1 files changed, 235 insertions, 127 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
index 1f040d85ac47..4f352ec9dec4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
@@ -104,10 +104,11 @@ int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
 		}
 		break;
 	case AMDGPU_HW_IP_DMA:
-		if (ring < 2) {
-			*out_ring = &adev->sdma[ring].ring;
+		if (ring < adev->sdma.num_instances) {
+			*out_ring = &adev->sdma.instance[ring].ring;
 		} else {
-			DRM_ERROR("only two SDMA rings are supported\n");
+			DRM_ERROR("only %d SDMA rings are supported\n",
+				  adev->sdma.num_instances);
 			return -EINVAL;
 		}
 		break;
@@ -130,42 +131,42 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 {
 	union drm_amdgpu_cs *cs = data;
 	uint64_t *chunk_array_user;
-	uint64_t *chunk_array = NULL;
+	uint64_t *chunk_array;
 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
-	unsigned size, i;
-	int r = 0;
+	unsigned size;
+	int i;
+	int ret;
 
-	if (!cs->in.num_chunks)
-		goto out;
+	if (cs->in.num_chunks == 0)
+		return 0;
+
+	chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
+	if (!chunk_array)
+		return -ENOMEM;
 
 	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
 	if (!p->ctx) {
-		r = -EINVAL;
-		goto out;
+		ret = -EINVAL;
+		goto free_chunk;
 	}
+
 	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
 
 	/* get chunks */
 	INIT_LIST_HEAD(&p->validated);
-	chunk_array = kcalloc(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
-	if (chunk_array == NULL) {
-		r = -ENOMEM;
-		goto out;
-	}
-
-	chunk_array_user = (uint64_t *)(unsigned long)(cs->in.chunks);
+	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
 	if (copy_from_user(chunk_array, chunk_array_user,
 			   sizeof(uint64_t)*cs->in.num_chunks)) {
-		r = -EFAULT;
-		goto out;
+		ret = -EFAULT;
+		goto put_bo_list;
 	}
 
 	p->nchunks = cs->in.num_chunks;
-	p->chunks = kcalloc(p->nchunks, sizeof(struct amdgpu_cs_chunk),
+	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
 			    GFP_KERNEL);
-	if (p->chunks == NULL) {
-		r = -ENOMEM;
-		goto out;
+	if (!p->chunks) {
+		ret = -ENOMEM;
+		goto put_bo_list;
 	}
 
 	for (i = 0; i < p->nchunks; i++) {
@@ -176,8 +177,9 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
 		if (copy_from_user(&user_chunk, chunk_ptr,
 				       sizeof(struct drm_amdgpu_cs_chunk))) {
-			r = -EFAULT;
-			goto out;
+			ret = -EFAULT;
+			i--;
+			goto free_partial_kdata;
 		}
 		p->chunks[i].chunk_id = user_chunk.chunk_id;
 		p->chunks[i].length_dw = user_chunk.length_dw;
@@ -188,13 +190,14 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 
 		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
 		if (p->chunks[i].kdata == NULL) {
-			r = -ENOMEM;
-			goto out;
+			ret = -ENOMEM;
+			i--;
+			goto free_partial_kdata;
 		}
 		size *= sizeof(uint32_t);
 		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
-			r = -EFAULT;
-			goto out;
+			ret = -EFAULT;
+			goto free_partial_kdata;
 		}
 
 		switch (p->chunks[i].chunk_id) {
@@ -214,15 +217,17 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 				gobj = drm_gem_object_lookup(p->adev->ddev,
 							     p->filp, handle);
 				if (gobj == NULL) {
-					r = -EINVAL;
-					goto out;
+					ret = -EINVAL;
+					goto free_partial_kdata;
 				}
 
 				p->uf.bo = gem_to_amdgpu_bo(gobj);
+				amdgpu_bo_ref(p->uf.bo);
+				drm_gem_object_unreference_unlocked(gobj);
 				p->uf.offset = fence_data->offset;
 			} else {
-				r = -EINVAL;
-				goto out;
+				ret = -EINVAL;
+				goto free_partial_kdata;
 			}
 			break;
 
