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.
27 #include <linux/pagemap.h>
28 #include <linux/sync_file.h>
30 #include <drm/amdgpu_drm.h>
31 #include <drm/drm_syncobj.h>
33 #include "amdgpu_trace.h"
35 static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
36 struct drm_amdgpu_cs_chunk_fence *data,
39 struct drm_gem_object *gobj;
42 gobj = drm_gem_object_lookup(p->filp, data->handle);
46 p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
47 p->uf_entry.priority = 0;
48 p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
49 p->uf_entry.tv.shared = true;
50 p->uf_entry.user_pages = NULL;
52 size = amdgpu_bo_size(p->uf_entry.robj);
53 if (size != PAGE_SIZE || (data->offset + 8) > size)
56 *offset = data->offset;
58 drm_gem_object_put_unlocked(gobj);
60 if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
61 amdgpu_bo_unref(&p->uf_entry.robj);
68 static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
70 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
71 struct amdgpu_vm *vm = &fpriv->vm;
72 union drm_amdgpu_cs *cs = data;
73 uint64_t *chunk_array_user;
74 uint64_t *chunk_array;
75 unsigned size, num_ibs = 0;
76 uint32_t uf_offset = 0;
80 if (cs->in.num_chunks == 0)
83 chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
87 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
93 /* skip guilty context job */
94 if (atomic_read(&p->ctx->guilty) == 1) {
99 mutex_lock(&p->ctx->lock);
102 chunk_array_user = u64_to_user_ptr(cs->in.chunks);
103 if (copy_from_user(chunk_array, chunk_array_user,
104 sizeof(uint64_t)*cs->in.num_chunks)) {
109 p->nchunks = cs->in.num_chunks;
110 p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
117 for (i = 0; i < p->nchunks; i++) {
118 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
119 struct drm_amdgpu_cs_chunk user_chunk;
120 uint32_t __user *cdata;
122 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
123 if (copy_from_user(&user_chunk, chunk_ptr,
124 sizeof(struct drm_amdgpu_cs_chunk))) {
127 goto free_partial_kdata;
129 p->chunks[i].chunk_id = user_chunk.chunk_id;
130 p->chunks[i].length_dw = user_chunk.length_dw;
132 size = p->chunks[i].length_dw;
133 cdata = u64_to_user_ptr(user_chunk.chunk_data);
135 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
136 if (p->chunks[i].kdata == NULL) {
139 goto free_partial_kdata;
141 size *= sizeof(uint32_t);
142 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
144 goto free_partial_kdata;
147 switch (p->chunks[i].chunk_id) {
148 case AMDGPU_CHUNK_ID_IB:
152 case AMDGPU_CHUNK_ID_FENCE:
153 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
154 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
156 goto free_partial_kdata;
159 ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
162 goto free_partial_kdata;
166 case AMDGPU_CHUNK_ID_DEPENDENCIES:
167 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
168 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
173 goto free_partial_kdata;
177 ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
181 if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
186 if (p->uf_entry.robj)
187 p->job->uf_addr = uf_offset;
195 kvfree(p->chunks[i].kdata);
205 /* Convert microseconds to bytes. */
206 static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
208 if (us <= 0 || !adev->mm_stats.log2_max_MBps)
211 /* Since accum_us is incremented by a million per second, just
212 * multiply it by the number of MB/s to get the number of bytes.
