[PATCH] vhost: add device op to offload the interrupt kick

[Eelco Chaudron]

This patch adds an operation callback which gets called every time the
library wants to call eventfd_write(). This eventfd_write() call could
result in a system call, which could potentially block the PMD thread.

The callback function can decide whether it's ok to handle the
eventfd_write() now or have the newly introduced function,
rte_vhost_notify_guest(), called at a later time.

This can be used by 3rd party applications, like OVS, to avoid system
calls being called as part of the PMD threads.

Signed-off-by: Eelco Chaudron <echaudro at redhat.com>
---
 lib/vhost/rte_vhost.h |   10 +++++++++-
 lib/vhost/vhost.c     |   21 +++++++++++++++++++++
 lib/vhost/vhost.h     |   43 ++++++++++++++++++++++++++++---------------
 3 files changed, 58 insertions(+), 16 deletions(-)

diff --git a/lib/vhost/rte_vhost.h b/lib/vhost/rte_vhost.h
index 58a5d4be92..af7a394d0f 100644
--- a/lib/vhost/rte_vhost.h
+++ b/lib/vhost/rte_vhost.h
@@ -298,7 +298,13 @@ struct rte_vhost_device_ops {
 	 */
 	void (*guest_notified)(int vid);
 
-	void *reserved[1]; /**< Reserved for future extension */
+	/**
+	 * If this callback is registered, notification to the guest can
+	 * be handled by the front-end calling rte_vhost_notify_guest().
+	 * If it's not handled, 'false' should be returned. This can be used
+	 * to remove the "slow" eventfd_write() syscall from the datapath.
+	 */
+	bool (*guest_notify)(int vid, uint16_t queue_id);
 };
 
 /**
@@ -433,6 +439,8 @@ void rte_vhost_log_used_vring(int vid, uint16_t vring_idx,
 
 int rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable);
 
+void rte_vhost_notify_guest(int vid, uint16_t queue_id);
+
 /**
  * Register vhost driver. path could be different for multiple
  * instance support.
diff --git a/lib/vhost/vhost.c b/lib/vhost/vhost.c
index ef37943817..ee090d78ef 100644
--- a/lib/vhost/vhost.c
+++ b/lib/vhost/vhost.c
@@ -1467,6 +1467,27 @@ rte_vhost_enable_guest_notification(int vid, uint16_t queue_id, int enable)
 	return ret;
 }
 
+void
+rte_vhost_notify_guest(int vid, uint16_t queue_id)
+{
+	struct virtio_net *dev = get_device(vid);
+	struct vhost_virtqueue *vq;
+
+	if (!dev ||  queue_id >= VHOST_MAX_VRING)
+		return;
+
+	vq = dev->virtqueue[queue_id];
+	if (!vq)
+		return;
+
+	rte_spinlock_lock(&vq->access_lock);
+
+	if (vq->callfd >= 0)
+		eventfd_write(vq->callfd, (eventfd_t)1);
+
+	rte_spinlock_unlock(&vq->access_lock);
+}
+
 void
 rte_vhost_log_write(int vid, uint64_t addr, uint64_t len)
 {
diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h
index 8fdab13c70..39ad8260a1 100644
--- a/lib/vhost/vhost.h
+++ b/lib/vhost/vhost.h
@@ -883,6 +883,30 @@ vhost_need_event(uint16_t event_idx, uint16_t new_idx, uint16_t old)
 	return (uint16_t)(new_idx - event_idx - 1) < (uint16_t)(new_idx - old);
 }
 
+static __rte_always_inline void
+vhost_vring_kick_guest(struct virtio_net *dev, struct vhost_virtqueue *vq)
+{
+	if (dev->notify_ops->guest_notify) {
+		uint16_t qid;
+		for (qid = 0;  qid < dev->nr_vring; qid++) {
+			if (dev->virtqueue[qid] == vq) {
+				if (dev->notify_ops->guest_notify(dev->vid,
+								  qid))
+					goto done;
+				break;
+			}
+		}
+	}
+	eventfd_write(vq->callfd, (eventfd_t) 1);
+
+done:
+	if (dev->flags & VIRTIO_DEV_STATS_ENABLED)
+		vq->stats.guest_notifications++;
+	if (dev->notify_ops->guest_notified)
+		dev->notify_ops->guest_notified(dev->vid);
+}
+
+
 static __rte_always_inline void
 vhost_vring_call_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 {
@@ -905,21 +929,13 @@ vhost_vring_call_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 		if ((vhost_need_event(vhost_used_event(vq), new, old) &&
 					(vq->callfd >= 0)) ||
 				unlikely(!signalled_used_valid)) {
-			eventfd_write(vq->callfd, (eventfd_t) 1);
-			if (dev->flags & VIRTIO_DEV_STATS_ENABLED)
-				vq->stats.guest_notifications++;
-			if (dev->notify_ops->guest_notified)
-				dev->notify_ops->guest_notified(dev->vid);
+			vhost_vring_kick_guest(dev, vq);
 		}
 	} else {
 		/* Kick the guest if necessary. */
 		if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 				&& (vq->callfd >= 0)) {
-			eventfd_write(vq->callfd, (eventfd_t)1);
-			if (dev->flags & VIRTIO_DEV_STATS_ENABLED)
-				vq->stats.guest_notifications++;
-			if (dev->notify_ops->guest_notified)
-				dev->notify_ops->guest_notified(dev->vid);
+			vhost_vring_kick_guest(dev, vq);
 		}
 	}
 }
@@ -971,11 +987,8 @@ vhost_vring_call_packed(struct virtio_net *dev, struct vhost_virtqueue *vq)
 	if (vhost_need_event(off, new, old))
 		kick = true;
 kick:
-	if (kick) {
-		eventfd_write(vq->callfd, (eventfd_t)1);
-		if (dev->notify_ops->guest_notified)
-			dev->notify_ops->guest_notified(dev->vid);
-	}
+	if (kick)
+		vhost_vring_kick_guest(dev, vq);
 }
 
 static __rte_always_inline void