On Fri, Sep 06, 2019 at 01:55:29PM +0100, Edward Cree wrote: > On 06/09/2019 11:56, Pablo Neira Ayuso wrote: > > On Fri, Sep 06, 2019 at 11:02:18AM +0100, Edward Cree wrote: > >> Still NAK for the same reasons as three versions ago (when it was called > >> "netfilter: payload mangling offload support"), you've never managed to > >> explain why this extra API complexity is useful. (Reducing LOC does not > >> mean you've reduced complexity.) > > Oh well... > > > > Patch 1) Mask is inverted for no reason, just because tc pedit needs > > this in this way. All drivers reverse this mask. > > > > Patch 2) All drivers mask out meaningless fields in the value field. > > To be clear: I have no issue with these two patches; they look fine to me. > (Though I'd like to see some comments on struct flow_action_entry specifying > the semantics of these fields, especially if they're going to differ from > the corresponding fields in struct tc_pedit_key.) OK, I can document this semantics, I need just _time_ to write that documentation. I was expecting this patch description is enough by now until I can get to finish that documentation. > > Patch 3) Without this patchset, offsets are on the 32-bit boundary. > > Drivers need to play with the 32-bit mask to infer what field they are > > supposed to mangle... eg. with 32-bit offset alignment, checking if > > the use want to alter the ttl/hop_limit need for helper structures to > > check the 32-bit mask. Mangling a IPv6 address comes with one single > > action... > > Drivers are still going to need to handle multiple pedit actions, in > case the original rule wanted to mangle two non-consecutive fields. > And you can't just coalesce all consecutive mangles, because if you > mangle two consecutive fields (e.g. UDP sport and dport) the driver > still needs to disentangle that if it works on a 'fields' (rather > than 'u32s') level. This infrastructure is _not_ coalescing two consecutive field, e.g. UDP sport and dport is _not_ coalesced. The coalesce routine does _not_ handle multiple tc pedit ex actions. I'll clarify this in the cover letter in the next patchset round. > So either have the core convert things into named protocol fields > (i.e. "set src IPv6 to 1234::5 and add 1 to UDP sport"), or leave > the current sequence-of-u32-mangles as it is. This in-between "we'll > coalesce things together despite not knowing what fields they are" is > neither fish nor fowl. The model you propose would still need this code for tc pedit to adjust offset/length and coalesce u32 fields.
