;
918 r = amdgpu_ib_get(adev, NULL, chunk_ib->ib_bytes, ib);
920 DRM_ERROR("Failed to get ib !\n");
924 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
925 amdgpu_bo_kunmap(aobj);
927 r = amdgpu_ib_get(adev, vm, 0, ib);
929 DRM_ERROR("Failed to get ib !\n");
933 ib->gpu_addr = chunk_ib->va_start;
936 ib->length_dw = chunk_ib->ib_bytes / 4;
937 ib->flags = chunk_ib->flags;
941 /* UVD & VCE fw doesn't support user fences */
942 if (parser->job->uf_addr && (
943 parser->job->ring->type == AMDGPU_RING_TYPE_UVD ||
944 parser->job->ring->type == AMDGPU_RING_TYPE_VCE))
950 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
951 struct amdgpu_cs_parser *p)
953 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
956 for (i = 0; i < p->nchunks; ++i) {
957 struct drm_amdgpu_cs_chunk_dep *deps;
958 struct amdgpu_cs_chunk *chunk;
961 chunk = &p->chunks[i];
963 if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
966 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
967 num_deps = chunk->length_dw * 4 /
968 sizeof(struct drm_amdgpu_cs_chunk_dep);
970 for (j = 0; j < num_deps; ++j) {
971 struct amdgpu_ring *ring;
972 struct amdgpu_ctx *ctx;
975 r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
977 deps[j].ring, &ring);
981 ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
985 fence = amdgpu_ctx_get_fence(ctx, ring,
993 r = amdgpu_sync_fence(adev, &p->job->sync,
1006 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1007 union drm_amdgpu_cs *cs)
1009 struct amdgpu_ring *ring = p->job->ring;
1010 struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1011 struct amdgpu_job *job;
1017 r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1019 amdgpu_job_free(job);
1023 job->owner = p->filp;
1024 job->fence_ctx = entity->fence_context;
1025 p->fence = fence_get(&job->base.s_fence->finished);
1026 cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, p->fence);
1027 job->uf_sequence = cs->out.handle;
1028 amdgpu_job_free_resources(job);
1030 trace_amdgpu_cs_ioctl(job);
1031 amd_sched_entity_push_job(&job->base);
1036 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1038 struct amdgpu_device *adev = dev->dev_private;
1039 union drm_amdgpu_cs *cs = data;
1040 struct amdgpu_cs_parser parser = {};
1041 bool reserved_buffers = false;
1044 if (!adev->accel_working)
1050 r = amdgpu_cs_parser_init(&parser, data);
1052 DRM_ERROR("Failed to initialize parser !\n");
1053 amdgpu_cs_parser_fini(&parser, r, false);
1054 r = amdgpu_cs_handle_lockup(adev, r);
1057 r = amdgpu_cs_parser_bos(&parser, data);
1059 DRM_ERROR("Not enough memory for command submission!\n");
1060 else if (r && r != -ERESTARTSYS)
1061 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1063 reserved_buffers = true;
1064 r = amdgpu_cs_ib_fill(adev, &parser);
1068 r = amdgpu_cs_dependencies(adev, &parser);
1070 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1076 for (i = 0; i < parser.job->num_ibs; i++)
1077 trace_amdgpu_cs(&parser, i);
1079 r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1083 r = amdgpu_cs_submit(&parser, cs);
1086 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1087 r = amdgpu_cs_handle_lockup(adev, r);
1092 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1095 * @data: data from userspace
1096 * @filp: file private
1098 * Wait for the command submission identified by handle to finish.
1100 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1101 struct drm_file *filp)
1103 union drm_amdgpu_wait_cs *wait = data;
1104 struct amdgpu_device *adev = dev->dev_private;
1105 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1106 struct amdgpu_ring *ring = NULL;
1107 struct amdgpu_ctx *ctx;
1108 struct fence *fence;
1111 r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
1112 wait->in.ring, &ring);
1116 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1120 fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
1124 r = fence_wait_timeout(fence, true, timeout);
1129 amdgpu_ctx_put(ctx);
1133 memset(wait, 0, sizeof(*wait));
1134 wait->out.status = (r == 0);
1140 * amdgpu_cs_find_bo_va - find bo_va for VM address
1142 * @parser: command submission parser context
1144 * @bo: resulting BO of the mapping found
1146 * Search the buffer objects in the command submission context for a certain
1147 * virtual memory address. Returns allocation structure when found, NULL
1150 struct amdgpu_bo_va_mapping *
1151 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1152 uint64_t addr, struct amdgpu_bo **bo)
1154 struct amdgpu_bo_va_mapping *mapping;
1157 if (!parser->bo_list)
1160 addr /= AMDGPU_GPU_PAGE_SIZE;
1162 for (i = 0; i < parser->bo_list->num_entries; i++) {
1163 struct amdgpu_bo_list_entry *lobj;
1165 lobj = &parser->bo_list->array[i];
1169 list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
1170 if (mapping->it.start > addr ||
1171 addr > mapping->it.last)
1174 *bo = lobj->bo_va->bo;
1178 list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
1179 if (mapping->it.start > addr ||
1180 addr > mapping->it.last)
1183 *bo = lobj->bo_va->bo;
1192 * amdgpu_cs_sysvm_access_required - make BOs accessible by the system VM
1194 * @parser: command submission parser context
1196 * Helper for UVD/VCE VM emulation, make sure BOs are accessible by the system VM.
1198 int amdgpu_cs_sysvm_access_required(struct amdgpu_cs_parser *parser)
1203 if (!parser->bo_list)
1206 for (i = 0; i < parser->bo_list->num_entries; i++) {
1207 struct amdgpu_bo *bo = parser->bo_list->array[i].robj;
1209 r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);