> On Wed, May 21, 2025 at 10:06:13AM +0800, Xuewei Niu wrote:
> >> On Mon, May 19, 2025 at 03:06:48PM +0800, Xuewei Niu wrote:
> >> >The virtio_vsock_sock has a new field called bytes_unread as the return
> >> >value of the SIOCINQ ioctl.
> >> >
> >> >Though the rx_bytes exists, we introduce a bytes_unread field to the
> >> >virtio_vsock_sock struct. The reason is that it will not be updated
> >> >until the skbuff is fully consumed, which causes inconsistency.
> >> >
> >> >The byte_unread is increased by the length of the skbuff when skbuff is
> >> >enqueued, and it is decreased when dequeued.
> >> >
> >> >Signed-off-by: Xuewei Niu <niuxuewei....@antgroup.com>
> >> >---
> >> > drivers/vhost/vsock.c | 1 +
> >> > include/linux/virtio_vsock.h | 2 ++
> >> > net/vmw_vsock/virtio_transport.c | 1 +
> >> > net/vmw_vsock/virtio_transport_common.c | 17 +++++++++++++++++
> >> > net/vmw_vsock/vsock_loopback.c | 1 +
> >> > 5 files changed, 22 insertions(+)
> >> >
> >> >diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c
> >> >index 802153e23073..0f20af6e5036 100644
> >> >--- a/drivers/vhost/vsock.c
> >> >+++ b/drivers/vhost/vsock.c
> >> >@@ -452,6 +452,7 @@ static struct virtio_transport vhost_transport = {
> >> > .notify_set_rcvlowat =
> >> > virtio_transport_notify_set_rcvlowat,
> >> >
> >> > .unsent_bytes = virtio_transport_unsent_bytes,
> >> >+ .unread_bytes = virtio_transport_unread_bytes,
> >> >
> >> > .read_skb = virtio_transport_read_skb,
> >> > },
> >> >diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
> >> >index 0387d64e2c66..0a7bd240113a 100644
> >> >--- a/include/linux/virtio_vsock.h
> >> >+++ b/include/linux/virtio_vsock.h
> >> >@@ -142,6 +142,7 @@ struct virtio_vsock_sock {
> >> > u32 buf_alloc;
> >> > struct sk_buff_head rx_queue;
> >> > u32 msg_count;
> >> >+ size_t bytes_unread;
> >>
> >> Can we just use `rx_bytes` field we already have?
> >>
> >> Thanks,
> >> Stefano
> >
> >I perfer not. The `rx_bytes` won't be updated until the skbuff is fully
> >consumed, causing inconsistency issues. If it is acceptable to you, I'll
> >reuse the field instead.
>
> I think here we found a little pre-existing issue that should be related
> also to what Arseniy (CCed) is trying to fix (low_rx_bytes).
>
> We basically have 2 counters:
> - rx_bytes, which we use internally to see if there are bytes to read
> and for sock_rcvlowat
> - fwd_cnt, which we use instead for the credit mechanism and informing
> the other peer whether we have space or not
>
> These are updated with virtio_transport_dec_rx_pkt() and
> virtio_transport_inc_rx_pkt()
>
> As far as I can see, from the beginning, we call
> virtio_transport_dec_rx_pkt() only when we consume the entire packet.
> This makes sense for `fwd_cnt`, because we still have occupied space in
> memory and we don't want to update the credit until we free all the
> space, but I think it makes no sense for `rx_bytes`, which is only used
> internally and should reflect the current situation of bytes to read.
