my crunch time is over so I can review these on time now :)
one small comment below, although it doesn't stop me from giving my R-B here:
Reviewed-by: Lyude Paul lyude@redhat.com
On Wed, 2020-06-10 at 16:47 +0300, Imre Deak wrote:
Atm, a pending delayed destroy work during module removal will be canceled, leaving behind MST ports, mstbs. Fix this by using a dedicated workqueue which will be drained of requeued items as well when destroying it.
v2:
- Check if wq is NULL before calling destroy_workqueue().
Cc: Lyude Paul lyude@redhat.com Cc: Stanislav Lisovskiy stanislav.lisovskiy@intel.com Reviewed-by: Stanislav Lisovskiy stanislav.lisovskiy@intel.com Signed-off-by: Imre Deak imre.deak@intel.com
drivers/gpu/drm/drm_dp_mst_topology.c | 19 ++++++++++++++++--- include/drm/drm_dp_mst_helper.h | 8 ++++++++ 2 files changed, 24 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/drm_dp_mst_topology.c b/drivers/gpu/drm/drm_dp_mst_topology.c index eff8d6ac0273..a5f67b9db7fa 100644 --- a/drivers/gpu/drm/drm_dp_mst_topology.c +++ b/drivers/gpu/drm/drm_dp_mst_topology.c @@ -1604,7 +1604,7 @@ static void drm_dp_destroy_mst_branch_device(struct kref *kref) mutex_lock(&mgr->delayed_destroy_lock); list_add(&mstb->destroy_next, &mgr->destroy_branch_device_list); mutex_unlock(&mgr->delayed_destroy_lock);
- schedule_work(&mgr->delayed_destroy_work);
- queue_work(mgr->delayed_destroy_wq, &mgr->delayed_destroy_work);
}
/** @@ -1721,7 +1721,7 @@ static void drm_dp_destroy_port(struct kref *kref) mutex_lock(&mgr->delayed_destroy_lock); list_add(&port->next, &mgr->destroy_port_list); mutex_unlock(&mgr->delayed_destroy_lock);
- schedule_work(&mgr->delayed_destroy_work);
- queue_work(mgr->delayed_destroy_wq, &mgr->delayed_destroy_work);
}
/** @@ -5182,6 +5182,15 @@ int drm_dp_mst_topology_mgr_init(struct drm_dp_mst_topology_mgr *mgr, INIT_LIST_HEAD(&mgr->destroy_port_list); INIT_LIST_HEAD(&mgr->destroy_branch_device_list); INIT_LIST_HEAD(&mgr->up_req_list);
- /*
* delayed_destroy_work will be queued on a dedicated WQ, so that any* requeuing will be also flushed when deiniting the topology manager.*/- mgr->delayed_destroy_wq = alloc_ordered_workqueue("drm_dp_mst_wq", 0);
- if (mgr->delayed_destroy_wq == NULL)
return -ENOMEM;- INIT_WORK(&mgr->work, drm_dp_mst_link_probe_work); INIT_WORK(&mgr->tx_work, drm_dp_tx_work); INIT_WORK(&mgr->delayed_destroy_work, drm_dp_delayed_destroy_work);
@@ -5226,7 +5235,11 @@ void drm_dp_mst_topology_mgr_destroy(struct drm_dp_mst_topology_mgr *mgr) { drm_dp_mst_topology_mgr_set_mst(mgr, false); flush_work(&mgr->work);
- cancel_work_sync(&mgr->delayed_destroy_work);
- /* The following will also drain any requeued work on the WQ. */
- if (mgr->delayed_destroy_wq) {
destroy_workqueue(mgr->delayed_destroy_wq);mgr->delayed_destroy_wq = NULL;- }
We should definitely cleanup the cleanup in this function, I don't mind submitting some patches to do it today if you poke me on IRC once you've got this pushed to drm-misc-next
mutex_lock(&mgr->payload_lock); kfree(mgr->payloads); mgr->payloads = NULL; diff --git a/include/drm/drm_dp_mst_helper.h b/include/drm/drm_dp_mst_helper.h index 9e1ffcd7cb68..17b568c6f4f8 100644 --- a/include/drm/drm_dp_mst_helper.h +++ b/include/drm/drm_dp_mst_helper.h @@ -672,6 +672,14 @@ struct drm_dp_mst_topology_mgr { * @destroy_branch_device_list. */ struct mutex delayed_destroy_lock;
- /**
* @delayed_destroy_wq: Workqueue used for delayed_destroy_work items.* A dedicated WQ makes it possible to drain any requeued work items* on it.*/- struct workqueue_struct *delayed_destroy_wq;
- /**
- @delayed_destroy_work: Work item to destroy MST port and branch
- devices, needed to avoid locking inversion.