在 2022/12/27 17:10, Heng Qi 写道:

在 2022/12/27 下午2:46, Jason Wang 写道:

quoted

在 2022/12/20 22:14, Heng Qi 写道:

quoted

Support xdp for multi buffer packets in mergeable mode.

Putting the first buffer as the linear part for xdp_buff,
and the rest of the buffers as non-linear fragments to struct
skb_shared_info in the tailroom belonging to xdp_buff.

Signed-off-by: Heng Qi <redacted>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
---
  drivers/net/virtio_net.c | 78 
++++++++++++++++++++++++++++++++++++++++
  1 file changed, 78 insertions(+)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 08f209d7b0bf..8fc3b1841d92 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -931,6 +931,84 @@ static struct sk_buff *receive_big(struct 

net_device *dev,
      return NULL;
  }
  +/* TODO: build xdp in big mode */
+static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
+                      struct virtnet_info *vi,
+                      struct receive_queue *rq,
+                      struct xdp_buff *xdp,
+                      void *buf,
+                      unsigned int len,
+                      unsigned int frame_sz,
+                      u16 *num_buf,
+                      unsigned int *xdp_frags_truesize,
+                      struct virtnet_rq_stats *stats)
+{
+    unsigned int tailroom = SKB_DATA_ALIGN(sizeof(struct 
skb_shared_info));
+    struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
+    unsigned int truesize, headroom;
+    struct skb_shared_info *shinfo;
+    unsigned int xdp_frags_truesz = 0;
+    unsigned int cur_frag_size;
+    struct page *page;
+    skb_frag_t *frag;
+    int offset;
+    void *ctx;
+
+    xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
+    xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM,
+             VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, 
true);
+
+    if (*num_buf > 1) {
+        shinfo = xdp_get_shared_info_from_buff(xdp);
+        shinfo->nr_frags = 0;
+        shinfo->xdp_frags_size = 0;
+    }
+
+    if ((*num_buf - 1) > MAX_SKB_FRAGS)
+        return -EINVAL;
+
+    while ((--*num_buf) >= 1) {
+        buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
+        if (unlikely(!buf)) {
+            pr_debug("%s: rx error: %d buffers out of %d missing\n",
+                 dev->name, *num_buf,
+                 virtio16_to_cpu(vi->vdev, hdr->num_buffers));
+            dev->stats.rx_length_errors++;
+            return -EINVAL;
+        }
+
+        if (!xdp_buff_has_frags(xdp))
+            xdp_buff_set_frags_flag(xdp);

Any reason to put this inside the loop?

I'll move it outside the loop in the next version.

quoted

quoted

+
+        stats->bytes += len;
+        page = virt_to_head_page(buf);
+        offset = buf - page_address(page);
+        truesize = mergeable_ctx_to_truesize(ctx);
+        headroom = mergeable_ctx_to_headroom(ctx);
+
+        cur_frag_size = truesize + (headroom ? (headroom + 
tailroom) : 0);
+        xdp_frags_truesz += cur_frag_size;

Not related to this patch, but it would easily confuse the future 
readers that the we need another math for truesize. I think at least 
we need some comments for this or

Yes it might, I'll add more comments on this.

quoted

I guess the root cause is in get_mergeable_buf_len:

static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
                                      struct ewma_pkt_len *avg_pkt_len,
                                          unsigned int room)
{
        struct virtnet_info *vi = rq->vq->vdev->priv;
        const size_t hdr_len = vi->hdr_len;
        unsigned int len;

        if (room)
        return PAGE_SIZE - room;

And we do

    len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);

    ...

    ctx = mergeable_len_to_ctx(len, headroom);


I wonder if it's better to pack the real truesize (PAGE_SIZE) here. 
This may ease a lot of things.

I don't know the historical reason for not packing, but I guess this 
is for the convenience of
comparing the actual length len given by the device with truesize. 
Therefore, I think it would
be better to keep the above practice for now?

The problem is:

We had

static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)

It means the truesize should be calculated before packing into the 
context. So having another math to get truesize is very confusing, I 
think it's better to fix that. Otherwise we may end up code that is hard 
to be reviewed even with the help of comments.

Thanks

Perhaps, I can add more explanation to the
above code to help future readers try not to get confused.

Thanks.

quoted

Thanks

quoted

+        if (unlikely(len > truesize || cur_frag_size > PAGE_SIZE)) {
+            pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
+                 dev->name, len, (unsigned long)ctx);
+            dev->stats.rx_length_errors++;
+            return -EINVAL;
+        }
+
+        frag = &shinfo->frags[shinfo->nr_frags++];
+        __skb_frag_set_page(frag, page);
+        skb_frag_off_set(frag, offset);
+        skb_frag_size_set(frag, len);
+        if (page_is_pfmemalloc(page))
+            xdp_buff_set_frag_pfmemalloc(xdp);
+
+        shinfo->xdp_frags_size += len;
+    }
+
+    *xdp_frags_truesize = xdp_frags_truesz;
+    return 0;
+}
+
  static struct sk_buff *receive_mergeable(struct net_device *dev,
                       struct virtnet_info *vi,
                       struct receive_queue *rq,