@@ -230,20 +235,35 @@ int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
 			break;
 
 		default:
-			r = -EINVAL;
-			goto out;
+			ret = -EINVAL;
+			goto free_partial_kdata;
 		}
 	}
 
+
 	p->ibs = kcalloc(p->num_ibs, sizeof(struct amdgpu_ib), GFP_KERNEL);
 	if (!p->ibs) {
-		r = -ENOMEM;
-		goto out;
+		ret = -ENOMEM;
+		goto free_all_kdata;
 	}
 
-out:
 	kfree(chunk_array);
-	return r;
+	return 0;
+
+free_all_kdata:
+	i = p->nchunks - 1;
+free_partial_kdata:
+	for (; i >= 0; i--)
+		drm_free_large(p->chunks[i].kdata);
+	kfree(p->chunks);
+put_bo_list:
+	if (p->bo_list)
+		amdgpu_bo_list_put(p->bo_list);
+	amdgpu_ctx_put(p->ctx);
+free_chunk:
+	kfree(chunk_array);
+
+	return ret;
 }
 
 /* Returns how many bytes TTM can move per IB.
@@ -298,25 +318,17 @@ static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
 	return max(bytes_moved_threshold, 1024*1024ull);
 }
 
-int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p)
+int amdgpu_cs_list_validate(struct amdgpu_device *adev,
+			    struct amdgpu_vm *vm,
+			    struct list_head *validated)
 {
-	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
-	struct amdgpu_vm *vm = &fpriv->vm;
-	struct amdgpu_device *adev = p->adev;
 	struct amdgpu_bo_list_entry *lobj;
-	struct list_head duplicates;
 	struct amdgpu_bo *bo;
 	u64 bytes_moved = 0, initial_bytes_moved;
 	u64 bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(adev);
 	int r;
 
-	INIT_LIST_HEAD(&duplicates);
-	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates);
-	if (unlikely(r != 0)) {
-		return r;
-	}
-
-	list_for_each_entry(lobj, &p->validated, tv.head) {
+	list_for_each_entry(lobj, validated, tv.head) {
 		bo = lobj->robj;
 		if (!bo->pin_count) {
 			u32 domain = lobj->prefered_domains;
@@ -331,7 +343,7 @@ int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p)
 			 * into account. We don't want to disallow buffer moves
 			 * completely.
 			 */
-			if (current_domain != AMDGPU_GEM_DOMAIN_CPU &&
+			if ((lobj->allowed_domains & current_domain) != 0 &&
 			    (domain & current_domain) == 0 && /* will be moved */
 			    bytes_moved > bytes_moved_threshold) {
 				/* don't move it */
@@ -350,7 +362,6 @@ int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p)
 					domain = lobj->allowed_domains;
 					goto retry;
 				}
-				ttm_eu_backoff_reservation(&p->ticket, &p->validated);
 				return r;
 			}
 		}
@@ -363,6 +374,7 @@ static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p)
 {
 	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
 	struct amdgpu_cs_buckets buckets;
+	struct list_head duplicates;
 	bool need_mmap_lock = false;
 	int i, r;
 
@@ -382,8 +394,22 @@ static int amdgpu_cs_parser_relocs(struct amdgpu_cs_parser *p)
 	if (need_mmap_lock)
 		down_read(&current->mm->mmap_sem);
 
-	r = amdgpu_cs_list_validate(p);
+	INIT_LIST_HEAD(&duplicates);
+	r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true, &duplicates);
+	if (unlikely(r != 0))
+		goto error_reserve;
+
+	r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &p->validated);
+	if (r)
+		goto error_validate;
 
+	r = amdgpu_cs_list_validate(p->adev, &fpriv->vm, &duplicates);
+
+error_validate:
+	if (r)
+		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
+
+error_reserve:
 	if (need_mmap_lock)
 		up_read(&current->mm->mmap_sem);
 
