Re: [Qemu-devel] tap devices not receiving packets from a bridge

[Michael S. Tsirkin]

On Tue, Feb 12, 2013 at 10:10:24AM +0100, Peter Lieven wrote:
> 
> Am 12.02.2013 um 10:08 schrieb "Michael S. Tsirkin" <address@hidden>:
> 
> > On Tue, Feb 12, 2013 at 08:06:04AM +0100, Peter Lieven wrote:
> >> On 23.01.2013 11:03, Michael S. Tsirkin wrote:
> >>> On Tue, Jan 22, 2013 at 10:04:07AM +0100, Peter Lieven wrote:
> >>>> On 23.11.2012 12:01, Michael S. Tsirkin wrote:
> >>>>> On Fri, Nov 23, 2012 at 10:41:21AM +0100, Peter Lieven wrote:
> >>>>>> 
> >>>>>> Am 23.11.2012 um 08:02 schrieb Stefan Hajnoczi:
> >>>>>> 
> >>>>>>> On Thu, Nov 22, 2012 at 03:29:52PM +0100, Peter Lieven wrote:
> >>>>>>>> is anyone aware of a problem with the linux network bridge that in 
> >>>>>>>> very rare circumstances stops
> >>>>>>>> a bridge from sending pakets to a tap device?
> >>>>>>>> 
> >>>>>>>> My problem occurs in conjunction with vanilla qemu-kvm-1.2.0 and 
> >>>>>>>> Ubuntu Kernel 3.2.0-34.53
> >>>>>>>> which is based on Linux 3.2.33.
> >>>>>>>> 
> >>>>>>>> I was not yet able to reproduce the issue, it happens in really rare 
> >>>>>>>> cases. The symptom is that
> >>>>>>>> the tap does not have any TX packets. RX is working fine. I see the 
> >>>>>>>> packets coming in at
> >>>>>>>> the physical interface on the host, but they are not forwarded to 
> >>>>>>>> the tap interface.
> >>>>>>>> The bridge itself has learnt the mac address of the vServer that is 
> >>>>>>>> connected to the tap interface.
> >>>>>>>> It does not help to toggle the bridge link status,  the tap 
> >>>>>>>> interface status or the interface in the vServer.
> >>>>>>>> It seems that problem occurs if a tap interface that has previously 
> >>>>>>>> been used, but set to nonpersistent
> >>>>>>>> is set persistent again and then is by chance assigned to the same 
> >>>>>>>> vServer (=same mac address on same
> >>>>>>>> bridge) again. Unfortunately it seems not to be reproducible.
> >>>>>>> 
> >>>>>>> Not sure but this patch from Michael Tsirkin may help - it solves an
> >>>>>>> issue with persistent tap devices:
> >>>>>>> 
> >>>>>>> http://patchwork.ozlabs.org/patch/198598/
> >>>>>> 
> >>>>>> Hi Stefan,
> >>>>>> 
> >>>>>> thanks for the pointer. I have seen this patch, but I have neglected 
> >>>>>> it because it was dealing
> >>>>>> with persistent taps. But maybe the taps in the kernel are not deleted 
> >>>>>> directly.
> >>>>>> Can you remember what the syptomps of the above issue have been? Sorry 
> >>>>>> for
> >>>>>> being vague, but I currently have no clue whats going on.
> >>>>>> 
> >>>>>> Can someone who has more internal knowledge of the bridging/tap code 
> >>>>>> say if qemu can
> >>>>>> be responsible at all if the tap device is not receiving packets from 
> >>>>>> the bridge.
> >>>>>> 
> >>>>>> If I have the following config. Lets say packets coming in via 
> >>>>>> physical interface eth1.123,
> >>>>>> and a bridge called br123.I further have a virtual machine with tap0. 
> >>>>>> Both eth1.123
> >>>>>> and tap0 are member of br123.
> >>>>>> 
> >>>>>> If the issue occurs the vServer has no network connectivity inbound. 
> >>>>>> If I sent a ping
> >>>>>> from the vServer I see it on tap0 and leaving on eth1.123. I see 
> >>>>>> further the arp reply coming
> >>>>>> in via eth1.123, but the reply can't be seen on tap0.
> >>>>>> 
> >>>>>> Peter
> >>>>> 
> >>>>> If guest is not consuming packets, a TX queue in tap device
