Re: [PATCH] RDMA/cma: Split apart the multiple uses of the same list heads

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Sep 16, 2021 at 01:11:14PM +0800, Mark Zhang wrote:
> > @@ -4928,10 +4930,10 @@ static void cma_process_remove(struct cma_device *cma_dev)
> >          mutex_lock(&lock);
> >          while (!list_empty(&cma_dev->id_list)) {
> >                  struct rdma_id_private *id_priv = list_first_entry(
> > -                       &cma_dev->id_list, struct rdma_id_private, list);
> > +                       &cma_dev->id_list, struct rdma_id_private, device_item);
> > 
> > -               list_del(&id_priv->listen_list);
> > -               list_del_init(&id_priv->list);
> > +               list_del_init(&id_priv->listen_item);
> 
> Should it still be
>     list_del(&id_priv->listen_list);
> as it isn't dev_id_priv?

Yes, probably should stay here, but it isn't entirely sane

The next code block must trigger the list_del or the
wait_for_completion() below will block forever.

The whole RDMA_CM_EVENT_DEVICE_REMOVAL bit is kind of insane and needs
some cleaning. For instance I think it is a bug if any ULP doesn't
return 1 from the event.

Jason