214 return us << adev->mm_stats.log2_max_MBps;
217 static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
219 if (!adev->mm_stats.log2_max_MBps)
222 return bytes >> adev->mm_stats.log2_max_MBps;
225 /* Returns how many bytes TTM can move right now. If no bytes can be moved,
226 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
227 * which means it can go over the threshold once. If that happens, the driver
228 * will be in debt and no other buffer migrations can be done until that debt
231 * This approach allows moving a buffer of any size (it's important to allow
234 * The currency is simply time in microseconds and it increases as the clock
235 * ticks. The accumulated microseconds (us) are converted to bytes and
238 static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
242 s64 time_us, increment_us;
243 u64 free_vram, total_vram, used_vram;
245 /* Allow a maximum of 200 accumulated ms. This is basically per-IB
248 * It means that in order to get full max MBps, at least 5 IBs per
249 * second must be submitted and not more than 200ms apart from each
252 const s64 us_upper_bound = 200000;
254 if (!adev->mm_stats.log2_max_MBps) {
260 total_vram = adev->gmc.real_vram_size - adev->vram_pin_size;
261 used_vram = amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
262 free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
264 spin_lock(&adev->mm_stats.lock);
266 /* Increase the amount of accumulated us. */
267 time_us = ktime_to_us(ktime_get());
268 increment_us = time_us - adev->mm_stats.last_update_us;
269 adev->mm_stats.last_update_us = time_us;
270 adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
273 /* This prevents the short period of low performance when the VRAM
274 * usage is low and the driver is in debt or doesn't have enough
275 * accumulated us to fill VRAM quickly.
277 * The situation can occur in these cases:
278 * - a lot of VRAM is freed by userspace
279 * - the presence of a big buffer causes a lot of evictions
280 * (solution: split buffers into smaller ones)
282 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
283 * accum_us to a positive number.
285 if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
288 /* Be more aggresive on dGPUs. Try to fill a portion of free
291 if (!(adev->flags & AMD_IS_APU))
292 min_us = bytes_to_us(adev, free_vram / 4);
294 min_us = 0; /* Reset accum_us on APUs. */
296 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
299 /* This is set to 0 if the driver is in debt to disallow (optional)
302 *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
304 /* Do the same for visible VRAM if half of it is free */
305 if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size) {
306 u64 total_vis_vram = adev->gmc.visible_vram_size;
308 amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
310 if (used_vis_vram < total_vis_vram) {
311 u64 free_vis_vram = total_vis_vram - used_vis_vram;
312 adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
313 increment_us, us_upper_bound);
315 if (free_vis_vram >= total_vis_vram / 2)
316 adev->mm_stats.accum_us_vis =
317 max(bytes_to_us(adev, free_vis_vram / 2),
318 adev->mm_stats.accum_us_vis);
321 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
326 spin_unlock(&adev->mm_stats.lock);
329 /* Report how many bytes have really been moved for the last command
330 * submission. This can result in a debt that can stop buffer migrations
333 void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
336 spin_lock(&adev->mm_stats.lock);
337 adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
338 adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
339 spin_unlock(&adev->mm_stats.lock);
342 static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
343 struct amdgpu_bo *bo)
345 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
346 struct ttm_operation_ctx ctx = {
347 .interruptible = true,
348 .no_wait_gpu = false,
349 .resv = bo->tbo.resv,
358 /* Don't move this buffer if we have depleted our allowance
359 * to move it. Don't move anything if the threshold is zero.
361 if (p->bytes_moved < p->bytes_moved_threshold) {
362 if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
363 (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
