[RFC PATCH 7/8] vhost, vhost-scsi: flush IO vqs then send TMF rsp

[Mike Christie <michael.christie@xxxxxxxxxx>]

With one worker we will always send the scsi cmd responses then send the
TMF rsp, because LIO will always complete the scsi cmds first which
calls vhost_scsi_release_cmd to add them to the work queue.

When the next patch adds multiple worker support, the worker threads
could still be sending their responses when the tmf's work is run.
So this patch has vhost-scsi flush the IO vqs on other worker threads
before we send the tmf response.

Signed-off-by: Mike Christie <michael.christie@xxxxxxxxxx>
---
 drivers/vhost/scsi.c  | 20 ++++++++++++++++++--
 drivers/vhost/vhost.c |  9 +++++++++
 drivers/vhost/vhost.h |  1 +
 3 files changed, 28 insertions(+), 2 deletions(-)

diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
index 08bc513..8005a7f 100644
--- a/drivers/vhost/scsi.c
+++ b/drivers/vhost/scsi.c
@@ -1179,12 +1179,28 @@ static void vhost_scsi_tmf_resp_work(struct vhost_work *work)
 {
 	struct vhost_scsi_tmf *tmf = container_of(work, struct vhost_scsi_tmf,
 						  vwork);
+	struct vhost_virtqueue *vq;
+	unsigned int cpu;
 	int resp_code;
+	int i;
 
-	if (tmf->scsi_resp == TMR_FUNCTION_COMPLETE)
+	if (tmf->scsi_resp == TMR_FUNCTION_COMPLETE) {
+		/*
+		 * When processing a TMF, lio completes the cmds then the
+		 * TMF, so with one worker the TMF always completes after
+		 * cmds. For multiple worker support, we must flush every
+		 * worker that runs on a different cpu than the EVT vq.
+		 */
+		cpu = tmf->vhost->vqs[VHOST_SCSI_VQ_CTL].vq.cpu;
+		for (i = VHOST_SCSI_VQ_IO; i < tmf->vhost->dev.nvqs; i++) {
+			vq = &tmf->vhost->vqs[i].vq;
+			if (cpu != vq->cpu)
+				vhost_vq_work_flush(vq);
+		}
 		resp_code = VIRTIO_SCSI_S_FUNCTION_SUCCEEDED;
-	else
+	} else {
 		resp_code = VIRTIO_SCSI_S_FUNCTION_REJECTED;
+	}
 
 	vhost_scsi_send_tmf_resp(tmf->vhost, &tmf->svq->vq, tmf->in_iovs,
 				 tmf->vq_desc, &tmf->resp_iov, resp_code);
diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
index f425d0f..4aae504 100644
--- a/drivers/vhost/vhost.c
+++ b/drivers/vhost/vhost.c
@@ -250,6 +250,15 @@ void vhost_work_dev_flush(struct vhost_dev *dev)
 }
 EXPORT_SYMBOL_GPL(vhost_work_dev_flush);
 
+void vhost_vq_work_flush(struct vhost_virtqueue *vq)
+{
+	if (vq->cpu != -1)
+		flush_work(&vq->work);
+	else
+		vhost_work_dev_flush(vq->dev);
+}
+EXPORT_SYMBOL_GPL(vhost_vq_work_flush);
+
 /* Flush any work that has been scheduled. When calling this, don't hold any
  * locks that are also used by the callback. */
 void vhost_poll_flush(struct vhost_poll *poll)
diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h
index 28ff4a2..2d306f8 100644
--- a/drivers/vhost/vhost.h
+++ b/drivers/vhost/vhost.h
@@ -40,6 +40,7 @@ struct vhost_poll {
 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work);
 bool vhost_has_work(struct vhost_dev *dev);
 void vhost_vq_work_queue(struct vhost_virtqueue *vq, struct vhost_work *work);
+void vhost_vq_work_flush(struct vhost_virtqueue *vq);
 
 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 		     __poll_t mask, struct vhost_dev *dev,
-- 
1.8.3.1