Lorenzo Bianconi <lorenzo@xxxxxxxxxx> writes:
>> Lorenzo Bianconi <lorenzo@xxxxxxxxxx> writes:
>>
>> >> On Thu, 4 Nov 2021 18:35:32 +0100 Lorenzo Bianconi wrote:
>> >> > This change adds support for tail growing and shrinking for XDP multi-buff.
>> >> >
>> >> > When called on a multi-buffer packet with a grow request, it will always
>> >> > work on the last fragment of the packet. So the maximum grow size is the
>> >> > last fragments tailroom, i.e. no new buffer will be allocated.
>> >> >
>> >> > When shrinking, it will work from the last fragment, all the way down to
>> >> > the base buffer depending on the shrinking size. It's important to mention
>> >> > that once you shrink down the fragment(s) are freed, so you can not grow
>> >> > again to the original size.
>> >>
>> >> > +static int bpf_xdp_mb_increase_tail(struct xdp_buff *xdp, int offset)
>> >> > +{
>> >> > + struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
>> >> > + skb_frag_t *frag = &sinfo->frags[sinfo->nr_frags - 1];
>> >> > + int size, tailroom;
>> >> > +
>> >> > + tailroom = xdp->frame_sz - skb_frag_size(frag) - skb_frag_off(frag);
>> >>
>> >> I know I complained about this before but the assumption that we can
>> >> use all the space up to xdp->frame_sz makes me uneasy.
>> >>
>> >> Drivers may not expect the idea that core may decide to extend the
>> >> last frag.. I don't think the skb path would ever do this.
>> >>
>> >> How do you feel about any of these options:
>> >> - dropping this part for now (return an error for increase)
>> >> - making this an rxq flag or reading the "reserved frag size"
>> >> from rxq (so that drivers explicitly opt-in)
>> >> - adding a test that can be run on real NICs
>> >> ?
>> >
>> > I think this has been added to be symmetric with bpf_xdp_adjust_tail().
>> > I do think there is a real use-case for it so far so I am fine to just
>> > support the shrink part.
>> >
>> > @Eelco, Jesper, Toke: any comments on it?
>>
>> Well, tail adjust is useful for things like encapsulations that need to
>> add a trailer. Don't see why that wouldn't be something people would
>> want to do for jumboframes as well?
>>
>
> I agree this would be useful for protocols that add a trailer.
>
>> Not sure I get what the issue is with this either? But having a test
>> that can be run to validate this on hardware would be great in any case,
>> I suppose - we've been discussing more general "compliance tests" for
>> XDP before...
>
> what about option 2? We can add a frag_size field to rxq [0] that is set by
> the driver initializing the xdp_buff. frag_size set to 0 means we can use
> all the buffer.
>
> Regards,
> Lorenzo
>
> [0] pahole -C xdp_rxq_info vmlinux
> struct xdp_rxq_info {
> struct net_device * dev; /* 0 8 */
> u32 queue_index; /* 8 4 */
> u32 reg_state; /* 12 4 */
> struct xdp_mem_info mem; /* 16 8 */
> unsigned int napi_id; /* 24 4 */
>
> /* size: 64, cachelines: 1, members: 5 */
> /* padding: 36 */
> } __attribute__((__aligned__(64)));
Works for me :)
-Toke