364 /* And don't move a CPU_ACCESS_REQUIRED BO to limited
365 * visible VRAM if we've depleted our allowance to do
368 if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
369 domain = bo->preferred_domains;
371 domain = bo->allowed_domains;
373 domain = bo->preferred_domains;
376 domain = bo->allowed_domains;
380 amdgpu_ttm_placement_from_domain(bo, domain);
381 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
383 p->bytes_moved += ctx.bytes_moved;
384 if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
385 amdgpu_bo_in_cpu_visible_vram(bo))
386 p->bytes_moved_vis += ctx.bytes_moved;
388 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains &&
389 !(bo->flags & AMDGPU_GEM_CREATE_NO_FALLBACK)) {
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 *validated)
401 uint32_t domain = validated->allowed_domains;
402 struct ttm_operation_ctx ctx = { true, false };
408 for (;&p->evictable->tv.head != &p->validated;
409 p->evictable = list_prev_entry(p->evictable, tv.head)) {
411 struct amdgpu_bo_list_entry *candidate = p->evictable;
412 struct amdgpu_bo *bo = candidate->robj;
413 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
414 bool update_bytes_moved_vis;
417 /* If we reached our current BO we can forget it */
418 if (candidate->robj == validated)
421 /* We can't move pinned BOs here */
425 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
427 /* Check if this BO is in one of the domains we need space for */
428 if (!(other & domain))
431 /* Check if we can move this BO somewhere else */
432 other = bo->allowed_domains & ~domain;
436 /* Good we can try to move this BO somewhere else */
437 update_bytes_moved_vis =
438 adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
439 amdgpu_bo_in_cpu_visible_vram(bo);
440 amdgpu_ttm_placement_from_domain(bo, other);
441 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
442 p->bytes_moved += ctx.bytes_moved;
443 if (update_bytes_moved_vis)
444 p->bytes_moved_vis += ctx.bytes_moved;
449 p->evictable = list_prev_entry(p->evictable, tv.head);
450 list_move(&candidate->tv.head, &p->validated);
458 static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
460 struct amdgpu_cs_parser *p = param;
464 r = amdgpu_cs_bo_validate(p, bo);
465 } while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo));
470 r = amdgpu_cs_bo_validate(p, bo->shadow);
475 static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
476 struct list_head *validated)
478 struct ttm_operation_ctx ctx = { true, false };
479 struct amdgpu_bo_list_entry *lobj;
482 list_for_each_entry(lobj, validated, tv.head) {
483 struct amdgpu_bo *bo = lobj->robj;
484 bool binding_userptr = false;
485 struct mm_struct *usermm;
487 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
488 if (usermm && usermm != current->mm)
491 /* Check if we have user pages and nobody bound the BO already */
492 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm) &&
494 amdgpu_ttm_placement_from_domain(bo,
495 AMDGPU_GEM_DOMAIN_CPU);
496 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
499 amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
501 binding_userptr = true;
504 if (p->evictable == lobj)
507 r = amdgpu_cs_validate(p, bo);
511 if (binding_userptr) {
512 kvfree(lobj->user_pages);
513 lobj->user_pages = NULL;
519 static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
520 union drm_amdgpu_cs *cs)
522 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
523 struct amdgpu_bo_list_entry *e;
524 struct list_head duplicates;
525 unsigned i, tries = 10;
528 INIT_LIST_HEAD(&p->validated);
530 p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
532 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
533 if (p->bo_list->first_userptr != p->bo_list->num_entries)
534 p->mn = amdgpu_mn_get(p->adev);
537 INIT_LIST_HEAD(&duplicates);
538 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
540 if (p->uf_entry.robj && !p->uf_entry.robj->parent)
541 list_add(&p->uf_entry.tv.head, &p->validated);
544 struct list_head need_pages;
547 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
549 if (unlikely(r != 0)) {
550 if (r != -ERESTARTSYS)
551 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
552 goto error_free_pages;
555 /* Without a BO list we don't have userptr BOs */
559 INIT_LIST_HEAD(&need_pages);
