[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] Timout packets in device's TX queue

Discussing this further with Paul, we came to the conclusion that probably the best solution would be to drop these packets in qdisc. Netback RX path stop accepting new packets if the target guest doesn't have enough room in the ring. Also (AFAIK) NIC drivers do the same if they don't have more resource for TX, and this is all good for us. Our problem is that the queue in qdisc layer (sorry if my terminology not clear!) can still accumulate these packets indefinitely. The drastic measure would be to reduce txqueuelen to 0 during setup for every affected device, but that's not really nice. Instead, we should be able to configure qdisc to timeout packets on those queues, at least the SKBs where (skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY). I'm not that familiar with it to know if that's already possible, or if not, then how good idea would it be to implement it. I've changed the subject and included netdev and Eric, maybe someone can shed some more light on this question.



On 14/11/13 09:42, Paul Durrant wrote:
-----Original Message-----
From: Zoltan Kiss
Sent: 13 November 2013 20:30
To: xen-devel@xxxxxxxxxxxxxxxxxxxx; Ian Campbell; Wei Liu; Paul Durrant;
Malcolm Crossley; David Vrabel
Cc: Jonathan Davies
Subject: netback: Delayed copy alternative


I'm trying to forward port delayed copy to my new grant mapping patches.
One important problem I've faced is that classic used
gnttab_copy_grant_page to replace the granted page with a local copy and
unmap the grant. And this function has never been upstreamed as only
netback used it. Unfortunately upstreaming it is not a very easy task,
as the kernel's grant table infrastructure doesn't track at the moment
whether the page is DMA mapped or not. It is required because we
shouldn't proceed with the copy and replace if a device already mapped
the page for DMA.
David came up with an alternative idea: we do this delayed copy because
we don't want the guest's page to get stucked in Dom0 indefinitely. The
only realistic case for that would be if the egress interface would be
an another guest's vif, where the guest (either due to a bug or as a
malicious attempt) doesn't empty its ring. I think it's a safe
assumption that Dom0 otherwise doesn't hold on to packets for too long.
Or if it does, then that's a bug we should fix instead of doing a copy
of the packet.
If we accept that only other vif's can keep the skb indefinitely, then
an easier solution would be to handle this problem on the RX side: the
RX thread can also check whether this skb hanged around for too long and
drop it. Actually, xenvif_start_xmit already checks if the guest
provided enough slots for us to do the grant copy. If I understand it
correctly. What do you think about such an approach?

Well, now that David fixed the DMA unmap tracking thing, I believe that another 
vif is *generally* the only place an skb can hang around for a long time. The 
problem is that there is an edge case... If a network driver turns off queue 
processing (for flow control reasons, and NB that 10G Ethernet requires the 
driver to do this if the PHY signals flow control and internal buffering is 
exhausted, 1G is allowed to be an open drain) then the skb can sit in the queue 
indefinitely and there's no way you can deal with this from the guest RX side 
of netback. You need to have a copy-aside option to handle this.


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.