Re: [RFC 07/12] nfp: add skb mark support to the bpf offload

[Jakub Kicinski <jakub.kicinski@netronome.com>]

On Thu, 02 Jun 2016 00:30:07 +0200, Daniel Borkmann wrote:
> On 06/02/2016 12:19 AM, Jakub Kicinski wrote:
> > On Wed, 1 Jun 2016 14:56:26 -0700, Alexei Starovoitov wrote:  
> >> On Wed, Jun 01, 2016 at 05:50:09PM +0100, Jakub Kicinski wrote:  
> >>> Skb marking should be set in designated register, FW will
> >>> prepend it to the packet for us.
> >>>
> >>> Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
> >>> Reviewed-by: Dinan Gunawardena <dgunawardena@netronome.com>
> >>> Reviewed-by: Simon Horman <simon.horman@netronome.com>
> >>> ---
> >>>   drivers/net/ethernet/netronome/nfp/nfp_bpf_jit.c    | 20 ++++++++++++++++++++
> >>>   drivers/net/ethernet/netronome/nfp/nfp_net.h        |  2 +-
> >>>   drivers/net/ethernet/netronome/nfp/nfp_net_common.c |  8 +++++++-
> >>>   3 files changed, 28 insertions(+), 2 deletions(-)
> >>>
> >>> diff --git a/drivers/net/ethernet/netronome/nfp/nfp_bpf_jit.c b/drivers/net/ethernet/netronome/nfp/nfp_bpf_jit.c
> >>> index d7eecfceba5c..b31e673a6fe8 100644
> >>> --- a/drivers/net/ethernet/netronome/nfp/nfp_bpf_jit.c
> >>> +++ b/drivers/net/ethernet/netronome/nfp/nfp_bpf_jit.c
> >>> @@ -46,6 +46,8 @@
> >>>
> >>>   #define REG_IMM0_N	30 /* Bank AB */
> >>>   #define REG_QNUM	29 /* Bank AB */
> >>> +#define REG_MARK	28 /* Bank A */
> >>> +#define REG_MARK_STS	28 /* Bank B */
> >>>
> >>>   /* --- NFP prog --- */
> >>>   /* Foreach "multiple" entries macros provide pos and next<n> pointers.
> >>> @@ -416,6 +418,15 @@ static int construct_data_ld(struct nfp_prog *nfp_prog, u16 offset, u8 size)
> >>>   	return construct_data_ind_ld(nfp_prog, offset, 0, false, size);
> >>>   }
> >>>
> >>> +static int wrp_skb_mark(struct nfp_prog *nfp_prog, u16 src)
> >>> +{
> >>> +	__emit_alu(nfp_prog, REG_MARK, ALU_DST_A, REG_NONE, ALU_OP_NONE, src,
> >>> +		   false, false);
> >>> +	__emit_immed(nfp_prog, REG_MARK_STS, ALU_DST_B, 1, false);
> >>> +
> >>> +	return 0;
> >>> +}
> >>> +
> >>>   static int
> >>>   construct_br_imm(struct nfp_prog *nfp_prog, u32 imm, u16 dst, u8 br, u16 off,
> >>>   		 enum alu_op alu_op, bool sw)
> >>> @@ -538,6 +549,14 @@ static int imm_ld8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
> >>>   	return 0;
> >>>   }
> >>>
> >>> +static int mem_stx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
> >>> +{
> >>> +	if (meta->insn.off == offsetof(struct sk_buff, mark))
> >>> +		return wrp_skb_mark(nfp_prog, meta->insn.src_reg * 2);  
> >>
> >> couldn't figure out from the diff or commit log...
> >> what is the meaning of 'skb->mark' for nfp?
> >> Looks like it's writing into magic register and fw will do something
> >> with that register?
> >> 'mark' is packet metadata. Could you explain how it's passing
> >> this metadata? Is it on the wire as well or somehow in the wire
> >> only between two nfps?
> >> Looks like interesting feature.  
> >
> > Oh, it's not a magic register, it just an "API" I have between the BPF
> > and the datapath firmware.  Whatever is put in that register will be
> > prepended to the packet (if the mark status register is set).  
> 
> That is very useful indeed!
> 
> Btw, do you later on plan to also add something similar like TC_ACT_REDIRECT,
> f.e. to push the packet same or different NIC port out again w/o leaving the
> HW?

I don't see any reason why we wouldn't be able to do that :)