@@ -232,12 +232,9 @@ void vhost_poll_stop(struct vhost_poll *poll)
}
EXPORT_SYMBOL_GPL(vhost_poll_stop);
-static void vhost_work_queue_on(struct vhost_worker *worker,
- struct vhost_work *work)
+static void vhost_worker_work_queue(struct vhost_worker *worker,
+ struct vhost_work *work)
{
- if (!worker)
- return;
-
if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
/* We can only add the work to the list after we're
* sure it was not in the list.
@@ -248,21 +245,32 @@ static void vhost_work_queue_on(struct vhost_worker *worker,
}
}
-static int vhost_workers_idr_flush_iter(int id, void *worker, void *dev)
+static void vhost_worker_flush(struct vhost_worker *worker)
{
struct vhost_flush_struct flush;
init_completion(&flush.wait_event);
vhost_work_init(&flush.work, vhost_flush_work);
- vhost_work_queue_on(worker, &flush.work);
+ vhost_worker_work_queue(worker, &flush.work);
wait_for_completion(&flush.wait_event);
+}
+
+static int vhost_workers_idr_flush_iter(int id, void *worker, void *dev)
+{
+ vhost_worker_flush(worker);
return 0;
}
void vhost_vq_work_queue(struct vhost_virtqueue *vq, struct vhost_work *work)
{
- vhost_work_queue_on(vq->worker, work);
+ struct vhost_worker *worker;
+
+ rcu_read_lock();
+ worker = rcu_dereference(vq->worker);
+ if (worker)
+ vhost_worker_work_queue(worker, work);
+ rcu_read_unlock();
}
EXPORT_SYMBOL_GPL(vhost_vq_work_queue);
@@ -282,7 +290,16 @@ EXPORT_SYMBOL_GPL(vhost_work_dev_flush);
/* A lockless hint for busy polling code to exit the loop */
bool vhost_vq_has_work(struct vhost_virtqueue *vq)
{
- return vq->worker && !llist_empty(&vq->worker->work_list);
+ struct vhost_worker *worker;
+ bool has_work = false;
+
+ rcu_read_lock();
+ worker = rcu_dereference(vq->worker);
+ if (worker && !llist_empty(&worker->work_list))
+ has_work = true;
+ rcu_read_unlock();
+
+ return has_work;
}
EXPORT_SYMBOL_GPL(vhost_vq_has_work);
@@ -507,7 +524,7 @@ void vhost_dev_init(struct vhost_dev *dev,
vq->log = NULL;
vq->indirect = NULL;
vq->heads = NULL;
- vq->worker = NULL;
+ rcu_assign_pointer(vq->worker, NULL);
vq->dev = dev;
mutex_init(&vq->mutex);
vhost_vq_reset(dev, vq);
@@ -587,11 +604,30 @@ static void vhost_worker_put(struct vhost_dev *dev, struct vhost_worker *worker)
kfree(worker);
}
-static void vhost_vq_detach_worker(struct vhost_virtqueue *vq)
+static void vhost_vq_swap_worker(struct vhost_virtqueue *vq,
+ struct vhost_worker *new_worker, bool flush)
{
- if (vq->worker)
- vhost_worker_put(vq->dev, vq->worker);
- vq->worker = NULL;
+ struct vhost_worker *old_worker;
+
+ old_worker = rcu_dereference_check(vq->worker,
+ lockdep_is_held(&vq->dev->mutex));
+ rcu_assign_pointer(vq->worker, new_worker);
+
+ if (!old_worker)
+ return;
+
+ if (flush) {
+ /*
+ * For dev cleanup we won't have work running, but for the
+ * dynamic attach case we might so make sure we see the new
+ * worker and there is no work in the old worker when we
+ * return.
+ */
+ synchronize_rcu();
+ vhost_worker_flush(old_worker);
+ }
+
+ vhost_worker_put(vq->dev, old_worker);
}
static int vhost_workers_idr_iter(int id, void *worker, void *dev)
@@ -608,7 +644,7 @@ static void vhost_workers_free(struct vhost_dev *dev)
return;
for (i = 0; i < dev->nvqs; i++)
- vhost_vq_detach_worker(dev->vqs[i]);
+ vhost_vq_swap_worker(dev->vqs[i], NULL, false);
idr_for_each(&dev->worker_idr, vhost_workers_idr_iter, dev);
}
@@ -664,18 +700,13 @@ static int vhost_vq_attach_worker(struct vhost_virtqueue *vq,
if (!dev->use_worker)
return -EINVAL;
- /* We don't support setting a worker on an active vq */
- if (vq->private_data)
- return -EBUSY;
-
worker = idr_find(&dev->worker_idr, info->worker_id);
if (!worker)
return -ENODEV;
refcount_inc(&worker->refcount);
- vhost_vq_detach_worker(vq);
- vq->worker = worker;
+ vhost_vq_swap_worker(vq, worker, true);
return 0;
}
@@ -1826,7 +1857,10 @@ long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *arg
r = -EFAULT;
break;
}
+ /* We might flush worker, so we can't hold the vq mutex. */
+ mutex_unlock(&vq->mutex);
r = vhost_vq_attach_worker(vq, &w);
+ mutex_lock(&vq->mutex);
if (!r && copy_to_user(argp, &w, sizeof(w)))
r = -EFAULT;
break;
@@ -81,7 +81,7 @@ struct vhost_vring_call {
/* The virtqueue structure describes a queue attached to a device. */
struct vhost_virtqueue {
struct vhost_dev *dev;
- struct vhost_worker *worker;
+ struct vhost_worker __rcu *worker;
/* The actual ring of buffers. */
struct mutex mutex;
@@ -88,7 +88,7 @@
#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 the virtqueue is active.
+ * virtqueues.
*/
#define VHOST_ATTACH_VRING_WORKER _IOR(VHOST_VIRTIO, 0x15, \
struct vhost_vring_worker)
This patch allows userspace to change the vq to worker mapping while it's in use so tools can do this setup post device creation if needed. Signed-off-by: Mike Christie <michael.christie@oracle.com> --- drivers/vhost/vhost.c | 76 +++++++++++++++++++++++++++----------- drivers/vhost/vhost.h | 2 +- include/uapi/linux/vhost.h | 2 +- 3 files changed, 57 insertions(+), 23 deletions(-)