[PATCH net-next] r8169: improve rtl8169_start_xmit

[PATCH net-next] r8169: improve rtl8169_start_xmit

[Heiner Kallweit]

Improve the following in rtl8169_start_xmit:
- tp->cur_tx can be accessed in parallel by rtl_tx(), therefore
  annotate the race by using WRITE_ONCE
- avoid checking stop_queue a second time by moving the doorbell check
- netif_stop_queue() uses atomic operation set_bit() that includes a
  full memory barrier on some platforms, therefore use
  smp_mb__after_atomic to avoid overhead

Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
---
 drivers/net/ethernet/realtek/r8169_main.c | 15 ++++++---------
 1 file changed, 6 insertions(+), 9 deletions(-)

diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
index 8910e900e..940fc6590 100644
--- a/drivers/net/ethernet/realtek/r8169_main.c
+++ b/drivers/net/ethernet/realtek/r8169_main.c
@@ -4226,7 +4226,7 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,
 	/* rtl_tx needs to see descriptor changes before updated tp->cur_tx */
 	smp_wmb();
 
-	tp->cur_tx += frags + 1;
+	WRITE_ONCE(tp->cur_tx, tp->cur_tx + frags + 1);
 
 	stop_queue = !rtl_tx_slots_avail(tp, MAX_SKB_FRAGS);
 	if (unlikely(stop_queue)) {
@@ -4235,13 +4235,6 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,
 		 */
 		smp_wmb();
 		netif_stop_queue(dev);
-		door_bell = true;
-	}
-
-	if (door_bell)
-		rtl8169_doorbell(tp);
-
-	if (unlikely(stop_queue)) {
 		/* Sync with rtl_tx:
 		 * - publish queue status and cur_tx ring index (write barrier)
 		 * - refresh dirty_tx ring index (read barrier).
@@ -4249,11 +4242,15 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,
 		 * status and forget to wake up queue, a racing rtl_tx thread
 		 * can't.
 		 */
-		smp_mb();
+		smp_mb__after_atomic();
 		if (rtl_tx_slots_avail(tp, MAX_SKB_FRAGS))
 			netif_start_queue(dev);
+		door_bell = true;
 	}
 
+	if (door_bell)
+		rtl8169_doorbell(tp);
+
 	return NETDEV_TX_OK;
 
 err_dma_1:
-- 
2.29.2

Re: [PATCH net-next] r8169: improve rtl8169_start_xmit

[Jakub Kicinski]

On Sat, 14 Nov 2020 21:49:53 +0100 Heiner Kallweit wrote:
> Improve the following in rtl8169_start_xmit:
> - tp->cur_tx can be accessed in parallel by rtl_tx(), therefore
>   annotate the race by using WRITE_ONCE
> - avoid checking stop_queue a second time by moving the doorbell check
> - netif_stop_queue() uses atomic operation set_bit() that includes a
>   full memory barrier on some platforms, therefore use
>   smp_mb__after_atomic to avoid overhead
> 
> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>

Applied, thanks.

Re: [PATCH net-next] r8169: improve rtl8169_start_xmit

[David Miller]

From: Heiner Kallweit <hkallweit1@gmail.com>
Date: Sat, 22 Feb 2020 17:02:51 +0100

> Only call rtl8169_xmit_frags() if the skb is actually fragmented.
> This avoid a small overhead for non-fragmented skb's, and it allows
> to simplify rtl8169_xmit_frags() a little.
> 
> Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>

Applied, thanks Heiner.

[PATCH net-next] r8169: improve rtl8169_start_xmit

[Heiner Kallweit]

Only call rtl8169_xmit_frags() if the skb is actually fragmented.
This avoid a small overhead for non-fragmented skb's, and it allows
to simplify rtl8169_xmit_frags() a little.

Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
---
 drivers/net/ethernet/realtek/r8169_main.c | 23 ++++++++++-------------
 1 file changed, 10 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
index cc4b6fd60..f081007a2 100644
--- a/drivers/net/ethernet/realtek/r8169_main.c
+++ b/drivers/net/ethernet/realtek/r8169_main.c
@@ -4087,12 +4087,10 @@ static int rtl8169_xmit_frags(struct rtl8169_private *tp, struct sk_buff *skb,
 		tp->tx_skb[entry].len = len;
 	}
 
-	if (cur_frag) {
-		tp->tx_skb[entry].skb = skb;
-		txd->opts1 |= cpu_to_le32(LastFrag);
-	}
+	tp->tx_skb[entry].skb = skb;
+	txd->opts1 |= cpu_to_le32(LastFrag);
 
-	return cur_frag;
+	return 0;
 
 err_out:
 	rtl8169_tx_clear_range(tp, tp->cur_tx + 1, cur_frag);
@@ -4217,6 +4215,7 @@ static void rtl8169_doorbell(struct rtl8169_private *tp)
 static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,
 				      struct net_device *dev)
 {
+	unsigned int frags = skb_shinfo(skb)->nr_frags;
 	struct rtl8169_private *tp = netdev_priv(dev);
 	unsigned int entry = tp->cur_tx % NUM_TX_DESC;
 	struct TxDesc *txd = tp->TxDescArray + entry;
@@ -4225,9 +4224,8 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,
 	u32 opts[2], len;
 	bool stop_queue;
 	bool door_bell;
-	int frags;
 
-	if (unlikely(!rtl_tx_slots_avail(tp, skb_shinfo(skb)->nr_frags))) {
+	if (unlikely(!rtl_tx_slots_avail(tp, frags))) {
 		netif_err(tp, drv, dev, "BUG! Tx Ring full when queue awake!\n");
 		goto err_stop_0;
 	}
@@ -4256,14 +4254,13 @@ static netdev_tx_t rtl8169_start_xmit(struct sk_buff *skb,
 	tp->tx_skb[entry].len = len;
 	txd->addr = cpu_to_le64(mapping);
 
-	frags = rtl8169_xmit_frags(tp, skb, opts);
-	if (frags < 0)
-		goto err_dma_1;
-	else if (frags)
-		opts[0] |= FirstFrag;
-	else {
+	if (!frags) {
 		opts[0] |= FirstFrag | LastFrag;
 		tp->tx_skb[entry].skb = skb;
+	} else {
+		if (rtl8169_xmit_frags(tp, skb, opts))
+			goto err_dma_1;
+		opts[0] |= FirstFrag;
 	}
 
 	txd->opts2 = cpu_to_le32(opts[1]);
-- 
2.25.1