Re: [PATCH 0/4 net-next] flow_offload: update mangle action representation

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Sun, Sep 01, 2019 at 01:47:54PM -0700, Jakub Kicinski wrote:
> On Sat, 31 Aug 2019 16:22:17 +0200, Pablo Neira Ayuso wrote:
[...]
> > > Please see the definitions of:
> > > 
> > > struct nfp_fl_set_eth
> > > struct nfp_fl_set_ip4_addrs
> > > struct nfp_fl_set_ip4_ttl_tos
> > > struct nfp_fl_set_ipv6_tc_hl_fl
> > > struct nfp_fl_set_ipv6_addr
> > > struct nfp_fl_set_tport
> > > 
> > > These are the programming primitives for header rewrites in the NFP.
> > > Since each of those contains more than just one field, we'll have to
> > > keep all the field coalescing logic in the driver, even if you coalesce
> > > while fields (i.e. IPv6 addresses).  
> > 
> > nfp has been updated in this patch series to deal with the new mangle
> > representation.
> 
> It has been updated to handle the trivial coalescing.
> 
> > > Perhaps it's not a serious blocker for the series, but it'd be nice if
> > > rewrite action grouping was handled in the core. Since you're already
> > > poking at that code..  
> > 
> > Rewrite action grouping is already handled from the core front-end in
> > this patch series.
> 
> If you did what I'm asking the functions nfp_fl_check_mangle_start()
> and nfp_fl_check_mangle_end() would no longer exist. They were not
> really needed before you "common flow API" changes.

Thanks for the pointer. This driver-level coalescing routine you are
refering to is specific to optimize your layout. I agree the core
could be updated to do more coalescing, but this would need a way to
express what coalescing the driver would like to see in place. I would
wait to see more drivers that can benefit from that. I can only make
incremental steps, it's already hard to navigate over all this code.