Re: [PATCH net-next 3/9] ipv6: fib_rules: Add flow label support

[Ido Schimmel <idosch@nvidia.com>]

On Tue, Dec 17, 2024 at 02:59:19PM +0100, Guillaume Nault wrote:
> On Mon, Dec 16, 2024 at 07:11:55PM +0200, Ido Schimmel wrote:
> > @@ -332,6 +334,9 @@ INDIRECT_CALLABLE_SCOPE int fib6_rule_match(struct fib_rule *rule,
> >  	if (r->dscp && r->dscp != ip6_dscp(fl6->flowlabel))
> >  		return 0;
> >  
> > +	if ((r->flowlabel ^ flowi6_get_flowlabel(fl6)) & r->flowlabel_mask)
> > +		return 0;
> > +
> 
> Personally, I'd find the following form easier to read:
> +	if ((flowi6_get_flowlabel(fl6) & r->flowlabel_mask) != r->flowlabel)
> +		return 0;

The FIB rule code already uses the XOR form for other masked matches
('fwmark' for example), so I used it here to be consistent.

> Does GCC produce better code with the xor form?

No big difference.

Original:

static inline __be32 flowi6_get_flowlabel(const struct flowi6 *fl6)
{
        return fl6->flowlabel & IPV6_FLOWLABEL_MASK;
     b85:       81 e2 00 0f ff ff       and    $0xffff0f00,%edx
        if ((r->flowlabel ^ flowi6_get_flowlabel(fl6)) & r->flowlabel_mask)
     b8b:       33 90 c0 00 00 00       xor    0xc0(%rax),%edx
     b91:       23 90 c4 00 00 00       and    0xc4(%rax),%edx
                return 0;
     b97:       41 b8 00 00 00 00       mov    $0x0,%r8d
        if ((r->flowlabel ^ flowi6_get_flowlabel(fl6)) & r->flowlabel_mask)
     b9d:       0f 85 31 ff ff ff       jne    ad4 <fib6_rule_match+0x34>

Modified:

        if ((flowi6_get_flowlabel(fl6) & r->flowlabel_mask) != r->flowlabel)
     b85:       23 90 c4 00 00 00       and    0xc4(%rax),%edx
                return 0;
     b8b:       45 31 c0                xor    %r8d,%r8d
        if ((flowi6_get_flowlabel(fl6) & r->flowlabel_mask) != r->flowlabel)
     b8e:       81 e2 00 0f ff ff       and    $0xffff0f00,%edx
     b94:       3b 90 c0 00 00 00       cmp    0xc0(%rax),%edx
     b9a:       0f 85 34 ff ff ff       jne    ad4 <fib6_rule_match+0x34>