560 for (i = p->bo_list->first_userptr;
561 i < p->bo_list->num_entries; ++i) {
562 struct amdgpu_bo *bo;
564 e = &p->bo_list->array[i];
567 if (amdgpu_ttm_tt_userptr_invalidated(bo->tbo.ttm,
568 &e->user_invalidated) && e->user_pages) {
570 /* We acquired a page array, but somebody
571 * invalidated it. Free it and try again
573 release_pages(e->user_pages,
574 bo->tbo.ttm->num_pages);
575 kvfree(e->user_pages);
576 e->user_pages = NULL;
579 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm) &&
581 list_del(&e->tv.head);
582 list_add(&e->tv.head, &need_pages);
584 amdgpu_bo_unreserve(e->robj);
588 if (list_empty(&need_pages))
591 /* Unreserve everything again. */
592 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
594 /* We tried too many times, just abort */
597 DRM_ERROR("deadlock in %s\n", __func__);
598 goto error_free_pages;
601 /* Fill the page arrays for all userptrs. */
602 list_for_each_entry(e, &need_pages, tv.head) {
603 struct ttm_tt *ttm = e->robj->tbo.ttm;
605 e->user_pages = kvmalloc_array(ttm->num_pages,
606 sizeof(struct page*),
607 GFP_KERNEL | __GFP_ZERO);
608 if (!e->user_pages) {
610 DRM_ERROR("calloc failure in %s\n", __func__);
611 goto error_free_pages;
614 r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
616 DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
617 kvfree(e->user_pages);
618 e->user_pages = NULL;
619 goto error_free_pages;
624 list_splice(&need_pages, &p->validated);
627 amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
628 &p->bytes_moved_vis_threshold);
630 p->bytes_moved_vis = 0;
631 p->evictable = list_last_entry(&p->validated,
632 struct amdgpu_bo_list_entry,
635 r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
636 amdgpu_cs_validate, p);
638 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
642 r = amdgpu_cs_list_validate(p, &duplicates);
644 DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
648 r = amdgpu_cs_list_validate(p, &p->validated);
650 DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
654 amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
657 struct amdgpu_bo *gds = p->bo_list->gds_obj;
658 struct amdgpu_bo *gws = p->bo_list->gws_obj;
659 struct amdgpu_bo *oa = p->bo_list->oa_obj;
660 struct amdgpu_vm *vm = &fpriv->vm;
663 for (i = 0; i < p->bo_list->num_entries; i++) {
664 struct amdgpu_bo *bo = p->bo_list->array[i].robj;
666 p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
670 p->job->gds_base = amdgpu_bo_gpu_offset(gds);
671 p->job->gds_size = amdgpu_bo_size(gds);
674 p->job->gws_base = amdgpu_bo_gpu_offset(gws);
675 p->job->gws_size = amdgpu_bo_size(gws);
678 p->job->oa_base = amdgpu_bo_gpu_offset(oa);
679 p->job->oa_size = amdgpu_bo_size(oa);
683 if (!r && p->uf_entry.robj) {
684 struct amdgpu_bo *uf = p->uf_entry.robj;
686 r = amdgpu_ttm_alloc_gart(&uf->tbo);
687 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
692 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
697 for (i = p->bo_list->first_userptr;
698 i < p->bo_list->num_entries; ++i) {
699 e = &p->bo_list->array[i];
704 release_pages(e->user_pages,
705 e->robj->tbo.ttm->num_pages);
706 kvfree(e->user_pages);
713 static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
715 struct amdgpu_bo_list_entry *e;
718 list_for_each_entry(e, &p->validated, tv.head) {
719 struct reservation_object *resv = e->robj->tbo.resv;
720 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp,
721 amdgpu_bo_explicit_sync(e->robj));
730 * cs_parser_fini() - clean parser states
731 * @parser: parser structure holding parsing context.
732 * @error: error number
734 * If error is set than unvalidate buffer, otherwise just free memory
735 * used by parsing context.
737 static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
742 if (error && backoff)
743 ttm_eu_backoff_reservation(&parser->ticket,
746 for (i = 0; i < parser->num_post_dep_syncobjs; i++)
747 drm_syncobj_put(parser->post_dep_syncobjs[i]);
748 kfree(parser->post_dep_syncobjs);
750 dma_fence_put(parser->fence);
753 mutex_unlock(&parser->ctx->lock);
754 amdgpu_ctx_put(parser->ctx);
757 amdgpu_bo_list_put(parser->bo_list);
759 for (i = 0; i < parser->nchunks; i++)
760 kvfree(parser->chunks[i].kdata);
761 kfree(parser->chunks);
763 amdgpu_job_free(parser->job);
