[PATCH 4/4] drm/amdgpu: add userqueue resume

Shashank Sharma shashank.sharma at amd.com
Wed May 8 20:28:37 UTC 2024


This patch adds support for userqueue resume. What it typically does is
this:
- adds a new delayed work for resuming all the queues.
- schedules this delayed work from the suspend work.
- validates the BOs and replaces the eviction fence before resuming all
  the queues running under this instance of userq manager.

Cc: Alex Deucher <alexander.deucher at amd.com>
Cc: Christian Koenig <christian.koenig at amd.com>
Signed-off-by: Shashank Sharma <shashank.sharma at amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c | 166 ++++++++++++++++++
 .../gpu/drm/amd/include/amdgpu_userqueue.h    |   1 +
 2 files changed, 167 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
index fdbd542e7f53..02ddd713d068 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
@@ -398,6 +398,167 @@ int amdgpu_userq_ioctl(struct drm_device *dev, void *data,
 	return r;
 }
 
+static int
+amdgpu_userqueue_resume_all(struct amdgpu_userq_mgr *uq_mgr)
+{
+	struct amdgpu_device *adev = uq_mgr->adev;
+	const struct amdgpu_userq_funcs *userq_funcs;
+	struct amdgpu_usermode_queue *queue;
+	int queue_id, ret;
+
+	userq_funcs = adev->userq_funcs[AMDGPU_HW_IP_GFX];
+
+	/* Resume all the queues for this process */
+	idr_for_each_entry(&uq_mgr->userq_idr, queue, queue_id) {
+		ret = userq_funcs->resume(uq_mgr, queue);
+		if (ret)
+			DRM_ERROR("Failed to resume queue %d\n", queue_id);
+	}
+
+	return ret;
+}
+
+static int
+amdgpu_userqueue_replace_ev_fence(struct amdgpu_userq_mgr *uq_mgr,
+				  struct drm_exec *exec)
+{
+	int ret;
+	struct amdgpu_fpriv *fpriv = uq_mgr_to_fpriv(uq_mgr);
+	struct amdgpu_vm *vm = &fpriv->vm;
+	struct amdgpu_eviction_fence *old_ef, *new_ef;
+	struct amdgpu_bo_va *bo_va, *tmp;
+
+	old_ef = fpriv->ev_fence;
+	new_ef = amdgpu_eviction_fence_create(fpriv);
+	if (!new_ef) {
+		DRM_ERROR("Failed to create new eviction fence\n");
+		return ret;
+	}
+
+	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
+		struct amdgpu_bo *bo = bo_va->base.bo;
+
+		/* Skip pinned BOs */
+		if (bo->tbo.pin_count)
+			continue;
+
+		ret = drm_exec_lock_obj(exec, &bo->tbo.base);
+		if (unlikely(ret)) {
+			DRM_ERROR("Failed to lock BO for eviction fence replacement\n");
+			goto free_err;
+		}
+
+		/* replace the old eviction fence with new one */
+		amdgpu_eviction_fence_detach(fpriv, old_ef, bo);
+		ret = amdgpu_eviction_fence_attach(new_ef, bo);
+		if (ret) {
+			DRM_ERROR("Failed to attch new eviction fence\n");
+			goto free_err;
+		}
+	}
+
+	/* Update the new eviction fence */
+	fpriv->ev_fence = new_ef;
+	kfree(old_ef);
+	return 0;
+
+free_err:
+	kfree(new_ef);
+	return ret;
+}
+
+/* Expects drm_exec_until_all_locked called on this exec */
+static int
+amdgpu_userqueue_validate_bos(struct amdgpu_userq_mgr *uq_mgr,
+			      struct drm_exec *exec)
+{
+	int ret;
+	struct amdgpu_bo *bo;
+	struct amdgpu_bo_va *bo_va, *tmp;
+	struct amdgpu_fpriv *fpriv = uq_mgr_to_fpriv(uq_mgr);
+	struct amdgpu_vm *vm = &fpriv->vm;
+
+	list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
+		bo = bo_va->base.bo;
+		ret = drm_exec_lock_obj(exec, &bo->tbo.base);
+		if (unlikely(ret)) {
+			DRM_ERROR("Failed to exec lock  for validation\n");
+			goto unlock_all;
+		}
+	}
+
+	list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
+		bo = bo_va->base.bo;
+		ret = drm_exec_lock_obj(exec, &bo->tbo.base);
+		if (unlikely(ret)) {
+			DRM_ERROR("Failed to lock BO for validation\n");
+			goto unlock_all;
+		}
+
+		ret = amdgpu_bo_reserve(bo, false);
+		if (unlikely(ret)) {
+			DRM_ERROR("Failed to reserve BO for validation\n");
+			goto unlock_all;
+		}
+
+		ret = amdgpu_userqueue_validate_bo(bo);
+		amdgpu_bo_unreserve(bo);
+		if (ret) {
+			DRM_ERROR("Failed to validate BO\n");
+			goto unlock_all;
+		}
+	}
+
+	ret = amdgpu_vm_handle_moved(uq_mgr->adev, vm, NULL);
+	if (ret)
+		DRM_ERROR("Failed to handle moved BOs\n");
+
+unlock_all:
+	return ret;
+}
+
+static void amdgpu_userqueue_resume_worker(struct work_struct *work)
+{
+	int ret;
+	struct amdgpu_userq_mgr *uq_mgr = work_to_uq_mgr(work, resume_work.work);
+	struct amdgpu_fpriv *fpriv = uq_mgr_to_fpriv(uq_mgr);
+	struct amdgpu_vm *vm = &fpriv->vm;
+	struct drm_exec exec;
+
+	mutex_lock(&uq_mgr->userq_mutex);
+
+	drm_exec_init(&exec, 0, 0);
+	drm_exec_until_all_locked(&exec) {
+		ret = amdgpu_vm_lock_pd(vm, &exec, 2);
+		if (unlikely(ret)) {
+			DRM_ERROR("Failed to lock PD\n");
+			goto unlock_all;
+		}
+	}
+
+	ret = amdgpu_userqueue_validate_bos(uq_mgr, &exec);
+	if (ret) {
+		DRM_ERROR("Failed to validate BOs to restore\n");
+		goto unlock_all;
+	}
+
+	ret = amdgpu_userqueue_replace_ev_fence(uq_mgr, &exec);
+	if (ret) {
+		DRM_ERROR("Failed to signal eviction\n");
+		goto unlock_all;
+	}
+
+	ret = amdgpu_userqueue_resume_all(uq_mgr);
+	if (ret) {
+		DRM_ERROR("Failed to resume all queues\n");
+		goto unlock_all;
+	}
+
+unlock_all:
+	drm_exec_fini(&exec);
+	mutex_unlock(&uq_mgr->userq_mutex);
+}
+
 static int
 amdgpu_userqueue_suspend_all(struct amdgpu_userq_mgr *uq_mgr)
 {
@@ -442,6 +603,10 @@ amdgpu_userqueue_suspend_worker(struct work_struct *work)
 		return;
 	}
 
