Re: [PATCH net] octeontx2-pf: fix FLOW_DIS_IS_FRAGMENT implementation

[Asbjørn Sloth Tønnesen]

Hi Suman,

On 4/12/24 5:34 AM, Suman Ghosh wrote:
> > 		if (match.mask->flags & FLOW_DIS_IS_FRAGMENT) {
> > +			val = match.key->flags & FLOW_DIS_IS_FRAGMENT;
> > 			if (ntohs(flow_spec->etype) == ETH_P_IP) {
> > -				flow_spec->ip_flag = IPV4_FLAG_MORE;
> > +				flow_spec->ip_flag = val ? IPV4_FLAG_MORE : 0;
> [Suman] Do we need this? If user provide the command "tc filter add .... ip_flags nofrags" then the above if check should not be hit right? If we are inside the check then we always want to set IPV4_FLAG_MORE right?

In iproute2, the "frag"/"nofrag" is parsed in flower_parse_matching_flags(),
it sets TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT in TCA_FLOWER_KEY_FLAGS and
TCA_FLOWER_KEY_FLAGS_MASK.

Back in the kernel, in fl_set_key_flags() (net/sched/cls_flower.c) then,
directly translates TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT into FLOW_DIS_IS_FRAGMENT,
while only setting the key bit, if the mask bit is being set.

There are therefore 3 possible cases:

- `tc flower ...` (no ip_flags frag or nofrag)
   (match.key->flags & FLOW_DIS_FIRST_FRAG)  is false
   (match.mask->flags & FLOW_DIS_FIRST_FRAG) is false

- `tc flower ... ip_flags nofrag`
   (match.key->flags & FLOW_DIS_FIRST_FRAG)  is false
   (match.mask->flags & FLOW_DIS_FIRST_FRAG) is true

- `tc flower ... ip_flags frag`
   (match.key->flags & FLOW_DIS_FIRST_FRAG)  is true
   (match.mask->flags & FLOW_DIS_FIRST_FRAG) is true

The `nofrag` case will still have the mask bit set, and hence pass the entry condition.


>> 				flow_mask->ip_flag = IPV4_FLAG_MORE;

Yes, you should always set IPV4_FLAG_MORE in flow_mask, but not always in flow_spec.


--
Best regards
Asbjørn Sloth Tønnesen
Network Engineer
Fiberby - AS42541