Re: [PATCH net-next v11] virtio_net: add page_pool support for buffer allocation

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Mar 23, 2026 at 11:01:31AM -0400, Omar Elghoul wrote:
> Hi,
> 
> I've been testing linux-next (tags later than 03/17) and hit new issues in
> virtio-net on s390x. I bisected the issue, and I found this patch to be the
> first buggy commit.
> 
> The issue seems to only be reproducible when running in Secure Execution.
> Tested in a KVM guest, the virtio-net performance appears greatly reduced,
> and the dmesg output shows many instances of the following error messages.
> 
> Partial relevant logs
> =====================
> [   49.332028] macvtap0: bad gso: type: 0, size: 0, flags 1 tunnel 0 tnl csum 0
> [   74.365668] macvtap0: bad gso: type: 2e, size: 27948, flags 0 tunnel 0 tnl csum 0
> [  403.302168] macvtap0: bad csum: flags: 2, gso_type: 23 rx_tnl_csum 0
> [  403.302271] macvtap0: bad csum: flags: 2, gso_type: e0 rx_tnl_csum 0
> [  403.302279] macvtap0: bad csum: flags: 2, gso_type: e1 rx_tnl_csum 0
> [  403.309492] macvtap0: bad csum: flags: 2, gso_type: 4c rx_tnl_csum 0
> [  403.317029] macvtap0: bad csum: flags: 2, gso_type: e0 rx_tnl_csum 0
> 
> Steps to reproduce
> ==================
> 1. Boot a Linux guest implementing this patch under QEMU/KVM (*) with SE
>    enabled and a virtio-net-ccw device attached.
> 2. Run dmesg. The error message is usually already present at boot time,
>    but if not, it can be reproduced by creating any network traffic.
> 
> (*) This patch was not tested in a non-KVM hypervisor environment.
> 
> I've further confirmed that reverting this patch onto its parent commit
> resolves the issue. Please let me know if you'd like me to test a fix or if
> you would need more information.
> 
> Thanks in advance.
> 
> Best,
> Omar

Well... I am not sure how I missed it. Obvious in hindsight:

static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
                        void *buf, unsigned int len, void **ctx,
                        unsigned int *xdp_xmit,
                        struct virtnet_rq_stats *stats)
{
        struct net_device *dev = vi->dev;
        struct sk_buff *skb;
        u8 flags;
                
        if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
                pr_debug("%s: short packet %i\n", dev->name, len);
                DEV_STATS_INC(dev, rx_length_errors);
                virtnet_rq_free_buf(vi, rq, buf);
                return;
        }
        
        /* About the flags below:
         * 1. Save the flags early, as the XDP program might overwrite them.
         * These flags ensure packets marked as VIRTIO_NET_HDR_F_DATA_VALID
         * stay valid after XDP processing.
         * 2. XDP doesn't work with partially checksummed packets (refer to
         * virtnet_xdp_set()), so packets marked as
         * VIRTIO_NET_HDR_F_NEEDS_CSUM get dropped during XDP processing.
         */
                
        if (vi->mergeable_rx_bufs) {
                flags = ((struct virtio_net_common_hdr *)buf)->hdr.flags;
                skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
                                        stats);
        } else if (vi->big_packets) {
                void *p = page_address((struct page *)buf);
                
                flags = ((struct virtio_net_common_hdr *)p)->hdr.flags;
                skb = receive_big(dev, vi, rq, buf, len, stats);
        } else {
                flags = ((struct virtio_net_common_hdr *)buf)->hdr.flags;
                skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
        }       


So we are reading the header, before dma sync, which is within
receive_mergeable and friends:

static struct sk_buff *receive_mergeable(struct net_device *dev,
                                         struct virtnet_info *vi,
                                         struct receive_queue *rq,
                                         void *buf,
                                         void *ctx,
                                         unsigned int len,
                                         unsigned int *xdp_xmit,
                                         struct virtnet_rq_stats *stats)
{               
        struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
        int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
        struct page *page = virt_to_head_page(buf);
        int offset = buf - page_address(page);
        struct sk_buff *head_skb, *curr_skb;     
        unsigned int truesize = mergeable_ctx_to_truesize(ctx);
        unsigned int headroom = mergeable_ctx_to_headroom(ctx);
                
        head_skb = NULL;
                
        if (rq->use_page_pool_dma)
                page_pool_dma_sync_for_cpu(rq->page_pool, page, offset, len);
                


Just as a test, the below should fix it (compiled only), but the real
fix is more complex since we need to be careful to avoid expensive syncing
twice.


diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 97035b49bae7..57b4f5954bed 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -931,9 +931,19 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 
 static void *virtnet_rq_get_buf(struct receive_queue *rq, u32 *len, void **ctx)
 {
+	void *buf;
+
 	BUG_ON(!rq->page_pool);
 
-	return virtqueue_get_buf_ctx(rq->vq, len, ctx);
+	buf = virtqueue_get_buf_ctx(rq->vq, len, ctx);
+	if (buf && rq->use_page_pool_dma && *len) {
+		struct page *page = virt_to_head_page(buf);
+		int offset = buf - page_address(page);
+
+		page_pool_dma_sync_for_cpu(rq->page_pool, page, offset, *len);
+	}
+
+	return buf;
 }
 
 static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)




-- 
MST