+	/* Schedule a work to restore userqueue after 100 ms */
+	schedule_delayed_work(&uq_mgr->resume_work,
+			      msecs_to_jiffies(AMDGPU_USERQ_RESUME_TIME_MS));
+
 unlock:
 	mutex_unlock(&uq_mgr->userq_mutex);
 }
@@ -475,6 +640,7 @@ int amdgpu_userq_mgr_init(struct amdgpu_userq_mgr *userq_mgr, struct amdgpu_devi
 	/* This reference is required for suspend work */
 	fpriv->ev_fence->uq_mgr = userq_mgr;
 	INIT_DELAYED_WORK(&userq_mgr->suspend_work, amdgpu_userqueue_suspend_worker);
+	INIT_DELAYED_WORK(&userq_mgr->resume_work, amdgpu_userqueue_resume_worker);
 	return 0;
 }
 
diff --git a/drivers/gpu/drm/amd/include/amdgpu_userqueue.h b/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
index 647e63bf03ab..2e3fe784188e 100644
--- a/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
+++ b/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
@@ -73,6 +73,7 @@ struct amdgpu_userq_mgr {
 	struct amdgpu_device		*adev;
 
 	struct delayed_work		suspend_work;
+	struct delayed_work		resume_work;
 };
 
 int amdgpu_userq_ioctl(struct drm_device *dev, void *data, struct drm_file *filp);
-- 
2.43.2



More information about the amd-gfx mailing list