Re: [PATCH dpdk v2 03/16] net: add structure for ipv6 addresses

["Robin Jarry" <rjarry@redhat.com>]

Morten Brørup, Oct 06, 2024 at 10:18:
> This has been discussed before, but I want to double check...
>
> If - sometime in the future - we want to add a union to offer a 2-byte 
> access variant and make the structure to 2-byte aligned (which is the 
> common case in Ethernet packets), it will break both API and ABI. This 
> seems unlikely to get accepted at a later time, so I think we are 
> - right now - in a situation where it's now or never:
>
> struct rte_ipv6_addr {
> 	__extension__
> 	union {
> 		unsigned char a[RTE_IPV6_ADDR_SIZE];
> 		uint16_t      w[RTE_IPV6_ADDR_SIZE / 2];
> 	};
> };
>
> Unless some of the CPU folks want the 2-byte aligned variant, stick 
> with what you offered.

I was also thinking the same. I could have added an unnamed union which 
only contains one field.

However, it does not make much sense if we never want to change the 
default alignment.

Important note: DPDK is compiled with the following C flags:

  -Wno-address-of-packed-member

Added in 2017 https://git.dpdk.org/dpdk/commit/?id=95cd37070af44

If we had struct rte_ipv6_addr aligned on 2 bytes (or more), 
applications that do not silence that warning would have a hard time. 
Example in grout:

../modules/ip6/datapath/ip6_input.c:99:54: error: taking address of packed member of ‘struct rte_ipv6_hdr’ may result in an unaligned pointer value [-Werror=address-of-packed-member]
   99 |                 nh = ip6_route_lookup(iface->vrf_id, &ip->dst_addr);
      |                                                      ^~~~~~~~~~~~~

I'd say that it is better to keep it simple :)