From mboxrd@z Thu Jan 1 00:00:00 1970 From: Konrad Rzeszutek Wilk Subject: Re: domU to domU networking issues in v3.7? (netserver/netperf failing to communicate) Date: Mon, 12 Nov 2012 09:28:35 -0500 Message-ID: <20121112142835.GG19860@phenom.dumpdata.com> References: <20121110135931.GD23686@localhost.localdomain> <1352714084.27833.137.camel@zakaz.uk.xensource.com> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: Content-Disposition: inline In-Reply-To: <1352714084.27833.137.camel@zakaz.uk.xensource.com> List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xen.org Errors-To: xen-devel-bounces@lists.xen.org To: Ian Campbell Cc: "annie.li@oracle.com" , "marcos.matsunaga@oracle.com" , "xen-devel@lists.xen.org" List-Id: xen-devel@lists.xenproject.org On Mon, Nov 12, 2012 at 09:54:44AM +0000, Ian Campbell wrote: > On Sat, 2012-11-10 at 13:59 +0000, Konrad Rzeszutek Wilk wrote: > > Hey Ian, Xen-devel mailingl list, > > > > I think the issue of 70% traffic lost was actually introduced in v3.6 or > > perhaps v3.5. Annie and Marcos (CC-ed here) are looking to see which of > > the releases introduced this. The issue we are seeing is that a domU > > to domU communication breaks - this is with netperf/netserver talking to > > each other. > > > > Anyhow, I think the 3.7 compound page exacerbated the problem and also > > (at least on some of my test hardware) exposed existing issues with > > drivers. The issue I have is that the 'skge' driver has a bug that has > > been there for ages (I tested way back to 3.0 and still saw it) were it > > can not work with SWIOTLB. It is probably missing an pci_dma_sync > > somewhere. > > > > Anyhow the compound page got me to look at Xen-SWIOTLB and that looks > > OK. Even with synthetic driver (the fake one I posted somewhere) it > > dealt with compound pages properly (with debug or non-debug Xen > > hypervisor). > > The debug build is probably most interesting since it deliberately > allocates a non 1-1 p-to-m mapping so as to catch exactly these sorts of > issues. Right. My test env runs with that. And so far it only has issues with the skge one. > > > So was wondering if you had looked at this in more details? Any > > ideas? Or would it be more prudent to ask that once we know for sure > > which Linux release introduced the communication failures between > > guests? > > I've not looked at it any further I'm afraid. > > If these changes (be they in 3.5 or later, or earlier) are exposing > driver bugs then I suspect the netdev chaps would want to know about it. Right. Annie (CC-ed here) mentioned to me that v3.5 looks to work ok. And is off checking v3.6. v3.7 is definitly a no go. > > FWIW I see the issue with tg3. I should double-check that but I didn't get any issues with the tg3 I had. I was running the netperf/netserver from/to to see if I could reproduce the issues (70% loss) you described but did not have much luck. > > Ian.