@@ -769,13 +769,17 @@ static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
break;
/* Nothing new? Wait for eventfd to tell us they refilled. */
if (head == vq->num) {
+ /* If interrupted while doing busy polling, requeue
+ * the handler to be fair handle_rx as well as other
+ * tasks waiting on cpu
+ */
if (unlikely(busyloop_intr)) {
vhost_poll_queue(&vq->poll);
- } else if (unlikely(vhost_enable_notify(&net->dev,
- vq))) {
- vhost_disable_notify(&net->dev, vq);
- continue;
}
+ /* Kicks are disabled at this point, break loop and
+ * process any remaining batched packets. Queue will
+ * be re-enabled afterwards.
+ */
break;
}
@@ -825,7 +829,14 @@ static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
++nvq->done_idx;
} while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
+ /* Kicks are still disabled, dispatch any remaining batched msgs. */
vhost_tx_batch(net, nvq, sock, &msg);
+
+ /* All of our work has been completed; however, before leaving the
+ * TX handler, do one last check for work, and requeue handler if
+ * necessary. If there is no work, queue will be reenabled.
+ */
+ vhost_net_busy_poll_try_queue(net, vq);
}
static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
In handle_tx_copy, TX batching processes packets below ~PAGE_SIZE and batches up to 64 messages before calling sock->sendmsg. Currently, when there are no more messages on the ring to dequeue, handle_tx_copy re-enables kicks on the ring *before* firing off the batch sendmsg. However, sock->sendmsg incurs a non-zero delay, especially if it needs to wake up a thread (e.g., another vhost worker). If the guest submits additional messages immediately after the last ring check and disablement, it triggers an EPT_MISCONFIG vmexit to attempt to kick the vhost worker. This may happen while the worker is still processing the sendmsg, leading to wasteful exit(s). This is particularly problematic for single-threaded guest submission threads, as they must exit, wait for the exit to be processed (potentially involving a TTWU), and then resume. In scenarios like a constant stream of UDP messages, this results in a sawtooth pattern where the submitter frequently vmexits, and the vhost-net worker alternates between sleeping and waking. A common solution is to configure vhost-net busy polling via userspace (e.g., qemu poll-us). However, treating the sendmsg as the "busy" period by keeping kicks disabled during the final sendmsg and performing one additional ring check afterward provides a significant performance improvement without any excess busy poll cycles. If messages are found in the ring after the final sendmsg, requeue the TX handler. This ensures fairness for the RX handler and allows vhost_run_work_list to cond_resched() as needed. Test Case TX VM: taskset -c 2 iperf3 -c rx-ip-here -t 60 -p 5200 -b 0 -u -i 5 RX VM: taskset -c 2 iperf3 -s -p 5200 -D 6.12.0, each worker backed by tun interface with IFF_NAPI setup. Note: TCP side is largely unchanged as that was copy bound 6.12.0 unpatched EPT_MISCONFIG/second: 5411 Datagrams/second: ~382k Interval Transfer Bitrate Lost/Total Datagrams 0.00-30.00 sec 15.5 GBytes 4.43 Gbits/sec 0/11481630 (0%) sender 6.12.0 patched EPT_MISCONFIG/second: 58 (~93x reduction) Datagrams/second: ~650k (~1.7x increase) Interval Transfer Bitrate Lost/Total Datagrams 0.00-30.00 sec 26.4 GBytes 7.55 Gbits/sec 0/19554720 (0%) sender Signed-off-by: Jon Kohler <jon@nutanix.com> --- drivers/vhost/net.c | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-)