764 amdgpu_bo_unref(&parser->uf_entry.robj);
767 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p)
769 struct amdgpu_device *adev = p->adev;
770 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
771 struct amdgpu_vm *vm = &fpriv->vm;
772 struct amdgpu_bo_va *bo_va;
773 struct amdgpu_bo *bo;
776 r = amdgpu_vm_clear_freed(adev, vm, NULL);
780 r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
784 r = amdgpu_sync_fence(adev, &p->job->sync,
785 fpriv->prt_va->last_pt_update, false);
789 if (amdgpu_sriov_vf(adev)) {
792 bo_va = fpriv->csa_va;
794 r = amdgpu_vm_bo_update(adev, bo_va, false);
798 f = bo_va->last_pt_update;
799 r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
805 for (i = 0; i < p->bo_list->num_entries; i++) {
808 /* ignore duplicates */
809 bo = p->bo_list->array[i].robj;
813 bo_va = p->bo_list->array[i].bo_va;
817 r = amdgpu_vm_bo_update(adev, bo_va, false);
821 f = bo_va->last_pt_update;
822 r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
829 r = amdgpu_vm_handle_moved(adev, vm);
833 r = amdgpu_vm_update_directories(adev, vm);
837 r = amdgpu_sync_fence(adev, &p->job->sync, vm->last_update, false);
841 if (amdgpu_vm_debug && p->bo_list) {
842 /* Invalidate all BOs to test for userspace bugs */
843 for (i = 0; i < p->bo_list->num_entries; i++) {
844 /* ignore duplicates */
845 bo = p->bo_list->array[i].robj;
849 amdgpu_vm_bo_invalidate(adev, bo, false);
856 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
857 struct amdgpu_cs_parser *p)
859 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
860 struct amdgpu_vm *vm = &fpriv->vm;
861 struct amdgpu_ring *ring = p->job->ring;
864 /* Only for UVD/VCE VM emulation */
865 if (p->job->ring->funcs->parse_cs) {
868 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
869 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
870 struct amdgpu_bo_va_mapping *m;
871 struct amdgpu_bo *aobj = NULL;
872 struct amdgpu_cs_chunk *chunk;
873 uint64_t offset, va_start;
874 struct amdgpu_ib *ib;
877 chunk = &p->chunks[i];
878 ib = &p->job->ibs[j];
879 chunk_ib = chunk->kdata;
881 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
884 va_start = chunk_ib->va_start & AMDGPU_VA_HOLE_MASK;
885 r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
887 DRM_ERROR("IB va_start is invalid\n");
891 if ((va_start + chunk_ib->ib_bytes) >
892 (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
893 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
897 /* the IB should be reserved at this point */
898 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
903 offset = m->start * AMDGPU_GPU_PAGE_SIZE;
904 kptr += va_start - offset;
906 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
907 amdgpu_bo_kunmap(aobj);
909 r = amdgpu_ring_parse_cs(ring, p, j);
918 p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->root.base.bo);
920 r = amdgpu_bo_vm_update_pte(p);
925 return amdgpu_cs_sync_rings(p);
928 static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
929 struct amdgpu_cs_parser *parser)
931 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
932 struct amdgpu_vm *vm = &fpriv->vm;
934 int r, ce_preempt = 0, de_preempt = 0;
936 for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
937 struct amdgpu_cs_chunk *chunk;
938 struct amdgpu_ib *ib;
939 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
940 struct amdgpu_ring *ring;
942 chunk = &parser->chunks[i];
943 ib = &parser->job->ibs[j];
944 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
946 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
949 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX && amdgpu_sriov_vf(adev)) {
950 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
951 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
957 /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
958 if (ce_preempt > 1 || de_preempt > 1)
962 r = amdgpu_queue_mgr_map(adev, &parser->ctx->queue_mgr, chunk_ib->ip_type,
963 chunk_ib->ip_instance, chunk_ib->ring, &ring);
967 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE) {
968 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
969 if (!parser->ctx->preamble_presented) {
970 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
971 parser->ctx->preamble_presented = true;