@@ -441,17 +467,19 @@ static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bo
 		list_sort(NULL, &parser->validated, cmp_size_smaller_first);
 
 		ttm_eu_fence_buffer_objects(&parser->ticket,
-				&parser->validated,
-				&parser->ibs[parser->num_ibs-1].fence->base);
+					    &parser->validated,
+					    parser->fence);
 	} else if (backoff) {
 		ttm_eu_backoff_reservation(&parser->ticket,
 					   &parser->validated);
 	}
+	fence_put(parser->fence);
 
 	if (parser->ctx)
 		amdgpu_ctx_put(parser->ctx);
 	if (parser->bo_list)
 		amdgpu_bo_list_put(parser->bo_list);
+
 	drm_free_large(parser->vm_bos);
 	for (i = 0; i < parser->nchunks; i++)
 		drm_free_large(parser->chunks[i].kdata);
@@ -461,7 +489,7 @@ static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bo
 			amdgpu_ib_free(parser->adev, &parser->ibs[i]);
 	kfree(parser->ibs);
 	if (parser->uf.bo)
-		drm_gem_object_unreference_unlocked(&parser->uf.bo->gem_base);
+		amdgpu_bo_unref(&parser->uf.bo);
 }
 
 static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
@@ -476,12 +504,18 @@ static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
 	if (r)
 		return r;
 
+	r = amdgpu_sync_fence(adev, &p->ibs[0].sync, vm->page_directory_fence);
+	if (r)
+		return r;
+
 	r = amdgpu_vm_clear_freed(adev, vm);
 	if (r)
 		return r;
 
 	if (p->bo_list) {
 		for (i = 0; i < p->bo_list->num_entries; i++) {
+			struct fence *f;
+
 			/* ignore duplicates */
 			bo = p->bo_list->array[i].robj;
 			if (!bo)
@@ -495,11 +529,29 @@ static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
 			if (r)
 				return r;
 
-			amdgpu_sync_fence(&p->ibs[0].sync, bo_va->last_pt_update);
+			f = bo_va->last_pt_update;
+			r = amdgpu_sync_fence(adev, &p->ibs[0].sync, f);
+			if (r)
+				return r;
+		}
+
+	}
+
+	r = amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
+
+	if (amdgpu_vm_debug && p->bo_list) {
+		/* Invalidate all BOs to test for userspace bugs */
+		for (i = 0; i < p->bo_list->num_entries; i++) {
+			/* ignore duplicates */
+			bo = p->bo_list->array[i].robj;
+			if (!bo)
+				continue;
+
+			amdgpu_vm_bo_invalidate(adev, bo);
 		}
 	}
 
-	return amdgpu_vm_clear_invalids(adev, vm, &p->ibs[0].sync);
+	return r;
 }
 
 static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
@@ -523,18 +575,10 @@ static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
 		}
 	}
 
-	mutex_lock(&vm->mutex);
 	r = amdgpu_bo_vm_update_pte(parser, vm);
-	if (r) {
-		goto out;
-	}
-	amdgpu_cs_sync_rings(parser);
-
-	r = amdgpu_ib_schedule(adev, parser->num_ibs, parser->ibs,
-			       parser->filp);
+	if (!r)
+		amdgpu_cs_sync_rings(parser);
 
-out:
-	mutex_unlock(&vm->mutex);
 	return r;
 }
 
@@ -650,7 +694,6 @@ static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
 			ib->oa_size = amdgpu_bo_size(oa);
 		}
 	}
-
 	/* wrap the last IB with user fence */
 	if (parser->uf.bo) {
 		struct amdgpu_ib *ib = &parser->ibs[parser->num_ibs - 1];
@@ -693,9 +736,9 @@ static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
 			sizeof(struct drm_amdgpu_cs_chunk_dep);
 
 		for (j = 0; j < num_deps; ++j) {
-			struct amdgpu_fence *fence;
 			struct amdgpu_ring *ring;
 			struct amdgpu_ctx *ctx;
+			struct fence *fence;
 
 			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
 					       deps[j].ip_instance,
@@ -707,86 +750,143 @@ static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
 			if (ctx == NULL)
 				return -EINVAL;
 