> >>>>> will with time overrun (there's space for 1000 packets there).
> >>>>> This is code from tun:
> >>>>> 
> >>>>>        if (skb_queue_len(&tfile->socket.sk->sk_receive_queue)
> >>>>>> = dev->tx_queue_len / tun->numqueues){
> >>>>>                if (!(tun->flags & TUN_ONE_QUEUE)) {
> >>>>>                        /* Normal queueing mode. */
> >>>>>                        /* Packet scheduler handles dropping of further
> >>>>> * packets. */
> >>>>>                        netif_stop_subqueue(dev, txq);
> >>>>> 
> >>>>>                        /* We won't see all dropped packets
> >>>>> * individually, so overrun
> >>>>>                         * error is more appropriate. */
> >>>>>                        dev->stats.tx_fifo_errors++;
> >>>>> 
> >>>>> 
> >>>>> So you can detect that this triggered by looking at fifo errors counter 
> >>>>> in device.
> >>>>> 
> >>>>> Once this happens TX queue is stopped, then you hit this path:
> >>>>> 
> >>>>>                        if (!netif_xmit_stopped(txq)) {
> >>>>>                                __this_cpu_inc(xmit_recursion);
> >>>>>                                rc = dev_hard_start_xmit(skb, dev, txq);
> >>>>>                                __this_cpu_dec(xmit_recursion);
> >>>>>                                if (dev_xmit_complete(rc)) {
> >>>>>                                        HARD_TX_UNLOCK(dev, txq);
> >>>>>                                        goto out;
> >>>>>                                }
> >>>>>                        }
> >>>>> 
> >>>>> so packets are not passed to device anymore.
> >>>>> It will stay this way until guest consumes some packets and
> >>>>> queue is restarted.
> >>>> 
> >>>> After some time I again have a vServer in this state. It seems not like 
> >>>> there
> >>>> are no TX errors.
> >>>> 
> >>>> # ifconfig tap10
> >>>> tap10     Link encap:Ethernet  HWaddr 7a:59:20:6f:e7:e5
> >>>>          inet6 addr: fe80::7859:20ff:fe6f:e7e5/64 Scope:Link
> >>>>          UP BROADCAST RUNNING PROMISC MULTICAST  MTU:1500  Metric:1
> >>>>          RX packets:197431 errors:0 dropped:0 overruns:0 frame:0
> >>>>          TX packets:264309 errors:0 dropped:0 overruns:2 carrier:0
> >>>>          collisions:0 txqueuelen:500
> >>>>          RX bytes:13842063 (13.8 MB)  TX bytes:35092821 (35.0 MB)
> >>>> 
> >>>> It seems like the bridge is not forwarding any packets to the tap device 
> >>>> anymore altough it has learnt
> >>>> the MAC-Adresses and there are also broadcast packets coming in.
> >>>> 
> >>>> Any more ideas where I could debug?
> >>>> 
> >>>> Peter
> >>>> 
> >>>>> 
> >>>>>>> 
> >>>>>>> Stefan
> >>> 
> >>> Hmm. So there are two overrun errors that triggered, so
> >>> it's possible after the second one the queue got stuck in an xoff state.
> >>> You'd have to use something like systemtap or kdb to poke at the
> >>> queue state to see whether xoff flag is set and/or look
> >>> at the receive queue length.
> >>> 
> >>> For future, we can try to set TUN_ONE_QUEUE flag on the interface,
> >>> or try applying this patch
> >>> 5d097109257c03a71845729f8db6b5770c4bbedc
> >>> in kernel see if this helps.
> >>> 
> >> 
> >> If have set this option for 2 weeks now and not seen this problem again.
> >> How does this flag work with the recently added tap multiqueue support?
> >> 
> >> Peter
> > 
> > This will be the only option in 3.8.
> 
> Ok, but wouldn`t it be good to set it in qemu for kernels <3.8?
> 
> Peter

Yes, probably a good idea. Patch?

> > 
> > -- 
> > MST