[PATCH RFC 0/9] sk_buff: optimize layout for GRO

[Paolo Abeni]

On Mon, 2021-07-26 at 22:51 -0400, Paul Moore wrote:
> On Mon, Jul 26, 2021 at 11:13 AM Casey Schaufler <casey at schaufler-ca.com> wrote:
> > On 7/25/2021 3:52 PM, Florian Westphal wrote:
> > > Casey Schaufler <casey at schaufler-ca.com> wrote:
> > > > RedHat and android use SELinux and will want this. Ubuntu doesn't
> > > > yet, but netfilter in in the AppArmor task list. Tizen definitely
> > > > uses it with Smack. The notion that security modules are only used
> > > > in fringe cases is antiquated.
> > > I was not talking about LSM in general, I was referring to the
> > > extended info that Paul mentioned.
> > > 
> > > If thats indeed going to be used on every distro then skb extensions
> > > are not suitable for this, it would result in extr akmalloc for every
> > > skb.
> > 
> > I am explicitly talking about the use of secmarks. All my
> > references are uses of secmarks.
> 
> I'm talking about a void* which would contain LSM specific data; as I
> said earlier, think of inodes.  This LSM specific data would include
> the existing secmark data as well as network peer security information
> which would finally (!!!) allow us to handle forwarded traffic and
> enable a number of other fixes and performance improvements.
> 
> (The details are a bit beyond this discussion but it basically
> revolves around us not having to investigate the import the packet
> headers every time we want to determine the network peer security
> attributes, we could store the resolved LSM information in the
> sk_buff.security blob.)

I've investigated the feasibility of extending the secmark field to
long/void*. I think that performance wise it should be doable on top of
this series: the amount of allocated memory for sk_buff will not
change, nor the amount of memory memseted at skb initialization time.

I stumbled upon some uAPIs issues, as CT/nft expose a secmark related
field via uAPI, changing that size without breaking esisting user-space 
looks hard to me.

Additionally, even patch 7/9 is problematic, as there are some in
kernel users accessing and using the inner_ field regardless skb-
>encapsulation. That works while inner_* field are always
initializared/zeored, but will break with the mentioned patch. The fix
is doable, but large and complex. 

To keep the scope of this series sane, I'll drop in the next iteration
all the problematic patches - that is: no sk_buff layout change at all.

If there is interest for such thing, it could still be added
incrementally.

Cheers,

Paolo