onic: add periodic draining

In unidirectional flows, like UDP, generated by kernel through onic driver, the
traffic can block because of numerous frames that are in HW and not released
fast enough by the driver.

This patch adds a timer that releases frames periodically, regardless of the
reception and transmission rates.

This mechanism was chosen instead of the workqueue mechanism because of its
higher priority, larger than RX and TX procedures.

Change-Id: I141fcb7c2857a726e5e3954cabefb923c8ac0aac
Signed-off-by: Marian Rotariu <marian.rotariu@freescale.com>
Reviewed-on: http://git.am.freescale.net:8181/13569
Tested-by: Review Code-CDREVIEW <CDREVIEW@freescale.com>
Reviewed-by: Madalin-Cristian Bucur <madalin.bucur@freescale.com>

1 files changed, 37 insertions, 19 deletions

diff --git a/drivers/net/ethernet/freescale/dpa/dpaa_eth_generic.c b/drivers/net/ethernet/freescale/dpa/dpaa_eth_generic.c
index 6b1fd0a..8f2ad0f 100644
--- a/drivers/net/ethernet/freescale/dpa/dpaa_eth_generic.c
+++ b/drivers/net/ethernet/freescale/dpa/dpaa_eth_generic.c
@@ -89,7 +89,7 @@ static void dpa_generic_ern(struct qman_portal *portal,
 			    const struct qm_mr_entry *msg);
 static int __hot dpa_generic_tx(struct sk_buff *skb,
 				struct net_device *netdev);
-static void dpa_generic_drain_bp(struct dpa_bp *bp);
+static void dpa_generic_drain_bp(struct dpa_bp *bp, u8 nbuf);
 
 static const struct net_device_ops dpa_generic_ops = {
 	.ndo_open = dpa_generic_start,
@@ -105,6 +105,17 @@ static const struct net_device_ops dpa_generic_ops = {
 	.ndo_change_mtu = dpa_change_mtu,
 };
 
+void dpa_generic_draining_timer(unsigned long arg)
+{
+	struct dpa_generic_priv_s *priv = (struct dpa_generic_priv_s *)arg;
+
+	/* drain in pairs of 4 buffers */
+	dpa_generic_drain_bp(priv->draining_tx_bp, 4);
+
+	if (atomic_read(&priv->ifup))
+		mod_timer(&(priv->timer), jiffies + 1);
+}
+
 struct rtnl_link_stats64 *__cold
 dpa_generic_get_stats64(struct net_device *netdev,
 			struct rtnl_link_stats64 *stats)
@@ -260,6 +271,9 @@ static int __cold dpa_generic_start(struct net_device *netdev)
 	dpaa_generic_napi_enable(priv);
 	netif_tx_start_all_queues(netdev);
 
+	mod_timer(&priv->timer, jiffies + 100);
+	atomic_dec(&priv->ifup);
+
 	return 0;
 }
 
@@ -270,6 +284,8 @@ static int __cold dpa_generic_stop(struct net_device *netdev)
 	netif_tx_stop_all_queues(netdev);
 	dpaa_generic_napi_disable(priv);
 
+	atomic_inc_not_zero(&priv->ifup);
+
 	return 0;
 }
 
@@ -365,7 +381,7 @@ dpa_generic_rx_dqrr(struct qman_portal *portal,
 	/* This is needed for TCP traffic as draining only on TX is not
 	 * enough
 	 */
-	dpa_generic_drain_bp(priv->draining_tx_bp);
+	dpa_generic_drain_bp(priv->draining_tx_bp, 1);
 
 	if (unlikely(dpaa_eth_napi_schedule(percpu_priv, portal)))
 		return qman_cb_dqrr_stop;
@@ -444,25 +460,27 @@ qman_consume:
 	return qman_cb_dqrr_consume;
 }
 
-static void dpa_generic_drain_bp(struct dpa_bp *bp)
+static void dpa_generic_drain_bp(struct dpa_bp *bp, u8 nbuf)
 {
-	int ret;
-	struct bm_buffer bmb;
+	int ret, i;
+	struct bm_buffer bmb[8];
 	dma_addr_t addr;
 	int *countptr = __this_cpu_ptr(bp->percpu_count);
 	int count = *countptr;
 	struct sk_buff **skbh;
 
 	do {
-		/* most likely, the bpool has only one buffer in it */
-		ret = bman_acquire(bp->pool, &bmb, 1, 0);
+		/* bman_acquire will fail if nbuf > 8 */
+		ret = bman_acquire(bp->pool, bmb, nbuf, 0);
 		if (ret > 0) {
-			addr = bm_buf_addr(&bmb);
-			skbh = (struct sk_buff **)phys_to_virt(addr);
-			dma_unmap_single(bp->dev, addr, bp->size,
-					DMA_TO_DEVICE);
-			dev_kfree_skb_any(*skbh);
-			count--;
+			for (i = 0; i < nbuf; i++) {
+				addr = bm_buf_addr(&bmb[i]);
+				skbh = (struct sk_buff **)phys_to_virt(addr);
+				dma_unmap_single(bp->dev, addr, bp->size,
+						DMA_TO_DEVICE);
+				dev_kfree_skb_any(*skbh);
+			}
+			count -= i;
 		}
 	} while (ret > 0);
 
@@ -639,12 +657,7 @@ static int __hot dpa_generic_tx(struct sk_buff *skb, struct net_device *netdev)
 	/* fd.cmd |= FM_FD_CMD_FCO; */
 	fd.bpid = bp->bpid;
 
-	/* (Partially) drain the Draining Buffer Pool pool; each core
-	 * acquires at most the number of buffers it put there; since
-	 * BMan allows for up to 8 buffers to be acquired at one time,
-	 * work in batches of 8 for efficiency reasons
-	 */
-	dpa_generic_drain_bp(bp);
+	dpa_generic_drain_bp(bp, 1);
 
 	queue_mapping = dpa_get_queue_mapping(skb);
 	egress_fq = priv->egress_fqs[queue_mapping];
@@ -1428,6 +1441,11 @@ static int dpa_generic_eth_probe(struct platform_device *_of_dev)
 	priv->msg_enable = netif_msg_init(debug, -1);
 	priv->tx_headroom = DPA_DEFAULT_TX_HEADROOM;
 
+	init_timer(&priv->timer);
+	atomic_set(&priv->ifup, 0);
+	priv->timer.data = (unsigned long)priv;
+	priv->timer.function = dpa_generic_draining_timer;
+
 	err = dpa_generic_bp_create(netdev, rx_bp_count, rx_bp, rx_buf_layout,
 			draining_tx_bp, tx_buf_layout);
 	if (err < 0)