Re: [PATCH -next] SUNRPC: Clean XPRT_CONGESTED of xprt->state when rpc task is killed

[Trond Myklebust <trondmy@hammerspace.com>]

On Mon, 2021-12-20 at 11:39 +0800, Bixuan Cui wrote:
> ping~
> 
> 在 2021/12/14 下午9:53, Bixuan Cui 写道:
> > When the values of tcp_max_slot_table_entries and
> > sunrpc.tcp_slot_table_entries are lower than the number of rpc
> > tasks,
> > xprt_dynamic_alloc_slot() in xprt_alloc_slot() will return -EAGAIN,
> > and
> > then set xprt->state to XPRT_CONGESTED:
> >    xprt_retry_reserve
> >      ->xprt_do_reserve
> >        ->xprt_alloc_slot
> >          ->xprt_dynamic_alloc_slot // return -EAGAIN and task-
> > >tk_rqstp is NULL
> >            ->xprt_add_backlog // set_bit(XPRT_CONGESTED, &xprt-
> > >state);
> > 
> > When rpc task is killed, XPRT_CONGESTED bit of xprt->state will not
> > be
> > cleaned up and nfs hangs:
> >    rpc_exit_task
> >      ->xprt_release // if (req == NULL) is true, then
> > XPRT_CONGESTED
> >                    // bit not clean
> > 
> > Add xprt_wake_up_backlog(xprt) to clean XPRT_CONGESTED bit in
> > xprt_release().

I'm not seeing how this explanation makes sense. If the task doesn't
hold a slot, then freeing that task isn't going to clear the congestion
caused by all the slots being in use.


-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com