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

[Jason Wang <jasowang@redhat.com>]

On 2018年05月22日 11:46, Michael S. Tsirkin wrote:
> On Tue, May 22, 2018 at 11:22:11AM +0800, Jason Wang wrote:
>>
>> On 2018年05月22日 06:08, Michael S. Tsirkin wrote:
>>> On Mon, May 21, 2018 at 11:47:42AM -0400, David Miller wrote:
>>>> From: Jason Wang <jasowang@redhat.com>
>>>> Date: Fri, 18 May 2018 21:00:43 +0800
>>>>
>>>>> 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@stressinduktion.org>
>>>>> Cc: Eric Dumazet <edumazet@google.com>
>>>>> Fixes: 1bd4978a88ac2 ("tun: honor IFF_UP in tun_get_user()")
>>>>> Signed-off-by: Jason Wang <jasowang@redhat.com>
>>>> I'm no so sure what to do with this patch.
>>>>
>>>> Like Michael says, this flag bit is only checks upon transmit which
>>>> may or may not happen after this point.  It doesn't seem to be
>>>> guaranteed.
>> The flag is checked in tun_chr_poll() as well.
>>
>>> Jason, can't we detect a link up transition and respond accordingly?
>>> What do you think?
>>>
>> I think we've already tried to do this, in tun_net_open() we call
>> write_space(). But the problem is the bit may not be set at that time.
>>
>> A second thought is to set the bit in tun_chr_poll() instead of -EIO like:
>>
>> diff --git a/drivers/net/tun.c b/drivers/net/tun.c
>> index d45ac37..46a1573 100644
>> --- a/drivers/net/tun.c
>> +++ b/drivers/net/tun.c
>> @@ -1423,6 +1423,13 @@ static void tun_net_init(struct net_device *dev)
>>          dev->max_mtu = MAX_MTU - dev->hard_header_len;
>>   }
>>
>> +static bool tun_sock_writeable(struct tun_struct *tun, struct tun_file
>> *tfile)
>> +{
>> +       struct sock *sk = tfile->socket.sk;
>> +
>> +       return (tun->dev->flags & IFF_UP) && sock_writeable(sk);
>> +}
>> +
>>   /* Character device part */
>>
>>   /* Poll */
>> @@ -1445,10 +1452,9 @@ static __poll_t tun_chr_poll(struct file *file,
>> poll_table *wait)
>>          if (!ptr_ring_empty(&tfile->tx_ring))
>>                  mask |= EPOLLIN | EPOLLRDNORM;
>>
>> -       if (tun->dev->flags & IFF_UP &&
>> -           (sock_writeable(sk) ||
>> -            (!test_and_set_bit(SOCKWQ_ASYNC_NOSPACE, &sk->sk_socket->flags)
>> &&
>> -             sock_writeable(sk))))
>> +       if (tun_sock_writeable(tun, tfile) ||
>> +           (!test_and_set_bit(SOCKWQ_ASYNC_NOSPACE, &sk->sk_socket->flags)
>> &&
>> +            tun_sock_writeable(tun, tfile)));
>>                  mask |= EPOLLOUT | EPOLLWRNORM;
>>
>>          if (tun->dev->reg_state != NETREG_REGISTERED)
>>
>> Does this make more sense?
>>
>> Thanks
> I just understood the motivation for doing it on EIO.
> Maybe there's a reason it makes sense here as well,
> but it's far from obvious. I suggest you repost adding
> an explanation in the comment. The original patch will
> be fine with an explanation as well.
>

Ok, let me add explanation on both code and commit log.

Thanks