Re: [PATCH v2] nbd: Don't use workqueue to handle recv work

From: Yongji Xie
Date: Tue Jan 04 2022 - 00:32:02 EST


On Tue, Jan 4, 2022 at 12:10 AM Josef Bacik <josef@xxxxxxxxxxxxxx> wrote:
>
> On Thu, Dec 30, 2021 at 12:01:23PM +0800, Yongji Xie wrote:
> > On Thu, Dec 30, 2021 at 1:35 AM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
> > >
> > > On Mon, Dec 27, 2021 at 05:12:41PM +0800, Xie Yongji wrote:
> > > > The rescuer thread might take over the works queued on
> > > > the workqueue when the worker thread creation timed out.
> > > > If this happens, we have no chance to create multiple
> > > > recv threads which causes I/O hung on this nbd device.
> > >
> > > If a workqueue is used there aren't really 'receive threads'.
> > > What is the deadlock here?
> >
> > We might have multiple recv works, and those recv works won't quit
> > unless the socket is closed. If the rescuer thread takes over those
> > works, only the first recv work can run. The I/O needed to be handled
> > in other recv works would be hung since no thread can handle them.
> >
>
> I'm not following this explanation. What is the rescuer thread you're talking

https://www.kernel.org/doc/html/latest/core-api/workqueue.html#c.rescuer_thread

> about? If there's an error we close the socket which will error out the recvmsg
> which will make the recv workqueue close down.

When to close the socket? The nbd daemon doesn't know what happens in
the kernel.

>
> > In that case, we can see below stacks in rescuer thread:
> >
> > __schedule
> > schedule
> > scheule_timeout
> > unix_stream_read_generic
> > unix_stream_recvmsg
> > sock_xmit
> > nbd_read_stat
> > recv_work
> > process_one_work
> > rescuer_thread
> > kthread
> > ret_from_fork
>
> This is just the thing hanging waiting for an incoming request, so this doesn't
> tell me anything. Thanks,
>

The point is the *recv_work* is handled in the *rescuer_thread*.
Normally it should be handled in *work_thread* like:

__schedule
schedule
scheule_timeout
unix_stream_read_generic
unix_stream_recvmsg
sock_xmit
nbd_read_stat
recv_work
process_one_work
*work_thread*
kthread
ret_from_fork

Thanks,
Yongji