Xuan Zhuo
2023-Mar-23 10:59 UTC
[PATCH net-next 2/8] virtio_net: mergeable xdp: introduce mergeable_xdp_prepare
On Thu, 23 Mar 2023 12:45:41 +0800, Yunsheng Lin <linyunsheng at huawei.com> wrote:> On 2023/3/23 9:45, Xuan Zhuo wrote: > > On Wed, 22 Mar 2023 19:52:48 +0800, Yunsheng Lin <linyunsheng at huawei.com> wrote: > >> On 2023/3/22 11:03, Xuan Zhuo wrote: > >>> Separating the logic of preparation for xdp from receive_mergeable. > >>> > >>> The purpose of this is to simplify the logic of execution of XDP. > >>> > >>> The main logic here is that when headroom is insufficient, we need to > >>> allocate a new page and calculate offset. It should be noted that if > >>> there is new page, the variable page will refer to the new page. > >>> > >>> Signed-off-by: Xuan Zhuo <xuanzhuo at linux.alibaba.com> > >>> --- > >>> drivers/net/virtio_net.c | 135 ++++++++++++++++++++++----------------- > >>> 1 file changed, 77 insertions(+), 58 deletions(-) > >>> > >>> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c > >>> index 4d2bf1ce0730..bb426958cdd4 100644 > >>> --- a/drivers/net/virtio_net.c > >>> +++ b/drivers/net/virtio_net.c > >>> @@ -1162,6 +1162,79 @@ static int virtnet_build_xdp_buff_mrg(struct net_device *dev, > >>> return 0; > >>> } > >>> > >>> +static void *mergeable_xdp_prepare(struct virtnet_info *vi, > >>> + struct receive_queue *rq, > >>> + struct bpf_prog *xdp_prog, > >>> + void *ctx, > >>> + unsigned int *frame_sz, > >>> + int *num_buf, > >>> + struct page **page, > >>> + int offset, > >>> + unsigned int *len, > >>> + struct virtio_net_hdr_mrg_rxbuf *hdr) > >> > >> The naming convention seems to be xdp_prepare_mergeable(). > > > > What convention? > > > > > >> > >>> +{ > >>> + unsigned int truesize = mergeable_ctx_to_truesize(ctx); > >>> + unsigned int headroom = mergeable_ctx_to_headroom(ctx); > >>> + struct page *xdp_page; > >>> + unsigned int xdp_room; > >>> + > >>> + /* Transient failure which in theory could occur if > >>> + * in-flight packets from before XDP was enabled reach > >>> + * the receive path after XDP is loaded. > >>> + */ > >>> + if (unlikely(hdr->hdr.gso_type)) > >>> + return NULL; > >>> + > >>> + /* Now XDP core assumes frag size is PAGE_SIZE, but buffers > >>> + * with headroom may add hole in truesize, which > >>> + * make their length exceed PAGE_SIZE. So we disabled the > >>> + * hole mechanism for xdp. See add_recvbuf_mergeable(). > >>> + */ > >>> + *frame_sz = truesize; > >>> + > >>> + /* This happens when headroom is not enough because > >>> + * of the buffer was prefilled before XDP is set. > >>> + * This should only happen for the first several packets. > >>> + * In fact, vq reset can be used here to help us clean up > >>> + * the prefilled buffers, but many existing devices do not > >>> + * support it, and we don't want to bother users who are > >>> + * using xdp normally. > >>> + */ > >>> + if (!xdp_prog->aux->xdp_has_frags && > >>> + (*num_buf > 1 || headroom < virtnet_get_headroom(vi))) { > >>> + /* linearize data for XDP */ > >>> + xdp_page = xdp_linearize_page(rq, num_buf, > >>> + *page, offset, > >>> + VIRTIO_XDP_HEADROOM, > >>> + len); > >>> + > >>> + if (!xdp_page) > >>> + return NULL; > >>> + } else if (unlikely(headroom < virtnet_get_headroom(vi))) { > >>> + xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM + > >>> + sizeof(struct skb_shared_info)); > >>> + if (*len + xdp_room > PAGE_SIZE) > >>> + return NULL; > >>> + > >>> + xdp_page = alloc_page(GFP_ATOMIC); > >>> + if (!xdp_page) > >>> + return NULL; > >>> + > >>> + memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM, > >>> + page_address(*page) + offset, *len); > >> > >> It seems the above 'else if' was not really tested even before this patch, > >> as there is no "--*num_buf" if xdp_linearize_page() is not called, which > >> may causes virtnet_build_xdp_buff_mrg() to comsume one more buffer than > >> expected? > > > > Why do you think so? > > > In first 'if' block, there is a "--*num_buf" before gotoing 'err_xdp' > for virtqueue_get_buf() failure in xdp_linearize_page(). > > But here there is no "--*num_buf" before gotoing 'err_xdp' for > alloc_page() failure.Inside err_xdp, we will get all bufs and free them util num_buf is 0. Thanks.> > So one of them has to be wrong, right? > > > > >> > >> Also, it seems better to split the xdp_linearize_page() to two functions > >> as pskb_expand_head() and __skb_linearize() do, one to expand the headroom, > >> the other one to do the linearizing. > > > > No skb here. > > I means following the semantics of pskb_expand_head() and __skb_linearize(), > not to combine the headroom expanding and linearizing into one function as > xdp_linearize_page() does now if we want a better refoctor result. > > > > > > >>