vhost: allow userspace to create workers

For vhost-scsi with 3 vqs or more and a workload that tries to use
them in parallel like:

fio --filename=/dev/sdb  --direct=1 --rw=randrw --bs=4k \
--ioengine=libaio --iodepth=128  --numjobs=3

the single vhost worker thread will become a bottlneck and we are stuck
at around 500K IOPs no matter how many jobs, virtqueues, and CPUs are
used.

To better utilize virtqueues and available CPUs, this patch allows
userspace to create workers and bind them to vqs. You can have N workers
per dev and also share N workers with M vqs on that dev.

This patch adds the interface related code and the next patch will hook
vhost-scsi into it. The patches do not try to hook net and vsock into
the interface because:

1. multiple workers don't seem to help vsock. The problem is that with
only 2 virtqueues we never fully use the existing worker when doing
bidirectional tests. This seems to match vhost-scsi where we don't see
the worker as a bottleneck until 3 virtqueues are used.

2. net already has a way to use multiple workers.

Signed-off-by: Mike Christie <michael.christie@oracle.com>
Message-Id: <20230626232307.97930-16-michael.christie@oracle.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>

1 files changed, 141 insertions, 1 deletions

diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index ffbaf7d32e2c..bfba91ecbd0a 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -626,6 +626,76 @@ free_worker:
 	return NULL;
 }
 
+/* Caller must have device mutex */
+static void __vhost_vq_attach_worker(struct vhost_virtqueue *vq,
+				     struct vhost_worker *worker)
+{
+	if (vq->worker)
+		vq->worker->attachment_cnt--;
+	worker->attachment_cnt++;
+	vq->worker = worker;
+}
+
+ /* Caller must have device and virtqueue mutex */
+static int vhost_vq_attach_worker(struct vhost_virtqueue *vq,
+				  struct vhost_vring_worker *info)
+{
+	unsigned long index = info->worker_id;
+	struct vhost_dev *dev = vq->dev;
+	struct vhost_worker *worker;
+
+	if (!dev->use_worker)
+		return -EINVAL;
+
+	/*
+	 * We only allow userspace to set a virtqueue's worker if it's not
+	 * active and polling is not enabled. We also assume drivers
+	 * supporting this will not be internally queueing works directly or
+	 * via calls like vhost_dev_flush at this time.
+	 */
+	if (vhost_vq_get_backend(vq) || vq->kick)
+		return -EBUSY;
+
+	worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
+	if (!worker || worker->id != info->worker_id)
+		return -ENODEV;
+
+	__vhost_vq_attach_worker(vq, worker);
+	return 0;
+}
+
+/* Caller must have device mutex */
+static int vhost_new_worker(struct vhost_dev *dev,
+			    struct vhost_worker_state *info)
+{
+	struct vhost_worker *worker;
+
+	worker = vhost_worker_create(dev);
+	if (!worker)
+		return -ENOMEM;
+
+	info->worker_id = worker->id;
+	return 0;
+}
+
+/* Caller must have device mutex */
+static int vhost_free_worker(struct vhost_dev *dev,
+			     struct vhost_worker_state *info)
+{
+	unsigned long index = info->worker_id;
+	struct vhost_worker *worker;
+
+	worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
+	if (!worker || worker->id != info->worker_id)
+		return -ENODEV;
+
+	if (worker->attachment_cnt)
+		return -EBUSY;
+
+	vhost_worker_destroy(dev, worker);
+	return 0;
+}
+
 static int vhost_get_vq_from_user(struct vhost_dev *dev, void __user *argp,
 				  struct vhost_virtqueue **vq, u32 *id)
 {
@@ -647,6 +717,76 @@ static int vhost_get_vq_from_user(struct vhost_dev *dev, void __user *argp,
 	return 0;
 }
 
+/* Caller must have device mutex */
+long vhost_worker_ioctl(struct vhost_dev *dev, unsigned int ioctl,
+			void __user *argp)
+{
+	struct vhost_vring_worker ring_worker;
+	struct vhost_worker_state state;
+	struct vhost_virtqueue *vq;
+	long ret;
+	u32 idx;
+
+	if (!dev->use_worker)
+		return -EINVAL;
+
+	if (!vhost_dev_has_owner(dev))
+		return -EINVAL;
+
+	ret = vhost_dev_check_owner(dev);
+	if (ret)
+		return ret;
+
+	switch (ioctl) {
+	/* dev worker ioctls */
+	case VHOST_NEW_WORKER:
+		ret = vhost_new_worker(dev, &state);
+		if (!ret && copy_to_user(argp, &state, sizeof(state)))
+			ret = -EFAULT;
+		return ret;
+	case VHOST_FREE_WORKER:
+		if (copy_from_user(&state, argp, sizeof(state)))
+			return -EFAULT;
+		return vhost_free_worker(dev, &state);
+	/* vring worker ioctls */
+	case VHOST_ATTACH_VRING_WORKER:
+	case VHOST_GET_VRING_WORKER:
+		break;
+	default:
+		return -ENOIOCTLCMD;
+	}
+
+	ret = vhost_get_vq_from_user(dev, argp, &vq, &idx);
+	if (ret)
+		return ret;
+
+	mutex_lock(&vq->mutex);
+	switch (ioctl) {
+	case VHOST_ATTACH_VRING_WORKER:
+		if (copy_from_user(&ring_worker, argp, sizeof(ring_worker))) {
+			ret = -EFAULT;
+			break;
+		}
+
+		ret = vhost_vq_attach_worker(vq, &ring_worker);
+		break;
+	case VHOST_GET_VRING_WORKER:
+		ring_worker.index = idx;
+		ring_worker.worker_id = vq->worker->id;
+
+		if (copy_to_user(argp, &ring_worker, sizeof(ring_worker)))
+			ret = -EFAULT;
+		break;
+	default:
+		ret = -ENOIOCTLCMD;
+		break;
+	}
+
+	mutex_unlock(&vq->mutex);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(vhost_worker_ioctl);
+
 /* Caller should have device mutex */
 long vhost_dev_set_owner(struct vhost_dev *dev)
 {
@@ -684,7 +824,7 @@ long vhost_dev_set_owner(struct vhost_dev *dev)
 		smp_wmb();
 
 		for (i = 0; i < dev->nvqs; i++)
-			dev->vqs[i]->worker = worker;
+			__vhost_vq_attach_worker(dev->vqs[i], worker);
 	}
 
 	return 0;