[PATCH V3 net-next 1/1] hv_netvsc: Use the xmit_more skb flag to optimize signaling the host

KY Srinivasan kys at microsoft.com
Wed May 6 19:55:03 UTC 2015



> -----Original Message-----
> From: Eric Dumazet [mailto:eric.dumazet at gmail.com]
> Sent: Wednesday, May 6, 2015 12:18 PM
> To: KY Srinivasan
> Cc: davem at davemloft.net; netdev at vger.kernel.org; linux-
> kernel at vger.kernel.org; devel at linuxdriverproject.org; olaf at aepfle.de;
> apw at canonical.com; jasowang at redhat.com
> Subject: Re: [PATCH V3 net-next 1/1] hv_netvsc: Use the xmit_more skb flag
> to optimize signaling the host
> 
> On Wed, 2015-05-06 at 18:28 +0000, KY Srinivasan wrote:
> 
> > Ah! I too was wondering how we could get into this situation. The condition
> you mention
> > is already handled in the lower level - if the attempt to put the last packet
> on vmbus were to
> > fail because the ring is full, we will notify the host independent of kick_q
> parameter - look
> > at the function vmbus_sendpacket_ctl() in drivers/hv/channel.c
> >
> > I will resend the patch by reverting this to version 2.
> 
> I am not really convinced by what you are saying ;)
> 
> problem is not when you fail to add an additional packet.
> 
> Problem is :
> 
> You queue a packet but not kick because skb->xmit_more = 1, and _then_
> you call netif_tx_stop_queue().
> 
> No further packet will be delivered to your driver, until
> netif_tx_wake_queue() is called again from netvsc_send_completion(),
> possibly milli seconds later.

Ok; I see your point. We stop the q in our driver under two conditions:

1. When the ring is full and we fail to queue. This situation is taken care of already.
2. We fall below a certain free space in the ring buffer - this is the case you are pointing out.

I will address this and resend the patch. Thanks for the review.

Regards,

K. Y
> 
> If qdisc was replaced by another one, no packet will kick again, unless
> other traffic comes.
> 
> Given complexity of your driver, I have no easy way to check you made
> this xmit_more conversion correctly, and one packet can not sit in the
> equivalent of TX ring buffer for unbound time.
> 
> Better add a full explanation into your changelog.
> 
> 



More information about the devel mailing list