Re: [PATCH net] tuntap: raise EPOLLOUT on device up

[Jason Wang]

On 2018å05æ18æ 22:46, Michael S. Tsirkin wrote:
> On Fri, May 18, 2018 at 10:11:54PM +0800, Jason Wang wrote:
> > On 2018å05æ18æ 22:06, Michael S. Tsirkin wrote:
> > > On Fri, May 18, 2018 at 10:00:31PM +0800, Jason Wang wrote:
> > > > On 2018å05æ18æ 21:26, Jason Wang wrote:
> > > > > On 2018å05æ18æ 21:13, Michael S. Tsirkin wrote:
> > > > > > On Fri, May 18, 2018 at 09:00:43PM +0800, Jason Wang wrote:
> > > > > > > We return -EIO on device down but can not raise EPOLLOUT after it was
> > > > > > > up. This may confuse user like vhost which expects tuntap to raise
> > > > > > > EPOLLOUT to re-enable its TX routine after tuntap is down. This could
> > > > > > > be easily reproduced by transmitting packets from VM while down and up
> > > > > > > the tap device. Fixing this by set SOCKWQ_ASYNC_NOSPACE on -EIO.
> > > > > > >
> > > > > > > Cc: Hannes Frederic Sowa <hannes@xxxxxxxxxxxxxxxxxxx>
> > > > > > > Cc: Eric Dumazet <edumazet@xxxxxxxxxx>
> > > > > > > Fixes: 1bd4978a88ac2 ("tun: honor IFF_UP in tun_get_user()")
> > > > > > > Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx>
> > > > > > > ---
> > > > > > >   Â drivers/net/tun.c | 4 +++-
> > > > > > >   Â 1 file changed, 3 insertions(+), 1 deletion(-)
> > > > > > >
> > > > > > > diff --git a/drivers/net/tun.c b/drivers/net/tun.c
> > > > > > > index d45ac37..1b29761 100644
> > > > > > > --- a/drivers/net/tun.c
> > > > > > > +++ b/drivers/net/tun.c
> > > > > > > @@ -1734,8 +1734,10 @@ static ssize_t tun_get_user(struct
> > > > > > > tun_struct *tun, struct tun_file *tfile,
> > > > > > >   ÂÂÂÂÂ int skb_xdp = 1;
> > > > > > >   ÂÂÂÂÂ bool frags = tun_napi_frags_enabled(tun);
> > > > > > >   Â -ÂÂÂ if (!(tun->dev->flags & IFF_UP))
> > > > > > > +ÂÂÂ if (!(tun->dev->flags & IFF_UP)) {
> > > > > > Isn't this racy? What if flag is cleared at this point?
> > > > > I think you mean "set at this point"? Then yes, so we probably need to
> > > > > set the bit during tun_net_close().
> > > > >
> > > > > Thanks
> > > > Looks no need, vhost will poll socket after it see EIO. So we are ok here?
> > > >
> > > > Thanks
> > > In fact I don't even understand why does this help any longer.
> > We disable tx polling and only enable it on demand for a better rx
> > performance. You may want to have a look at :
> >
> > commit feb8892cb441c742d4220cf7ced001e7fa070731
> > Author: Jason Wang <jasowang@xxxxxxxxxx>
> > Date:ÂÂ Mon Nov 13 11:45:34 2017 +0800
> >
> >  ÂÂÂ vhost_net: conditionally enable tx polling
> >
> > Thanks
>
> Question is, what looks at SOCKWQ_ASYNC_NOSPACE.
> I think it's tested when packet is transmitted,
> but there is no guarantee here any packet will
> ever be transmitted.

Well, actually, I do plan to disable vq polling from the beginning. But 
looks like you do not want this:

See https://patchwork.kernel.org/patch/10034025/

Thanks