-			r = amdgpu_fence_recreate(ring, p->filp,
-						  deps[j].handle,
-						  &fence);
-			if (r) {
+			fence = amdgpu_ctx_get_fence(ctx, ring,
+						     deps[j].handle);
+			if (IS_ERR(fence)) {
+				r = PTR_ERR(fence);
 				amdgpu_ctx_put(ctx);
 				return r;
-			}
 
-			amdgpu_sync_fence(&ib->sync, fence);
-			amdgpu_fence_unref(&fence);
-			amdgpu_ctx_put(ctx);
+			} else if (fence) {
+				r = amdgpu_sync_fence(adev, &ib->sync, fence);
+				fence_put(fence);
+				amdgpu_ctx_put(ctx);
+				if (r)
+					return r;
+			}
 		}
 	}
 
 	return 0;
 }
 
+static int amdgpu_cs_free_job(struct amdgpu_job *job)
+{
+	int i;
+	if (job->ibs)
+		for (i = 0; i < job->num_ibs; i++)
+			amdgpu_ib_free(job->adev, &job->ibs[i]);
+	kfree(job->ibs);
+	if (job->uf.bo)
+		amdgpu_bo_unref(&job->uf.bo);
+	return 0;
+}
+
 int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 {
 	struct amdgpu_device *adev = dev->dev_private;
 	union drm_amdgpu_cs *cs = data;
-	struct amdgpu_cs_parser parser;
-	int r, i;
+	struct amdgpu_cs_parser parser = {};
 	bool reserved_buffers = false;
+	int i, r;
 
-	down_read(&adev->exclusive_lock);
-	if (!adev->accel_working) {
-		up_read(&adev->exclusive_lock);
+	if (!adev->accel_working)
 		return -EBUSY;
-	}
-	/* initialize parser */
-	memset(&parser, 0, sizeof(struct amdgpu_cs_parser));
-	parser.filp = filp;
+
 	parser.adev = adev;
+	parser.filp = filp;
+
 	r = amdgpu_cs_parser_init(&parser, data);
 	if (r) {
 		DRM_ERROR("Failed to initialize parser !\n");
 		amdgpu_cs_parser_fini(&parser, r, false);
-		up_read(&adev->exclusive_lock);
 		r = amdgpu_cs_handle_lockup(adev, r);
 		return r;
 	}
-
 	r = amdgpu_cs_parser_relocs(&parser);
-	if (r) {
-		if (r != -ERESTARTSYS) {
-			if (r == -ENOMEM)
-				DRM_ERROR("Not enough memory for command submission!\n");
-			else
-				DRM_ERROR("Failed to process the buffer list %d!\n", r);
-		}
-	}
-
-	if (!r) {
+	if (r == -ENOMEM)
+		DRM_ERROR("Not enough memory for command submission!\n");
+	else if (r && r != -ERESTARTSYS)
+		DRM_ERROR("Failed to process the buffer list %d!\n", r);
+	else if (!r) {
 		reserved_buffers = true;
 		r = amdgpu_cs_ib_fill(adev, &parser);
 	}
 
-	if (!r)
+	if (!r) {
 		r = amdgpu_cs_dependencies(adev, &parser);
-
-	if (r) {
-		amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
-		up_read(&adev->exclusive_lock);
-		r = amdgpu_cs_handle_lockup(adev, r);
-		return r;
+		if (r)
+			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
 	}
 
+	if (r)
+		goto out;
+
 	for (i = 0; i < parser.num_ibs; i++)
 		trace_amdgpu_cs(&parser, i);
 
