vhost_net: remove tx polling state
authorJason Wang <jasowang@redhat.com>
Wed, 10 Apr 2013 20:50:48 +0000 (20:50 +0000)
committerDavid S. Miller <davem@davemloft.net>
Thu, 11 Apr 2013 20:16:22 +0000 (16:16 -0400)
After commit 2b8b328b61c799957a456a5a8dab8cc7dea68575 (vhost_net: handle polling
errors when setting backend), we in fact track the polling state through
poll->wqh, so there's no need to duplicate the work with an extra
vhost_net_polling_state. So this patch removes this and make the code simpler.

This patch also removes the all tx starting/stopping code in tx path according
to Michael's suggestion.

Netperf test shows almost the same result in stream test, but gets improvements
on TCP_RR tests (both zerocopy or copy) especially on low load cases.

Tested between multiqueue kvm guest and external host with two direct
connected 82599s.

zerocopy disabled:

sessions|transaction rates|normalize|
before/after/+improvements
1 | 9510.24/11727.29/+23.3%    | 693.54/887.68/+28.0%   |
25| 192931.50/241729.87/+25.3% | 2376.80/2771.70/+16.6% |
50| 277634.64/291905.76/+5%    | 3118.36/3230.11/+3.6%  |

zerocopy enabled:

sessions|transaction rates|normalize|
before/after/+improvements
1 | 7318.33/11929.76/+63.0%    | 521.86/843.30/+61.6%   |
25| 167264.88/242422.15/+44.9% | 2181.60/2788.16/+27.8% |
50| 272181.02/294347.04/+8.1%  | 3071.56/3257.85/+6.1%  |

Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/vhost/net.c
drivers/vhost/vhost.c

index ec6fb3fa59bb5962281bb6277220941b91df053c..87c216c1e54ede98ac69c8bc36d9e9fa9c940170 100644 (file)
@@ -64,20 +64,10 @@ enum {
        VHOST_NET_VQ_MAX = 2,
 };
 
-enum vhost_net_poll_state {
-       VHOST_NET_POLL_DISABLED = 0,
-       VHOST_NET_POLL_STARTED = 1,
-       VHOST_NET_POLL_STOPPED = 2,
-};
-
 struct vhost_net {
        struct vhost_dev dev;
        struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
        struct vhost_poll poll[VHOST_NET_VQ_MAX];
-       /* Tells us whether we are polling a socket for TX.
-        * We only do this when socket buffer fills up.
-        * Protected by tx vq lock. */
-       enum vhost_net_poll_state tx_poll_state;
        /* Number of TX recently submitted.
         * Protected by tx vq lock. */
        unsigned tx_packets;
@@ -155,28 +145,6 @@ static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
        }
 }
 
-/* Caller must have TX VQ lock */
-static void tx_poll_stop(struct vhost_net *net)
-{
-       if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
-               return;
-       vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
-       net->tx_poll_state = VHOST_NET_POLL_STOPPED;
-}
-
-/* Caller must have TX VQ lock */
-static int tx_poll_start(struct vhost_net *net, struct socket *sock)
-{
-       int ret;
-
-       if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
-               return 0;
-       ret = vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
-       if (!ret)
-               net->tx_poll_state = VHOST_NET_POLL_STARTED;
-       return ret;
-}
-
 /* In case of DMA done not in order in lower device driver for some reason.
  * upend_idx is used to track end of used idx, done_idx is used to track head
  * of used idx. Once lower device DMA done contiguously, we will signal KVM
@@ -242,7 +210,7 @@ static void handle_tx(struct vhost_net *net)
                .msg_flags = MSG_DONTWAIT,
        };
        size_t len, total_len = 0;
-       int err, wmem;
+       int err;
        size_t hdr_size;
        struct socket *sock;
        struct vhost_ubuf_ref *uninitialized_var(ubufs);
@@ -253,19 +221,9 @@ static void handle_tx(struct vhost_net *net)
        if (!sock)
                return;
 
