> On Mon, Apr 24, 2023 at 07:58:07PM +0800, Yunsheng Lin wrote:
> > On 2023/4/24 17:17, Lorenzo Bianconi wrote:
> > >> On 2023/4/23 22:20, Lorenzo Bianconi wrote:
> > >>>> On 2023/4/23 2:54, Lorenzo Bianconi wrote:
> > >>>>> struct veth_priv {
> > >>>>> @@ -727,17 +729,20 @@ static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
> > >>>>> goto drop;
> > >>>>>
> > >>>>> /* Allocate skb head */
> > >>>>> - page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
> > >>>>> + page = page_pool_dev_alloc_pages(rq->page_pool);
> > >>>>> if (!page)
> > >>>>> goto drop;
> > >>>>>
> > >>>>> nskb = build_skb(page_address(page), PAGE_SIZE);
> > >>>>
> > >>>> If page pool is used with PP_FLAG_PAGE_FRAG, maybe there is some additional
> > >>>> improvement for the MTU 1500B case, it seem a 4K page is able to hold two skb.
> > >>>> And we can reduce the memory usage too, which is a significant saving if page
> > >>>> size is 64K.
> > >>>
> > >>> please correct if I am wrong but I think the 1500B MTU case does not fit in the
> > >>> half-page buffer size since we need to take into account VETH_XDP_HEADROOM.
> > >>> In particular:
> > >>>
> > >>> - VETH_BUF_SIZE = 2048
> > >>> - VETH_XDP_HEADROOM = 256 + 2 = 258
> > >>
> > >> On some arch the NET_IP_ALIGN is zero.
> > >>
> > >> I suppose XDP_PACKET_HEADROOM are for xdp_frame and data_meta, it seems
> > >> xdp_frame is only 40 bytes for 64 bit arch and max size of metalen is 32
> > >> as xdp_metalen_invalid() suggest, is there any other reason why we need
> > >> 256 bytes here?
> > >
> > > XDP_PACKET_HEADROOM must be greater than (40 + 32)B because you may want push
> > > new data at the beginning of the xdp_buffer/xdp_frame running
> > > bpf_xdp_adjust_head() helper.
> > > I think 256B has been selected for XDP_PACKET_HEADROOM since it is 4 cachelines
> > > (but I can be wrong).
> > > There was a discussion in the past to reduce XDP_PACKET_HEADROOM to 192B but
> > > this is not merged yet and it is not related to this series. We can address
> > > your comments in a follow-up patch when XDP_PACKET_HEADROOM series is merged.
>
> Intel drivers still work just fine at 192 headroom and split the page but
> it makes it problematic for BIG TCP where MAX_SKB_FRAGS from shinfo needs
> to be increased. So it's the tailroom that becomes the bottleneck, not the
> headroom. I believe at some point we will convert our drivers to page_pool
> with full 4k page dedicated for a single frame.
>
> >
> > It worth mentioning that the performance gain in this patch is at the cost of
> > more memory usage, at most of VETH_RING_SIZE(256) + PP_ALLOC_CACHE_SIZE(128)
> > pages is used.
> >
> > IMHO, it seems better to limit the memory usage as much as possible, or provide a
> > way to disable/enable page pool for user.
>
> I think that this argument is valuable due to the purpose that veth can
> serve, IMHO it wouldn't be an argument for a standard PF driver. It would
> be interesting to see how veth would work with PP_FLAG_PAGE_FRAG.
actually I already have a PoC for using page_pool with PP_FLAG_PAGE_FRAG
flag in veth driver but if we do not reduce XDP_PACKET_HEADROOM size there
will not be any difference in the memory footprint since we will need two
fragments (so a full-page) for a 1500B frame. Moreover, we will have lower
performance since we will need to spread the data on two skb buffers
(linear area and frag0 in this case).
Regards,
Lorenzo
>
Attachment:
signature.asc
Description: PGP signature
