Re: [PATCH V5 2/6 net-next] netdevice.h: Add zero-copy flag in netdevice

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, May 18, 2011 at 01:40:29PM +0200, Michał Mirosław wrote:
> W dniu 18 maja 2011 13:17 użytkownik Michael S. Tsirkin
> <mst@redhat.com> napisał:
> > On Wed, May 18, 2011 at 01:10:50PM +0200, Michał Mirosław wrote:
> >> 2011/5/18 Michael S. Tsirkin <mst@redhat.com>:
> >> > On Tue, May 17, 2011 at 03:28:38PM -0700, Shirley Ma wrote:
> >> >> On Tue, 2011-05-17 at 23:48 +0200, Michał Mirosław wrote:
> >> >> > 2011/5/17 Shirley Ma <mashirle@us.ibm.com>:
> >> >> > > Hello Michael,
> >> >> > >
> >> >> > > Looks like to use a new flag requires more time/work. I am thinking
> >> >> > > whether we can just use HIGHDMA flag to enable zero-copy in macvtap
> >> >> > to
> >> >> > > avoid the new flag for now since mavctap uses real NICs as lower
> >> >> > device?
> >> >> >
> >> >> > Is there any other restriction besides requiring driver to not recycle
> >> >> > the skb? Are there any drivers that recycle TX skbs?
> >> >
> >> > Not just recycling skbs, keeping reference to any of the pages in the
> >> > skb. Another requirement is to invoke the callback
> >> > in a timely fashion.  For example virtio-net doesn't limit the time until
> >> > that happens (skbs are only freed when some other packet is
> >> > transmitted), so we need to avoid zcopy for such (nested-virt)
> >> > scenarious, right?
> >>
> >> Hmm. But every hardware driver supporting SG will keep reference to
> >> the pages until the packet is sent (or DMA'd to the device). This can
> >> take a long time if hardware queue happens to stall for some reason.
> >
> > That's a fundamental property of zero copy transmit.
> > You can't let the application/guest reuse the memory until
> > no one looks at it anymore.
> >
> >> Is it that you mean keeping a reference after all skbs pointing to the
> >> pages are released?
> > No one should reference the pages after the callback is invoked, yes.
> 
> >> >> Not more other restrictions, skb clone is OK. pskb_expand_head() looks
> >> >> OK to me from code review.
> >> > Hmm. pskb_expand_head calls skb_release_data while keeping
> >> > references to pages. How is that ok? What do I miss?
> >> It's making copy of the skb_shinfo earlier, so the pages refcount
> >> stays the same.
> > Exactly. But the callback is invoked so the guest thinks it's ok to
> > change this memory. If it does a corrupted packet will be sent out.
> 
> Hmm. I tool a quick look at skb_clone(), and it looks like this
> sequence will break this scheme:
> 
> skb2 = skb_clone(skb...);
> kfree_skb(skb) or pskb_expand_head(skb);  /* callback called */
> [use skb2, pages still referenced]
> kfree_skb(skb); /* callback called again */
> This sequence is common in bridge, might be in other places.
> 
> Maybe this ubuf thing should just track clones? This will make it work
> on all devices then.
> 
> Best Regards,
> Michał Mirosław

Well bridge has the problem that packet might get anywhere and it's
really hard to track. Same for tun - it can get queued forever.
veth, loopback are all a problem I think.

IOW we really want to limit this to real physical NICs
which mostly all DTRT. Whitelisting them with a new flag
is likely the most concervative approach, no?

-- 
MST