commit e0287c4ab87905dd4a2e45cf791f8e0a87fe602e
author Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com> Sun May 12 14:43:36 2013 +0300
committer John W. Linville <linville@tuxdriver.com> Wed May 22 15:08:35 2013 -0400
tree edd55224fb599e71c7752c3df57e65b4dcffde81
parent 98658095623109bdace46f21bece028c904fb900

wil6210: use NAPI

Introduce NAPI for Rx and Tx completion.

This fixes packet reordering that happens when Rx handled right in
the IRQ: netif_rx puts packet in 'percpu' queue, then network stack
fetches packets from 'percpu' queues for processing, with different
pattern of queue switching. As result, network stack see packets
in different order. This causes hard to understand TCP throughput
degradation in about 30min

Complete polling if only one packet was processed - this eliminates
empty polls that would be otherwise done at the end of each burst

Signed-off-by: Vladimir Kondratiev <qca_vkondrat@qca.qualcomm.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>

drivers/net/wireless/ath/wil6210/netdev.c

diff --git a/drivers/net/wireless/ath/wil6210/netdev.c b/drivers/net/wireless/ath/wil6210/netdev.c
index 098a8ec..29dd1e5 100644
--- a/drivers/net/wireless/ath/wil6210/netdev.c
+++ b/drivers/net/wireless/ath/wil6210/netdev.c
@@ -40,6 +40,55 @@
 	.ndo_validate_addr	= eth_validate_addr,
 };
 
+static int wil6210_netdev_poll_rx(struct napi_struct *napi, int budget)
+{
+	struct wil6210_priv *wil = container_of(napi, struct wil6210_priv,
+						napi_rx);
+	int quota = budget;
+	int done;
+
+	wil_rx_handle(wil, &quota);
+	done = budget - quota;
+
+	if (done <= 1) { /* burst ends - only one packet processed */
+		napi_complete(napi);
+		wil6210_unmask_irq_rx(wil);
+		wil_dbg_txrx(wil, "NAPI RX complete\n");
+	}
+
+	wil_dbg_txrx(wil, "NAPI RX poll(%d) done %d\n", budget, done);
+
+	return done;
+}
+
+static int wil6210_netdev_poll_tx(struct napi_struct *napi, int budget)
+{
+	struct wil6210_priv *wil = container_of(napi, struct wil6210_priv,
+						napi_tx);
+	int tx_done = 0;
+	uint i;
+
+	/* always process ALL Tx complete, regardless budget - it is fast */
+	for (i = 0; i < WIL6210_MAX_TX_RINGS; i++) {
+		struct vring *vring = &wil->vring_tx[i];
+
+		if (!vring->va)
+			continue;
+
+		tx_done += wil_tx_complete(wil, i);
+	}
+
+	if (tx_done <= 1) { /* burst ends - only one packet processed */
+		napi_complete(napi);
+		wil6210_unmask_irq_tx(wil);
+		wil_dbg_txrx(wil, "NAPI TX complete\n");
+	}
+
+	wil_dbg_txrx(wil, "NAPI TX poll(%d) done %d\n", budget, tx_done);
+
+	return min(tx_done, budget);
+}
+
 void *wil_if_alloc(struct device *dev, void __iomem *csr)
 {
 	struct net_device *ndev;
@@ -81,6 +130,11 @@
 	SET_NETDEV_DEV(ndev, wiphy_dev(wdev->wiphy));
 	wdev->netdev = ndev;
 
+	netif_napi_add(ndev, &wil->napi_rx, wil6210_netdev_poll_rx,
+		       WIL6210_NAPI_BUDGET);
+	netif_napi_add(ndev, &wil->napi_tx, wil6210_netdev_poll_tx,
+		       WIL6210_NAPI_BUDGET);
+
 	wil_link_off(wil);
 
 	return wil;