-       wmem = atomic_read(&sock->sk->sk_wmem_alloc);
-       if (wmem >= sock->sk->sk_sndbuf) {
-               mutex_lock(&vq->mutex);
-               tx_poll_start(net, sock);
-               mutex_unlock(&vq->mutex);
-               return;
-       }
-
        mutex_lock(&vq->mutex);
        vhost_disable_notify(&net->dev, vq);
 
-       if (wmem < sock->sk->sk_sndbuf / 2)
-               tx_poll_stop(net);
        hdr_size = vq->vhost_hlen;
        zcopy = vq->ubufs;
 
@@ -285,23 +243,14 @@ static void handle_tx(struct vhost_net *net)
                if (head == vq->num) {
                        int num_pends;
 
-                       wmem = atomic_read(&sock->sk->sk_wmem_alloc);
-                       if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
-                               tx_poll_start(net, sock);
-                               set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
-                               break;
-                       }
                        /* If more outstanding DMAs, queue the work.
                         * Handle upend_idx wrap around
                         */
                        num_pends = likely(vq->upend_idx >= vq->done_idx) ?
                                    (vq->upend_idx - vq->done_idx) :
                                    (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
-                       if (unlikely(num_pends > VHOST_MAX_PEND)) {
-                               tx_poll_start(net, sock);
-                               set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
+                       if (unlikely(num_pends > VHOST_MAX_PEND))
                                break;
-                       }
                        if (unlikely(vhost_enable_notify(&net->dev, vq))) {
                                vhost_disable_notify(&net->dev, vq);
                                continue;
@@ -364,8 +313,6 @@ static void handle_tx(struct vhost_net *net)
                                        UIO_MAXIOV;
                        }
                        vhost_discard_vq_desc(vq, 1);
-                       if (err == -EAGAIN || err == -ENOBUFS)
-                               tx_poll_start(net, sock);
                        break;
                }
                if (err != len)
@@ -628,7 +575,6 @@ static int vhost_net_open(struct inode *inode, struct file *f)
 
        vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
        vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
-       n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 
        f->private_data = n;
 
@@ -638,32 +584,24 @@ static int vhost_net_open(struct inode *inode, struct file *f)
 static void vhost_net_disable_vq(struct vhost_net *n,
                                 struct vhost_virtqueue *vq)
 {
+       struct vhost_poll *poll = n->poll + (vq - n->vqs);
        if (!vq->private_data)
                return;
-       if (vq == n->vqs + VHOST_NET_VQ_TX) {
-               tx_poll_stop(n);
-               n->tx_poll_state = VHOST_NET_POLL_DISABLED;
-       } else
-               vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
+       vhost_poll_stop(poll);
 }
 
 static int vhost_net_enable_vq(struct vhost_net *n,
                                struct vhost_virtqueue *vq)
 {
+       struct vhost_poll *poll = n->poll + (vq - n->vqs);
        struct socket *sock;
-       int ret;
 
        sock = rcu_dereference_protected(vq->private_data,
                                         lockdep_is_held(&vq->mutex));
        if (!sock)
                return 0;
-       if (vq == n->vqs + VHOST_NET_VQ_TX) {
-               n->tx_poll_state = VHOST_NET_POLL_STOPPED;
-               ret = tx_poll_start(n, sock);
-       } else
-               ret = vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 
-       return ret;
+       return vhost_poll_start(poll, sock->file);
 }
 
 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
index 9759249e6d908867cf72f53e5bb8e4ecc30db8ae..4eecdb867d53187179aec514072e756e17397719 100644 (file)
@@ -89,6 +89,9 @@ int vhost_poll_start(struct vhost_poll *poll, struct file *file)
        unsigned long mask;
        int ret = 0;
 
+       if (poll->wqh)
+               return 0;
+
        mask = file->f_op->poll(file, &poll->table);
        if (mask)
                vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);