 	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
-	if (r) {
+	if (r)
 		goto out;
+
+	if (amdgpu_enable_scheduler && parser.num_ibs) {
+		struct amdgpu_ring * ring = parser.ibs->ring;
+		struct amd_sched_fence *fence;
+		struct amdgpu_job *job;
+
+		job = kzalloc(sizeof(struct amdgpu_job), GFP_KERNEL);
+		if (!job) {
+			r = -ENOMEM;
+			goto out;
+		}
+
+		job->base.sched = &ring->sched;
+		job->base.s_entity = &parser.ctx->rings[ring->idx].entity;
+		job->adev = parser.adev;
+		job->owner = parser.filp;
+		job->free_job = amdgpu_cs_free_job;
+
+		job->ibs = parser.ibs;
+		job->num_ibs = parser.num_ibs;
+		parser.ibs = NULL;
+		parser.num_ibs = 0;
+
+		if (job->ibs[job->num_ibs - 1].user) {
+			job->uf = parser.uf;
+			job->ibs[job->num_ibs - 1].user = &job->uf;
+			parser.uf.bo = NULL;
+		}
+
+		fence = amd_sched_fence_create(job->base.s_entity,
+					       parser.filp);
+		if (!fence) {
+			r = -ENOMEM;
+			amdgpu_cs_free_job(job);
+			kfree(job);
+			goto out;
+		}
+		job->base.s_fence = fence;
+		parser.fence = fence_get(&fence->base);
+
+		cs->out.handle = amdgpu_ctx_add_fence(parser.ctx, ring,
+						      &fence->base);
+		job->ibs[job->num_ibs - 1].sequence = cs->out.handle;
+
+		trace_amdgpu_cs_ioctl(job);
+		amd_sched_entity_push_job(&job->base);
+
+	} else {
+		struct amdgpu_fence *fence;
+
+		r = amdgpu_ib_schedule(adev, parser.num_ibs, parser.ibs,
+				       parser.filp);
+		fence = parser.ibs[parser.num_ibs - 1].fence;
+		parser.fence = fence_get(&fence->base);
+		cs->out.handle = parser.ibs[parser.num_ibs - 1].sequence;
 	}
 
-	cs->out.handle = parser.ibs[parser.num_ibs - 1].fence->seq;
 out:
-	amdgpu_cs_parser_fini(&parser, r, true);
-	up_read(&adev->exclusive_lock);
+	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
 	r = amdgpu_cs_handle_lockup(adev, r);
 	return r;
 }
@@ -806,30 +906,29 @@ int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
 	union drm_amdgpu_wait_cs *wait = data;
 	struct amdgpu_device *adev = dev->dev_private;
 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
-	struct amdgpu_fence *fence = NULL;
 	struct amdgpu_ring *ring = NULL;
 	struct amdgpu_ctx *ctx;
+	struct fence *fence;
 	long r;
 
-	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
-	if (ctx == NULL)
-		return -EINVAL;
-
 	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
 			       wait->in.ring, &ring);
-	if (r) {
-		amdgpu_ctx_put(ctx);
+	if (r)
 		return r;
-	}
 
-	r = amdgpu_fence_recreate(ring, filp, wait->in.handle, &fence);
-	if (r) {
-		amdgpu_ctx_put(ctx);
-		return r;
-	}
+	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
+	if (ctx == NULL)
+		return -EINVAL;
+
+	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
+	if (IS_ERR(fence))
+		r = PTR_ERR(fence);
+	else if (fence) {
+		r = fence_wait_timeout(fence, true, timeout);
+		fence_put(fence);
+	} else
+		r = 1;
 
-	r = fence_wait_timeout(&fence->base, true, timeout);
-	amdgpu_fence_unref(&fence);
 	amdgpu_ctx_put(ctx);
 	if (r < 0)
 		return r;
@@ -864,7 +963,16 @@ amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
 		if (!reloc->bo_va)
 			continue;
 
-		list_for_each_entry(mapping, &reloc->bo_va->mappings, list) {
+		list_for_each_entry(mapping, &reloc->bo_va->valids, list) {
+			if (mapping->it.start > addr ||
+			    addr > mapping->it.last)
+				continue;
+
+			*bo = reloc->bo_va->bo;
+			return mapping;
+		}
+
+		list_for_each_entry(mapping, &reloc->bo_va->invalids, list) {
 			if (mapping->it.start > addr ||
 			    addr > mapping->it.last)
 				continue;