975 if (parser->job->ring && parser->job->ring != ring)
978 parser->job->ring = ring;
980 r = amdgpu_ib_get(adev, vm,
981 ring->funcs->parse_cs ? chunk_ib->ib_bytes : 0,
984 DRM_ERROR("Failed to get ib !\n");
988 ib->gpu_addr = chunk_ib->va_start;
989 ib->length_dw = chunk_ib->ib_bytes / 4;
990 ib->flags = chunk_ib->flags;
995 /* UVD & VCE fw doesn't support user fences */
996 if (parser->job->uf_addr && (
997 parser->job->ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
998 parser->job->ring->funcs->type == AMDGPU_RING_TYPE_VCE))
1001 return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->job->ring->idx);
1004 static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
1005 struct amdgpu_cs_chunk *chunk)
1007 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1010 struct drm_amdgpu_cs_chunk_dep *deps;
1012 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
1013 num_deps = chunk->length_dw * 4 /
1014 sizeof(struct drm_amdgpu_cs_chunk_dep);
1016 for (i = 0; i < num_deps; ++i) {
1017 struct amdgpu_ring *ring;
1018 struct amdgpu_ctx *ctx;
1019 struct dma_fence *fence;
1021 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
1025 r = amdgpu_queue_mgr_map(p->adev, &ctx->queue_mgr,
1027 deps[i].ip_instance,
1028 deps[i].ring, &ring);
1030 amdgpu_ctx_put(ctx);
1034 fence = amdgpu_ctx_get_fence(ctx, ring,
1036 if (IS_ERR(fence)) {
1038 amdgpu_ctx_put(ctx);
1041 r = amdgpu_sync_fence(p->adev, &p->job->sync, fence,
1043 dma_fence_put(fence);
1044 amdgpu_ctx_put(ctx);
1052 static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1056 struct dma_fence *fence;
1057 r = drm_syncobj_find_fence(p->filp, handle, &fence);
1061 r = amdgpu_sync_fence(p->adev, &p->job->sync, fence, true);
1062 dma_fence_put(fence);
1067 static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1068 struct amdgpu_cs_chunk *chunk)
1072 struct drm_amdgpu_cs_chunk_sem *deps;
1074 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1075 num_deps = chunk->length_dw * 4 /
1076 sizeof(struct drm_amdgpu_cs_chunk_sem);
1078 for (i = 0; i < num_deps; ++i) {
1079 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle);
1086 static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1087 struct amdgpu_cs_chunk *chunk)
1091 struct drm_amdgpu_cs_chunk_sem *deps;
1092 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1093 num_deps = chunk->length_dw * 4 /
1094 sizeof(struct drm_amdgpu_cs_chunk_sem);
1096 p->post_dep_syncobjs = kmalloc_array(num_deps,
1097 sizeof(struct drm_syncobj *),
1099 p->num_post_dep_syncobjs = 0;
1101 if (!p->post_dep_syncobjs)
1104 for (i = 0; i < num_deps; ++i) {
1105 p->post_dep_syncobjs[i] = drm_syncobj_find(p->filp, deps[i].handle);
1106 if (!p->post_dep_syncobjs[i])
1108 p->num_post_dep_syncobjs++;
1113 static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1114 struct amdgpu_cs_parser *p)
1118 for (i = 0; i < p->nchunks; ++i) {
1119 struct amdgpu_cs_chunk *chunk;
1121 chunk = &p->chunks[i];
1123 if (chunk->chunk_id == AMDGPU_CHUNK_ID_DEPENDENCIES) {
1124 r = amdgpu_cs_process_fence_dep(p, chunk);
1127 } else if (chunk->chunk_id == AMDGPU_CHUNK_ID_SYNCOBJ_IN) {
1128 r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1131 } else if (chunk->chunk_id == AMDGPU_CHUNK_ID_SYNCOBJ_OUT) {
1132 r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1141 static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1145 for (i = 0; i < p->num_post_dep_syncobjs; ++i)
1146 drm_syncobj_replace_fence(p->post_dep_syncobjs[i], p->fence);
1149 static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1150 union drm_amdgpu_cs *cs)
1152 struct amdgpu_ring *ring = p->job->ring;
1153 struct drm_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1154 struct amdgpu_job *job;
1160 amdgpu_mn_lock(p->mn);
1162 for (i = p->bo_list->first_userptr;
1163 i < p->bo_list->num_entries; ++i) {
1164 struct amdgpu_bo *bo = p->bo_list->array[i].robj;
1166 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm)) {
1167 amdgpu_mn_unlock(p->mn);
1168 return -ERESTARTSYS;
1176 r = drm_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1178 amdgpu_job_free(job);
1179 amdgpu_mn_unlock(p->mn);
1183 job->owner = p->filp;
1184 job->fence_ctx = entity->fence_context;
1185 p->fence = dma_fence_get(&job->base.s_fence->finished);
