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>

2 files changed, 49 insertions, 0 deletions

diff --git a/include/uapi/linux/vhost.h b/include/uapi/linux/vhost.h
index 92e1b700b51c..96dc146c2d15 100644
--- a/include/uapi/linux/vhost.h
+++ b/include/uapi/linux/vhost.h
@@ -45,6 +45,25 @@
 #define VHOST_SET_LOG_BASE _IOW(VHOST_VIRTIO, 0x04, __u64)
 /* Specify an eventfd file descriptor to signal on log write. */
 #define VHOST_SET_LOG_FD _IOW(VHOST_VIRTIO, 0x07, int)
+/* By default, a device gets one vhost_worker that its virtqueues share. This
+ * command allows the owner of the device to create an additional vhost_worker
+ * for the device. It can later be bound to 1 or more of its virtqueues using
+ * the VHOST_ATTACH_VRING_WORKER command.
+ *
+ * This must be called after VHOST_SET_OWNER and the caller must be the owner
+ * of the device. The new thread will inherit caller's cgroups and namespaces,
+ * and will share the caller's memory space. The new thread will also be
+ * counted against the caller's RLIMIT_NPROC value.
+ *
+ * The worker's ID used in other commands will be returned in
+ * vhost_worker_state.
+ */
+#define VHOST_NEW_WORKER _IOR(VHOST_VIRTIO, 0x8, struct vhost_worker_state)
+/* Free a worker created with VHOST_NEW_WORKER if it's not attached to any
+ * virtqueue. If userspace is not able to call this for workers its created,
+ * the kernel will free all the device's workers when the device is closed.
+ */
+#define VHOST_FREE_WORKER _IOW(VHOST_VIRTIO, 0x9, struct vhost_worker_state)
 
 /* Ring setup. */
 /* Set number of descriptors in ring. This parameter can not
@@ -70,6 +89,20 @@
 #define VHOST_VRING_BIG_ENDIAN 1
 #define VHOST_SET_VRING_ENDIAN _IOW(VHOST_VIRTIO, 0x13, struct vhost_vring_state)
 #define VHOST_GET_VRING_ENDIAN _IOW(VHOST_VIRTIO, 0x14, struct vhost_vring_state)
+/* Attach a vhost_worker created with VHOST_NEW_WORKER to one of the device's
+ * virtqueues. This must be done before VHOST_SET_VRING_KICK and the driver
+ * specific ioctl to activate the virtqueue (VHOST_SCSI_SET_ENDPOINT,
+ * VHOST_NET_SET_BACKEND, VHOST_VSOCK_SET_RUNNING) has been run.
+ *
+ * This will replace the virtqueue's existing worker. If the replaced worker
+ * is no longer attached to any virtqueues, it can be freed with
+ * VHOST_FREE_WORKER.
+ */
+#define VHOST_ATTACH_VRING_WORKER _IOW(VHOST_VIRTIO, 0x15,		\
+				       struct vhost_vring_worker)
+/* Return the vring worker's ID */
+#define VHOST_GET_VRING_WORKER _IOWR(VHOST_VIRTIO, 0x16,		\
+				     struct vhost_vring_worker)
 
 /* The following ioctls use eventfd file descriptors to signal and poll
  * for events. */
diff --git a/include/uapi/linux/vhost_types.h b/include/uapi/linux/vhost_types.h
index c5690a8992d8..d3aad12ad1fa 100644
--- a/include/uapi/linux/vhost_types.h
+++ b/include/uapi/linux/vhost_types.h
@@ -47,6 +47,22 @@ struct vhost_vring_addr {
 	__u64 log_guest_addr;
 };
 
+struct vhost_worker_state {
+	/*
+	 * For VHOST_NEW_WORKER the kernel will return the new vhost_worker id.
+	 * For VHOST_FREE_WORKER this must be set to the id of the vhost_worker
+	 * to free.
+	 */
+	unsigned int worker_id;
+};
+
+struct vhost_vring_worker {
+	/* vring index */
+	unsigned int index;
+	/* The id of the vhost_worker returned from VHOST_NEW_WORKER */
+	unsigned int worker_id;
+};
+
 /* no alignment requirement */
 struct vhost_iotlb_msg {
 	__u64 iova;