Re: [PATCH net-next v2 2/3] vsock/virtio: add SIOCOUTQ support for all virtio based transports
From: Stefano Garzarella <sgarzare@redhat.com>
Date: 2024-05-07 10:12:07
Also in:
kvm, virtualization
Subsystem:
networking [general], the rest, virtio and vhost vsock driver, virtio core, vm sockets (af_vsock) · Maintainers:
"David S. Miller", Eric Dumazet, Jakub Kicinski, Paolo Abeni, Linus Torvalds, Stefan Hajnoczi, Stefano Garzarella, "Michael S. Tsirkin", Jason Wang
On Thu, Apr 11, 2024 at 09:09:49AM GMT, Paolo Abeni wrote:
On Mon, 2024-04-08 at 15:37 +0200, Luigi Leonardi wrote:quoted
This patch introduce support for stream_bytes_unsent and seqpacket_bytes_unsent ioctl for virtio_transport, vhost_vsock and vsock_loopback. For all transports the unsent bytes counter is incremented in virtio_transport_send_pkt_info. In the virtio_transport (G2H) the counter is decremented each time the host notifies the guest that it consumed the skbuffs. In vhost-vsock (H2G) the counter is decremented after the skbuff is queued in the virtqueue. In vsock_loopback the counter is decremented after the skbuff is dequeued. Signed-off-by: Luigi Leonardi <redacted>I think this deserve an explicit ack from Stefano, and Stefano can't review patches in the next few weeks. If it's not urgent this will have to wait a bit.quoted
--- drivers/vhost/vsock.c | 4 ++- include/linux/virtio_vsock.h | 7 ++++++ net/vmw_vsock/virtio_transport.c | 4 ++- net/vmw_vsock/virtio_transport_common.c | 33 +++++++++++++++++++++++++ net/vmw_vsock/vsock_loopback.c | 7 ++++++ 5 files changed, 53 insertions(+), 2 deletions(-)diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c index ec20ecff85c7..dba8b3ea37bf 100644 --- a/drivers/vhost/vsock.c +++ b/drivers/vhost/vsock.c@@ -244,7 +244,7 @@ vhost_transport_do_send_pkt(struct vhost_vsock *vsock, restart_tx = true; } - consume_skb(skb); + virtio_transport_consume_skb_sent(skb, true); } } while(likely(!vhost_exceeds_weight(vq, ++pkts, total_len))); if (added)@@ -451,6 +451,8 @@ static struct virtio_transport vhost_transport = { .notify_buffer_size = virtio_transport_notify_buffer_size, .notify_set_rcvlowat = virtio_transport_notify_set_rcvlowat, + .unsent_bytes = virtio_transport_bytes_unsent, + .read_skb = virtio_transport_read_skb, },diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h index c82089dee0c8..dbb22d45d203 100644 --- a/include/linux/virtio_vsock.h +++ b/include/linux/virtio_vsock.h@@ -134,6 +134,8 @@ struct virtio_vsock_sock { u32 peer_fwd_cnt; u32 peer_buf_alloc; + atomic_t bytes_unsent;This will add 2 atomic operations per packet, possibly on contended cachelines. Have you considered leveraging the existing transport-level lock to protect the counter updates?
Good point! Maybe we can handle it together with `tx_cnt` in virtio_transport_get_credit()/virtio_transport_put_credit(). Eventually these are called exactly to count the payload we are sending (`tx_cnt` is a counter that only grows, virtio_transport_put_credit() is called only to return unused credit, so we can't use it directly but always need a new variable like `bytes_unsent`). I mean something like this (untested at all):
diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
index dbb22d45d203..713197c16b7f 100644
--- a/include/linux/virtio_vsock.h
+++ b/include/linux/virtio_vsock.h@@ -133,8 +133,7 @@ struct virtio_vsock_sock { u32 tx_cnt; u32 peer_fwd_cnt; u32 peer_buf_alloc; - - atomic_t bytes_unsent; + u32 bytes_unsent; /* Protected by rx_lock */ u32 fwd_cnt;
diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
index 82a31a13dc32..b1a51db616cf 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c@@ -419,13 +419,6 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk, */ rest_len -= ret; - /* Avoid to perform an atomic_add on 0 bytes. - * This is equivalent to check on VIRTIO_VSOCK_OP_RW - * as is the only packet type with payload. - */ - if (ret) - atomic_add(ret, &vvs->bytes_unsent); - if (WARN_ONCE(ret != skb_len, "'send_pkt()' returns %i, but %zu expected\n", ret, skb_len))
@@ -479,7 +472,10 @@ void virtio_transport_consume_skb_sent(struct sk_buff *skb, bool consume) struct virtio_vsock_sock *vvs; vvs = vs->trans; - atomic_sub(skb->len, &vvs->bytes_unsent); + + spin_lock_bh(&vvs->tx_lock); + vvs->bytes_unsent -= skb->len; + spin_unlock_bh(&vvs->tx_lock); } if (consume)
@@ -499,6 +495,7 @@ u32 virtio_transport_get_credit(struct virtio_vsock_sock *vvs, u32 credit) if (ret > credit) ret = credit; vvs->tx_cnt += ret; + vvs->bytes_unsent += ret; spin_unlock_bh(&vvs->tx_lock); return ret;
@@ -512,6 +509,7 @@ void virtio_transport_put_credit(struct virtio_vsock_sock *vvs, u32 credit) spin_lock_bh(&vvs->tx_lock); vvs->tx_cnt -= credit; + vvs->bytes_unsent -= ret; spin_unlock_bh(&vvs->tx_lock); } EXPORT_SYMBOL_GPL(virtio_transport_put_credit);
@@ -915,7 +913,6 @@ int virtio_transport_do_socket_init(struct vsock_sock *vsk, vsk->buffer_size = VIRTIO_VSOCK_MAX_BUF_SIZE; vvs->buf_alloc = vsk->buffer_size; - atomic_set(&vvs->bytes_unsent, 0); spin_lock_init(&vvs->rx_lock); spin_lock_init(&vvs->tx_lock);
@@ -1118,8 +1115,13 @@ EXPORT_SYMBOL_GPL(virtio_transport_destruct); int virtio_transport_bytes_unsent(struct vsock_sock *vsk) { struct virtio_vsock_sock *vvs = vsk->trans; + int ret; - return atomic_read(&vvs->bytes_unsent); + spin_lock_bh(&vvs->tx_lock); + ret = vvs->bytes_unsent; + spin_unlock_bh(&vvs->tx_lock); + + return ret; } EXPORT_SYMBOL_GPL(virtio_transport_bytes_unsent);
WDYT? Should virtio_transport_bytes_unsent() returns size_t? Thanks, Stefano