>
> So in my opinion we should fix it this way (untested):
>
> diff --git a/net/vmw_vsock/virtio_transport_common.c
> b/net/vmw_vsock/virtio_transport_common.c
> index 11eae88c60fc..ee70cb114328 100644
> --- a/net/vmw_vsock/virtio_transport_common.c
> +++ b/net/vmw_vsock/virtio_transport_common.c
> @@ -449,10 +449,10 @@ static bool virtio_transport_inc_rx_pkt(struct
> virtio_vsock_sock *vvs,
> }
>
> static void virtio_transport_dec_rx_pkt(struct virtio_vsock_sock *vvs,
> - u32 len)
> + u32 bytes_read, u32 bytes_dequeued)
> {
> - vvs->rx_bytes -= len;
> - vvs->fwd_cnt += len;
> + vvs->rx_bytes -= bytes_read;
> + vvs->fwd_cnt += bytes_dequeued;
> }
>
> void virtio_transport_inc_tx_pkt(struct virtio_vsock_sock *vvs, struct
> sk_buff *skb)
> @@ -581,11 +581,11 @@ virtio_transport_stream_do_dequeue(struct vsock_sock
> *vsk,
> size_t len)
> {
> struct virtio_vsock_sock *vvs = vsk->trans;
> - size_t bytes, total = 0;
> struct sk_buff *skb;
> u32 fwd_cnt_delta;
> bool low_rx_bytes;
> int err = -EFAULT;
> + size_t total = 0;
> u32 free_space;
>
> spin_lock_bh(&vvs->rx_lock);
> @@ -597,6 +597,8 @@ virtio_transport_stream_do_dequeue(struct vsock_sock *vsk,
> }
>
> while (total < len && !skb_queue_empty(&vvs->rx_queue)) {
> + size_t bytes, dequeued = 0;
> +
> skb = skb_peek(&vvs->rx_queue);
>
> bytes = min_t(size_t, len - total,
> @@ -620,12 +622,12 @@ virtio_transport_stream_do_dequeue(struct vsock_sock
> *vsk,
> VIRTIO_VSOCK_SKB_CB(skb)->offset += bytes;
>
> if (skb->len == VIRTIO_VSOCK_SKB_CB(skb)->offset) {
> - u32 pkt_len = le32_to_cpu(virtio_vsock_hdr(skb)->len);
> -
> - virtio_transport_dec_rx_pkt(vvs, pkt_len);
> + dequeued = le32_to_cpu(virtio_vsock_hdr(skb)->len);
> __skb_unlink(skb, &vvs->rx_queue);
> consume_skb(skb);
> }
> +
> + virtio_transport_dec_rx_pkt(vvs, bytes, dequeued);
> }
>
> fwd_cnt_delta = vvs->fwd_cnt - vvs->last_fwd_cnt;
> @@ -782,7 +784,7 @@ static int virtio_transport_seqpacket_do_dequeue(struct
> vsock_sock *vsk,
> msg->msg_flags |= MSG_EOR;
> }
>
> - virtio_transport_dec_rx_pkt(vvs, pkt_len);
> + virtio_transport_dec_rx_pkt(vvs, pkt_len, pkt_len);
> vvs->bytes_unread -= pkt_len;
> kfree_skb(skb);
> }
> @@ -1752,6 +1754,7 @@ int virtio_transport_read_skb(struct vsock_sock *vsk,
> skb_read_actor_t recv_acto
> struct sock *sk = sk_vsock(vsk);
> struct virtio_vsock_hdr *hdr;
> struct sk_buff *skb;
> + u32 pkt_len;
> int off = 0;
> int err;
>
> @@ -1769,7 +1772,8 @@ int virtio_transport_read_skb(struct vsock_sock *vsk,
> skb_read_actor_t recv_acto
> if (le32_to_cpu(hdr->flags) & VIRTIO_VSOCK_SEQ_EOM)
> vvs->msg_count--;
>
> - virtio_transport_dec_rx_pkt(vvs, le32_to_cpu(hdr->len));
> + pkt_len = le32_to_cpu(hdr->len);
> + virtio_transport_dec_rx_pkt(vvs, pkt_len, pkt_len);
> spin_unlock_bh(&vvs->rx_lock);
>
> virtio_transport_send_credit_update(vsk);
>
> @Arseniy WDYT?
> I will test it and send a proper patch.
>
> @Xuewei with that fixed, I think you can use `rx_bytes`, right?
I've seen your patch, and looks good to me. This will greatly simplify the
SIOCINQ ioctl implementation. I'll rework after your patch gets merged.
Thanks,
Xuewei
> Also because you missed for example `virtio_transport_read_skb()` used
> by ebpf (see commit 3543152f2d33 ("vsock: Update rx_bytes on
> read_skb()")).
>
> Thanks,
> Stefano
