virtio-net: Limit number of packets sent per TX flush

If virtio_net_flush_tx() is called with notification disabled, we can
race with the guest, processing packets at the same rate as they
get produced.  The trouble is that this means we have no guaranteed
exit condition from the function and can spend minutes in there.
Currently flush_tx is only called with notification on, which seems
to limit us to one pass through the queue per call.  An upcoming
patch changes this.

Also add an option to set this value on the command line as different
workloads may wish to use different values.  We can't necessarily
support any random value, so this is a developer option: x-txburst=
Usage:

-device virtio-net-pci,x-txburst=64 # 64 packets per tx flush

One pass through the queue (256) seems to be a good default value
for this, balancing latency with throughput.  We use a signed int
for x-txburst because 2^31 packets in a burst would take many, many
minutes to process and it allows us to easily return a negative
value value from virtio_net_flush_tx() to indicate a back-off
or error condition.

Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>

5 files changed, 29 insertions, 6 deletions

diff --git a/hw/s390-virtio-bus.c b/hw/s390-virtio-bus.c
index d5cb24e..092e65f 100644
--- a/hw/s390-virtio-bus.c
+++ b/hw/s390-virtio-bus.c
@@ -330,6 +330,8 @@ static VirtIOS390DeviceInfo s390_virtio_net = {
         DEFINE_NIC_PROPERTIES(VirtIOS390Device, nic),
         DEFINE_PROP_UINT32("x-txtimer", VirtIOS390Device,
                            net.txtimer, TX_TIMER_INTERVAL),
+        DEFINE_PROP_INT32("x-txburst", VirtIOS390Device,
+                          net.txburst, TX_BURST),
         DEFINE_PROP_END_OF_LIST(),
     },
 };
diff --git a/hw/syborg_virtio.c b/hw/syborg_virtio.c
index 5665189..3c3f3b0 100644
--- a/hw/syborg_virtio.c
+++ b/hw/syborg_virtio.c
@@ -298,6 +298,8 @@ static SysBusDeviceInfo syborg_virtio_net_info = {
         DEFINE_VIRTIO_NET_FEATURES(SyborgVirtIOProxy, host_features),
         DEFINE_PROP_UINT32("x-txtimer", SyborgVirtIOProxy,
                            net.txtimer, TX_TIMER_INTERVAL),
+        DEFINE_PROP_INT32("x-txburst", SyborgVirtIOProxy,
+                          net.txburst, TX_BURST),
         DEFINE_PROP_END_OF_LIST(),
     }
 };
diff --git a/hw/virtio-net.c b/hw/virtio-net.c
index d5b03ab..55f3d94 100644
--- a/hw/virtio-net.c
+++ b/hw/virtio-net.c
@@ -37,6 +37,7 @@ typedef struct VirtIONet
     NICState *nic;
     QEMUTimer *tx_timer;
     uint32_t tx_timeout;
+    int32_t tx_burst;
     int tx_timer_active;
     uint32_t has_vnet_hdr;
     uint8_t has_ufo;
@@ -620,7 +621,7 @@ static ssize_t virtio_net_receive(VLANClientState *nc, const uint8_t *buf, size_
     return size;
 }
 
-static void virtio_net_flush_tx(VirtIONet *n, VirtQueue *vq);
+static int32_t virtio_net_flush_tx(VirtIONet *n, VirtQueue *vq);
 
 static void virtio_net_tx_complete(VLANClientState *nc, ssize_t len)
 {
@@ -636,16 +637,18 @@ static void virtio_net_tx_complete(VLANClientState *nc, ssize_t len)
 }
 
 /* TX */
-static void virtio_net_flush_tx(VirtIONet *n, VirtQueue *vq)
+static int32_t virtio_net_flush_tx(VirtIONet *n, VirtQueue *vq)
 {
     VirtQueueElement elem;
+    int32_t num_packets = 0;
 
-    if (!(n->vdev.status & VIRTIO_CONFIG_S_DRIVER_OK))
-        return;
+    if (!(n->vdev.status & VIRTIO_CONFIG_S_DRIVER_OK)) {
+        return num_packets;
+    }
 
     if (n->async_tx.elem.out_num) {
         virtio_queue_set_notification(n->tx_vq, 0);
-        return;
+        return num_packets;
     }
 
     while (virtqueue_pop(vq, &elem)) {
@@ -682,14 +685,19 @@ static void virtio_net_flush_tx(VirtIONet *n, VirtQueue *vq)
             virtio_queue_set_notification(n->tx_vq, 0);
             n->async_tx.elem = elem;
             n->async_tx.len  = len;
-            return;
+            return -EBUSY;
         }
 
         len += ret;
 
         virtqueue_push(vq, &elem, len);
         virtio_notify(&n->vdev, vq);
+
+        if (++num_packets >= n->tx_burst) {
+            break;
+        }
     }
+    return num_packets;
 }
 
 static void virtio_net_handle_tx(VirtIODevice *vdev, VirtQueue *vq)
@@ -934,6 +942,7 @@ VirtIODevice *virtio_net_init(DeviceState *dev, NICConf *conf,
     n->tx_timer = qemu_new_timer(vm_clock, virtio_net_tx_timer, n);
     n->tx_timer_active = 0;
     n->tx_timeout = net->txtimer;
+    n->tx_burst = net->txburst;
     n->mergeable_rx_bufs = 0;
     n->promisc = 1; /* for compatibility */
 
diff --git a/hw/virtio-net.h b/hw/virtio-net.h
index 46a2e1c..a2d1545 100644
--- a/hw/virtio-net.h
+++ b/hw/virtio-net.h
@@ -49,9 +49,17 @@
 
 #define TX_TIMER_INTERVAL 150000 /* 150 us */
 
+/* Limit the number of packets that can be sent via a single flush
+ * of the TX queue.  This gives us a guaranteed exit condition and
+ * ensures fairness in the io path.  256 conveniently matches the
+ * length of the TX queue and shows a good balance of performance
+ * and latency. */
+#define TX_BURST 256
+
 typedef struct virtio_net_conf
 {
     uint32_t txtimer;
+    int32_t txburst;
 } virtio_net_conf;
 
 /* Maximum packet size we can receive from tap device: header + 64k */
diff --git a/hw/virtio-pci.c b/hw/virtio-pci.c
index 1af48e2..3a5b3e6 100644
--- a/hw/virtio-pci.c
+++ b/hw/virtio-pci.c
@@ -693,6 +693,8 @@ static PCIDeviceInfo virtio_info[] = {
             DEFINE_NIC_PROPERTIES(VirtIOPCIProxy, nic),
             DEFINE_PROP_UINT32("x-txtimer", VirtIOPCIProxy,
                                net.txtimer, TX_TIMER_INTERVAL),
+            DEFINE_PROP_INT32("x-txburst", VirtIOPCIProxy,
+                              net.txburst, TX_BURST),
             DEFINE_PROP_END_OF_LIST(),
         },
         .qdev.reset = virtio_pci_reset,