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

[Yongji Xie <xieyongji@bytedance.com>]

On Wed, Mar 23, 2022 at 4:11 AM Josef Bacik <josef@toxicpanda.com> 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.
> >
> > To fix it, we can not simply remove the WQ_MEM_RECLAIM
> > flag since the recv work is in the memory reclaim path.
> > So this patch tries to create kthreads directly to
> > handle the recv work instead of using workqueue.
> >
>
> I still don't understand why we can't drop WQ_MEM_RECLAIM.  IIRC your argument
> is that we need it because a reconnect could happen under memory pressure and we
> need to be able to queue work for that.  However your code makes it so we're
> just doing a kthread_create(), which isn't coming out of some emergency pool, so
> it's just as likely to fail as a !WQ_MEM_RECLAIM workqueue.  Thanks,
>

I think the key point is the context in which the work thread is
created. It's the context of the nbd process if using kthread_create()
to create a workthread (might do some allocation). Then we can benefit
from the PR_SET_IO_FLUSHER flag, so memory reclaim would never hit the
page cache on the nbd device. But using queue_work() to create a
workthread, the actual thread creation happens in the context of the
work thread rather than the nbd process, so we can't rely on the
PR_SET_IO_FLUSHER flag to avoid deadlock.

Thanks,
Yongji