Re: Kernel BUG when registering the ring

[Jann Horn <jannh@google.com>]

On Tue, Feb 11, 2020 at 7:58 PM Jens Axboe <axboe@kernel.dk> wrote:
[...]
> @@ -849,6 +857,8 @@ void io_wq_cancel_all(struct io_wq *wq)
>         for_each_node(node) {
>                 struct io_wqe *wqe = wq->wqes[node];
>
> +               if (!node_online(node))
> +                       continue;
>                 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
>         }
>         rcu_read_unlock();

What is this going to do if a NUMA node is marked as offline (through
a call to node_set_offline() from try_offline_node()) while it has a
worker running, and then afterwards, with the worker still running,
io_wq_cancel_all() is executed? Is that going to potentially hang
because some op is still executing on that node's worker? Or is there
a reason why that can't happen?

[...]
> @@ -1084,6 +1100,8 @@ void io_wq_flush(struct io_wq *wq)
>         for_each_node(node) {
>                 struct io_wqe *wqe = wq->wqes[node];
>
> +               if (!node_online(node))
> +                       continue;
>                 init_completion(&data.done);
>                 INIT_IO_WORK(&data.work, io_wq_flush_func);
>                 data.work.flags |= IO_WQ_WORK_INTERNAL;

(io_wq_flush() is dead code since 05f3fb3c5397, right? Are there plans
to use it again?)