1187 r = amdgpu_ctx_add_fence(p->ctx, ring, p->fence, &seq);
1189 dma_fence_put(p->fence);
1190 dma_fence_put(&job->base.s_fence->finished);
1191 amdgpu_job_free(job);
1192 amdgpu_mn_unlock(p->mn);
1196 amdgpu_cs_post_dependencies(p);
1198 cs->out.handle = seq;
1199 job->uf_sequence = seq;
1201 amdgpu_job_free_resources(job);
1202 amdgpu_ring_priority_get(job->ring, job->base.s_priority);
1204 trace_amdgpu_cs_ioctl(job);
1205 drm_sched_entity_push_job(&job->base, entity);
1207 ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1208 amdgpu_mn_unlock(p->mn);
1213 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1215 struct amdgpu_device *adev = dev->dev_private;
1216 union drm_amdgpu_cs *cs = data;
1217 struct amdgpu_cs_parser parser = {};
1218 bool reserved_buffers = false;
1221 if (!adev->accel_working)
1227 r = amdgpu_cs_parser_init(&parser, data);
1229 DRM_ERROR("Failed to initialize parser !\n");
1233 r = amdgpu_cs_ib_fill(adev, &parser);
1237 r = amdgpu_cs_parser_bos(&parser, data);
1240 DRM_ERROR("Not enough memory for command submission!\n");
1241 else if (r != -ERESTARTSYS)
1242 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1246 reserved_buffers = true;
1248 r = amdgpu_cs_dependencies(adev, &parser);
1250 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1254 for (i = 0; i < parser.job->num_ibs; i++)
1255 trace_amdgpu_cs(&parser, i);
1257 r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1261 r = amdgpu_cs_submit(&parser, cs);
1264 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1269 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1272 * @data: data from userspace
1273 * @filp: file private
1275 * Wait for the command submission identified by handle to finish.
1277 int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1278 struct drm_file *filp)
1280 union drm_amdgpu_wait_cs *wait = data;
1281 struct amdgpu_device *adev = dev->dev_private;
1282 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1283 struct amdgpu_ring *ring = NULL;
1284 struct amdgpu_ctx *ctx;
1285 struct dma_fence *fence;
1288 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1292 r = amdgpu_queue_mgr_map(adev, &ctx->queue_mgr,
1293 wait->in.ip_type, wait->in.ip_instance,
1294 wait->in.ring, &ring);
1296 amdgpu_ctx_put(ctx);
1300 fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
1304 r = dma_fence_wait_timeout(fence, true, timeout);
1305 if (r > 0 && fence->error)
1307 dma_fence_put(fence);
1311 amdgpu_ctx_put(ctx);
1315 memset(wait, 0, sizeof(*wait));
1316 wait->out.status = (r == 0);
1322 * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1324 * @adev: amdgpu device
1325 * @filp: file private
1326 * @user: drm_amdgpu_fence copied from user space
1328 static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1329 struct drm_file *filp,
1330 struct drm_amdgpu_fence *user)
1332 struct amdgpu_ring *ring;
1333 struct amdgpu_ctx *ctx;
1334 struct dma_fence *fence;
1337 ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1339 return ERR_PTR(-EINVAL);
1341 r = amdgpu_queue_mgr_map(adev, &ctx->queue_mgr, user->ip_type,
1342 user->ip_instance, user->ring, &ring);
1344 amdgpu_ctx_put(ctx);
1348 fence = amdgpu_ctx_get_fence(ctx, ring, user->seq_no);
1349 amdgpu_ctx_put(ctx);
1354 int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1355 struct drm_file *filp)
1357 struct amdgpu_device *adev = dev->dev_private;
1358 union drm_amdgpu_fence_to_handle *info = data;
1359 struct dma_fence *fence;
1360 struct drm_syncobj *syncobj;
1361 struct sync_file *sync_file;
1364 fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1366 return PTR_ERR(fence);
1368 switch (info->in.what) {
1369 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1370 r = drm_syncobj_create(&syncobj, 0, fence);
1371 dma_fence_put(fence);
1374 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1375 drm_syncobj_put(syncobj);
1378 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1379 r = drm_syncobj_create(&syncobj, 0, fence);
1380 dma_fence_put(fence);
1383 r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1384 drm_syncobj_put(syncobj);
1387 case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1388 fd = get_unused_fd_flags(O_CLOEXEC);
1390 dma_fence_put(fence);
1394 sync_file = sync_file_create(fence);
