net: mana: Add RX fencing

RX fencing allows the driver to know that any prior change to the RQs has
finished, e.g. when the RQs are disabled/enabled or the hashkey/indirection
table are changed, RX fencing is required.

Remove the previous workaround "ssleep(1)" and add the real support for
RX fencing as the PF driver supports the MANA_FENCE_RQ request now (any
old PF driver not supporting the request won't be used in production).

Signed-off-by: Dexuan Cui <decui@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Link: https://lore.kernel.org/r/20211216001748.8751-1-decui@microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

drivers/net/ethernet/microsoft/mana/mana.h

@@ -289,6 +289,8 @@ struct mana_rxq {
 
 	struct mana_cq rx_cq;
 
+	struct completion fence_event;
+
 	struct net_device *ndev;
 
 	/* Total number of receive buffers to be allocated */

drivers/net/ethernet/microsoft/mana/mana_en.c

@@ -750,6 +750,61 @@ out:
 	return err;
 }
 
+static int mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq)
+{
+	struct mana_fence_rq_resp resp = {};
+	struct mana_fence_rq_req req = {};
+	int err;
+
+	init_completion(&rxq->fence_event);
+
+	mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ,
+			     sizeof(req), sizeof(resp));
+	req.wq_obj_handle =  rxq->rxobj;
+
+	err = mana_send_request(apc->ac, &req, sizeof(req), &resp,
+				sizeof(resp));
+	if (err) {
+		netdev_err(apc->ndev, "Failed to fence RQ %u: %d\n",
+			   rxq->rxq_idx, err);
+		return err;
+	}
+
+	err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp));
+	if (err || resp.hdr.status) {
+		netdev_err(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n",
+			   rxq->rxq_idx, err, resp.hdr.status);
+		if (!err)
+			err = -EPROTO;
+
+		return err;
+	}
+
+	if (wait_for_completion_timeout(&rxq->fence_event, 10 * HZ) == 0) {
+		netdev_err(apc->ndev, "Failed to fence RQ %u: timed out\n",
+			   rxq->rxq_idx);
+		return -ETIMEDOUT;
+	}
+
+	return 0;
+}
+
+static void mana_fence_rqs(struct mana_port_context *apc)
+{
+	unsigned int rxq_idx;
+	struct mana_rxq *rxq;
+	int err;
+
+	for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) {
+		rxq = apc->rxqs[rxq_idx];
+		err = mana_fence_rq(apc, rxq);
+
+		/* In case of any error, use sleep instead. */
+		if (err)
+			msleep(100);
+	}
+}
+
 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units)
 {
 	u32 used_space_old;
@@ -1023,7 +1078,7 @@ static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq,
 		return;
 
 	case CQE_RX_OBJECT_FENCE:
-		netdev_err(ndev, "RX Fencing is unsupported\n");
+		complete(&rxq->fence_event);
 		return;
 
 	default:
@@ -1617,6 +1672,7 @@ int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
 		    bool update_hash, bool update_tab)
 {
 	u32 queue_idx;
+	int err;
 	int i;
 
 	if (update_tab) {
@@ -1626,7 +1682,13 @@ int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx,
 		}
 	}
 
-	return mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
+	err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab);
+	if (err)
+		return err;
+
+	mana_fence_rqs(apc);
+
+	return 0;
 }
 
 static int mana_init_port(struct net_device *ndev)
@@ -1773,9 +1835,6 @@ static int mana_dealloc_queues(struct net_device *ndev)
 		return err;
 	}
 
-	/* TODO: Implement RX fencing */
-	ssleep(1);
-
 	mana_destroy_vport(apc);
 
 	return 0;