Re: [dpdk-dev] [PATCH v2] net/af_packet: fix vlan_insert corruption

[Bruce Richardson <bruce.richardson@intel.com>]

On Tue, Apr 16, 2019 at 08:03:36AM -0700, Stephen Hemminger wrote:
> On Tue, 16 Apr 2019 10:42:13 +0100
> Bruce Richardson <bruce.richardson@intel.com> wrote:
> 
> > On Tue, Apr 16, 2019 at 10:37:07AM +0100, Ferruh Yigit wrote:
> > > On 4/12/2019 11:08 PM, Stephen Hemminger wrote:  
> > > > On Fri, 12 Apr 2019 17:28:17 +0100
> > > > Ferruh Yigit <ferruh.yigit@intel.com> wrote:
> > > >   
> > > >> On 4/8/2019 5:41 PM, Stephen Hemminger wrote:  
> > > >>> If the af_packet transmit is sending a VLAN packet,
> > > >>> and the transmit path to the kernel os full, then it would
> > > >>> mismanage the outgoing mbuf. The original mbuf would end up
> > > >>> being freed twice, once by AF_PACKET PMD and once by caller.    
> > > >>
> > > >> This comment is valid with your new patch [1] that updates 'rte_vlan_insert()'
> > > >> to duplicate the mbuf, right?
> > > >>
> > > >> That patch looks like won't make the release, so I suggest this one wait that
> > > >> patch, although this is harmless on its own, commit log is misleading.
> > > >>
> > > >> [1]
> > > >> https://patches.dpdk.org/patch/51870/  
> > > > 
> > > > It was always true, even with existing vlan_insert.
> > > > Existing vlan_insert has issues if it ever creates a clone packet.
> > > > Existing vlan_insert can duplicate mbuf through clone
> > > >   
> > > 
> > > Right, existing vlan_insert has same issue on af_packet.
> > > 
> > > But, should vlan_insert try to duplicate the mbuf when it is shared, does it
> > > worth the complexity it brings? And when that support removed this patch won't
> > > be needed.  
> > 
> > I don't think vlan insert or other mbuf manipulation APIs should be
> > checking for shared state or not - that's the job of the app. We could have
> > cases where the user does want to modify a shared mbuf.
> > 
> > Regards,
> > /Bruce
> 
> The vlan_insert code is called on transmit, and there are lots of
> cases where a transmit mbuf might be shared (like TCP stack). And in that
> case inserting the vlan must be non-destructive to the original mbuf.
> 
> Whether you want to push the problem to the driver or do it in the
> library, it is still a problem.

Yes, I agree it's a problem. I'd prefer see it done in the driver than in the
library, since it's higher in the SW stack and has more context information
as to what is safe or not.

/Bruce.