1395 dma_fence_put(fence);
1401 fd_install(fd, sync_file->file);
1402 info->out.handle = fd;
1411 * amdgpu_cs_wait_all_fence - wait on all fences to signal
1413 * @adev: amdgpu device
1414 * @filp: file private
1415 * @wait: wait parameters
1416 * @fences: array of drm_amdgpu_fence
1418 static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1419 struct drm_file *filp,
1420 union drm_amdgpu_wait_fences *wait,
1421 struct drm_amdgpu_fence *fences)
1423 uint32_t fence_count = wait->in.fence_count;
1427 for (i = 0; i < fence_count; i++) {
1428 struct dma_fence *fence;
1429 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1431 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1433 return PTR_ERR(fence);
1437 r = dma_fence_wait_timeout(fence, true, timeout);
1438 dma_fence_put(fence);
1446 return fence->error;
1449 memset(wait, 0, sizeof(*wait));
1450 wait->out.status = (r > 0);
1456 * amdgpu_cs_wait_any_fence - wait on any fence to signal
1458 * @adev: amdgpu device
1459 * @filp: file private
1460 * @wait: wait parameters
1461 * @fences: array of drm_amdgpu_fence
1463 static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1464 struct drm_file *filp,
1465 union drm_amdgpu_wait_fences *wait,
1466 struct drm_amdgpu_fence *fences)
1468 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1469 uint32_t fence_count = wait->in.fence_count;
1470 uint32_t first = ~0;
1471 struct dma_fence **array;
1475 /* Prepare the fence array */
1476 array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1481 for (i = 0; i < fence_count; i++) {
1482 struct dma_fence *fence;
1484 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1485 if (IS_ERR(fence)) {
1487 goto err_free_fence_array;
1490 } else { /* NULL, the fence has been already signaled */
1497 r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1500 goto err_free_fence_array;
1503 memset(wait, 0, sizeof(*wait));
1504 wait->out.status = (r > 0);
1505 wait->out.first_signaled = first;
1507 if (first < fence_count && array[first])
1508 r = array[first]->error;
1512 err_free_fence_array:
1513 for (i = 0; i < fence_count; i++)
1514 dma_fence_put(array[i]);
1521 * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1524 * @data: data from userspace
1525 * @filp: file private
1527 int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1528 struct drm_file *filp)
1530 struct amdgpu_device *adev = dev->dev_private;
1531 union drm_amdgpu_wait_fences *wait = data;
1532 uint32_t fence_count = wait->in.fence_count;
1533 struct drm_amdgpu_fence *fences_user;
1534 struct drm_amdgpu_fence *fences;
1537 /* Get the fences from userspace */
1538 fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1543 fences_user = u64_to_user_ptr(wait->in.fences);
1544 if (copy_from_user(fences, fences_user,
1545 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1547 goto err_free_fences;
1550 if (wait->in.wait_all)
1551 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1553 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1562 * amdgpu_cs_find_bo_va - find bo_va for VM address
1564 * @parser: command submission parser context
1566 * @bo: resulting BO of the mapping found
1568 * Search the buffer objects in the command submission context for a certain
1569 * virtual memory address. Returns allocation structure when found, NULL
1572 int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1573 uint64_t addr, struct amdgpu_bo **bo,
1574 struct amdgpu_bo_va_mapping **map)
1576 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1577 struct ttm_operation_ctx ctx = { false, false };
1578 struct amdgpu_vm *vm = &fpriv->vm;
1579 struct amdgpu_bo_va_mapping *mapping;
1582 addr /= AMDGPU_GPU_PAGE_SIZE;
1584 mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1585 if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1588 *bo = mapping->bo_va->base.bo;
1591 /* Double check that the BO is reserved by this CS */
1592 if (READ_ONCE((*bo)->tbo.resv->lock.ctx) != &parser->ticket)
1595 if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1596 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1597 amdgpu_ttm_placement_from_domain(*bo, (*bo)->allowed_domains);
1598 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1603 return amdgpu_ttm_alloc_gart(&(*bo)->tbo);