> -----Original Message----- > From: Stephen Hemminger [mailto:shemminger@xxxxxxxxxx] > Sent: Monday, March 19, 2012 1:49 PM > To: Eric Dumazet > Cc: Haiyang Zhang; KY Srinivasan; davem@xxxxxxxxxxxxx; > netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; > devel@xxxxxxxxxxxxxxxxxxxxxx > Subject: Re: [PATCH 1/1] net/hyperv: Fix the code handling tx busy > > On Mon, 19 Mar 2012 10:11:58 -0700 > Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote: > > > On Mon, 2012-03-19 at 10:02 -0700, Haiyang Zhang wrote: > > > Instead of dropping the packet, we keep the skb buffer, and return > > > NETDEV_TX_BUSY to let upper layer retry send. This will not cause > > > endless loop, because the host is taking data away from ring buffer. > > > > > > Signed-off-by: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx> > > > Reviewed-by: K. Y. Srinivasan <kys@xxxxxxxxxxxxx> > > > --- > > > drivers/net/hyperv/netvsc_drv.c | 5 +---- > > > 1 files changed, 1 insertions(+), 4 deletions(-) > > > > > > diff --git a/drivers/net/hyperv/netvsc_drv.c > > > b/drivers/net/hyperv/netvsc_drv.c index 2517d20..dd29478 100644 > > > --- a/drivers/net/hyperv/netvsc_drv.c > > > +++ b/drivers/net/hyperv/netvsc_drv.c > > > @@ -223,13 +223,10 @@ static int netvsc_start_xmit(struct sk_buff *skb, > struct net_device *net) > > > net->stats.tx_bytes += skb->len; > > > net->stats.tx_packets++; > > > } else { > > > - /* we are shutting down or bus overloaded, just drop packet > */ > > > - net->stats.tx_dropped++; > > > kfree(packet); > > > - dev_kfree_skb_any(skb); > > > } > > > > > > - return NETDEV_TX_OK; > > > + return ret ? NETDEV_TX_BUSY : NETDEV_TX_OK; > > > } > > > > > > /* > > > > Thats simply not true at all. > > > > A start_xmit() cannot do that. > > > > TX_BUSY should never be returned at all, its a deprecated code, for > > pretty good reasons. (assuming queue is not stopped) > > > > Try this on a machine with one CPU, I am pretty sure this can trigger > > complete freezes. > > > > Once softirq loops in your start_xmit(), how do you think one process > > can help you now ? > > Eric is right, look how devices with real physical rings work. > They test for space left at end of start xmit and stop the transmit queue with > netif_stop_queue. The transmit done code then re-enables when enough > space is netif_wake_queue. Think of it as classic high/low water mark on a > FIFO. As in my previous reply to Eric -- We actually stop queue when the ring buffer is busy, see the code in netvsc.c I have tested with one CPU. After NETDEV_TX_BUSY is returned, the Linux guest OS continues to respond without any problem. Thanks, - Haiyang _______________________________________________ devel mailing list devel@xxxxxxxxxxxxxxxxxxxxxx http://driverdev.linuxdriverproject.org/mailman/listinfo/devel
