[dpdk-dev] [PATCH v2] net: adjust the header length parse size

[Wang, Haiyue]

Hi Stephen,

> -----Original Message-----
> From: Stephen Hemminger <stephen at networkplumber.org>
> Sent: Sunday, September 6, 2020 00:56
> To: Wang, Haiyue <haiyue.wang at intel.com>
> Cc: dev at dpdk.org; Yigit, Ferruh <ferruh.yigit at intel.com>; Olivier Matz <olivier.matz at 6wind.com>
> Subject: Re: [PATCH v2] net: adjust the header length parse size
> 
> On Sat,  5 Sep 2020 11:06:46 +0800
> Haiyue Wang <haiyue.wang at intel.com> wrote:
> 
> > Align to the rte_mbuf's design about Tx header length data size for the
> > header length parse result.
> > 	struct {
> > 		uint64_t   l2_len:7;             /*    88: 0  8 */
> > 		uint64_t   l3_len:9;             /*    88: 7  8 */
> > 		uint64_t   l4_len:8;             /*    88:16  8 */
> > 		uint64_t   tso_segsz:16;         /*    88:24  8 */
> > 		uint64_t   outer_l3_len:9;       /*    88:40  8 */
> > 		uint64_t   outer_l2_len:7;       /*    88:49  8 */
> > 	};
> >
> > Now the IPv6 can support bigger extension header.
> >
> > The below is the structure hole analysis result:
> >
> > Before:
> > struct rte_net_hdr_lens {
> >         uint8_t                    l2_len;               /*     0     1 */
> >         uint8_t                    l3_len;               /*     1     1 */
> >         uint8_t                    l4_len;               /*     2     1 */
> >         uint8_t                    tunnel_len;           /*     3     1 */
> >         uint8_t                    inner_l2_len;         /*     4     1 */
> >         uint8_t                    inner_l3_len;         /*     5     1 */
> >         uint8_t                    inner_l4_len;         /*     6     1 */
> >
> >         /* size: 7, cachelines: 1, members: 7 */
> >         /* last cacheline: 7 bytes */
> > };
> >
> > Now:
> > struct rte_net_hdr_lens {
> >         uint64_t                   l2_len:7;             /*     0: 0  8 */
> >         uint64_t                   l3_len:9;             /*     0: 7  8 */
> >         uint64_t                   l4_len:8;             /*     0:16  8 */
> >         uint64_t                   tunnel_len:8;         /*     0:24  8 */
> >         uint64_t                   inner_l2_len:7;       /*     0:32  8 */
> >         uint64_t                   inner_l3_len:9;       /*     0:39  8 */
> >         uint64_t                   inner_l4_len:8;       /*     0:48  8 */
> >
> >         /* size: 8, cachelines: 1, members: 7 */
> >         /* bit_padding: 8 bits */
> >         /* last cacheline: 8 bytes */
> > };
> >
> > Signed-off-by: Haiyue Wang <haiyue.wang at intel.com>
> 
> Bitfields are slow to access, compiler has to do mask/shift operations.

Yes, you are right, I use rdtsc to trace the rte_net_get_ptype run clock
about 100000 times, bitfields is near 40, but original is about 30.


> And there is no requirement that structure be the same size.
> 
> There is no requirement that fields be ordered the same as
> the protocol header. Also tunnel length might get big.
> Why not:
> 
> struct rte_net_hdr_lens {
> 	uint8_t l2_len;
> 	uint8_t inner_l2_len;
> 	uint16_t l3_len;
> 	uint16_t inner_l3_len;
> 	uint16_t tunnel_len;
> 	uint8_t l4_len;
> 	uint8_t inner_l4_len;
> };
> 

Thanks for your comment, this is better, and in v3. ;-) The run clock is nearly
the same as original type.