diff mbox

[(net-next.git),13/18] stmmac: perf, remove modulo in stmmac_rx()

Message ID 1449650274-14896-14-git-send-email-peppe.cavallaro@st.com
State Changes Requested, archived
Delegated to: David Miller
Headers show

Commit Message

Giuseppe CAVALLARO Dec. 9, 2015, 8:37 a.m. UTC 
The indexes into the ring buffer are always incremented, and
the entry is accessed via doing a modulo to find the "real" index.
Modulo is an expensive operation.

This patch replaces the modulo with a simple if clamp.
It helps improve stmmac RX path as it's being called inside RX loop.

Signed-off-by: Fabrice Gasnier <fabrice.gasnier@st.com>
Signed-off-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
---
 drivers/net/ethernet/stmicro/stmmac/stmmac_main.c |   28 ++++++++++++++++++---
 1 files changed, 24 insertions(+), 4 deletions(-)

Comments

David Laight Dec. 9, 2015, 5:21 p.m. UTC | #1 
From: Giuseppe Cavallaro
> Sent: 09 December 2015 08:38
> The indexes into the ring buffer are always incremented, and
> the entry is accessed via doing a modulo to find the "real" index.
> Modulo is an expensive operation.
> 
> This patch replaces the modulo with a simple if clamp.
> It helps improve stmmac RX path as it's being called inside RX loop.

Is dma_rx_size always a power of 2 ?
If so you can replace the % by & and remove the conditionals.

If you have changed the read and write values to indexes
then you need to be certain that the 'ring full' and 'ring empty'
cases are correctly distinguished.

	David
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Giuseppe CAVALLARO Dec. 10, 2015, 5:11 a.m. UTC | #2 
On 12/9/2015 6:21 PM, David Laight wrote:
> From: Giuseppe Cavallaro
>> Sent: 09 December 2015 08:38
>> The indexes into the ring buffer are always incremented, and
>> the entry is accessed via doing a modulo to find the "real" index.
>> Modulo is an expensive operation.
>>
>> This patch replaces the modulo with a simple if clamp.
>> It helps improve stmmac RX path as it's being called inside RX loop.
>
> Is dma_rx_size always a power of 2 ?

no

> If so you can replace the % by & and remove the conditionals.
>
> If you have changed the read and write values to indexes
> then you need to be certain that the 'ring full' and 'ring empty'
> cases are correctly distinguished.

this is implicitly managed by dirty and curr index.

thx
peppe

>
> 	David
>

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Giuseppe CAVALLARO Dec. 10, 2015, 5:51 a.m. UTC | #3 
On 12/10/2015 6:11 AM, Giuseppe CAVALLARO wrote:
> On 12/9/2015 6:21 PM, David Laight wrote:
>> From: Giuseppe Cavallaro
>>> Sent: 09 December 2015 08:38
>>> The indexes into the ring buffer are always incremented, and
>>> the entry is accessed via doing a modulo to find the "real" index.
>>> Modulo is an expensive operation.
>>>
>>> This patch replaces the modulo with a simple if clamp.
>>> It helps improve stmmac RX path as it's being called inside RX loop.
>>
>> Is dma_rx_size always a power of 2 ?
>
> no
>
>> If so you can replace the % by & and remove the conditionals.
>>
>> If you have changed the read and write values to indexes
>> then you need to be certain that the 'ring full' and 'ring empty'
>> cases are correctly distinguished.
>
> this is implicitly managed by dirty and curr index.

   ... also the dma will fail with unavailable buffer.

peppe
>
> thx
> peppe
>
>>
>>     David
>>
>
>
>

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
index 0a39331..52ac648 100644
--- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
+++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
@@ -211,6 +211,18 @@  static inline u32 stmmac_tx_avail(struct stmmac_priv *priv)
 	return avail;
 }
 
+static inline u32 stmmac_rx_dirty(struct stmmac_priv *priv)
+{
+	unsigned dirty;
+
+	if (priv->dirty_rx <= priv->cur_rx)
+		dirty = priv->cur_rx - priv->dirty_rx;
+	else
+		dirty = priv->dma_rx_size - priv->dirty_rx + priv->cur_rx;
+
+	return dirty;
+}
+
 /**
  * stmmac_hw_fix_mac_speed - callback for speed selection
  * @priv: driver private structure
@@ -2165,9 +2177,10 @@  static inline void stmmac_rx_refill(struct stmmac_priv *priv)
 {
 	unsigned int rxsize = priv->dma_rx_size;
 	int bfsize = priv->dma_buf_sz;
+	unsigned int entry = priv->dirty_rx;
+	int dirty = stmmac_rx_dirty(priv);
 
-	for (; priv->cur_rx - priv->dirty_rx > 0; priv->dirty_rx++) {
-		unsigned int entry = priv->dirty_rx % rxsize;
+	while (dirty-- > 0) {
 		struct dma_desc *p;
 
 		if (priv->extend_desc)
@@ -2203,6 +2216,10 @@  static inline void stmmac_rx_refill(struct stmmac_priv *priv)
 		wmb();
 		priv->hw->desc->set_rx_owner(p);
 		wmb();
+
+		if (unlikely(++priv->dirty_rx >= rxsize))
+			priv->dirty_rx = 0;
+		entry = priv->dirty_rx;
 	}
 }
 
@@ -2216,7 +2233,7 @@  static inline void stmmac_rx_refill(struct stmmac_priv *priv)
 static int stmmac_rx(struct stmmac_priv *priv, int limit)
 {
 	unsigned int rxsize = priv->dma_rx_size;
-	unsigned int entry = priv->cur_rx % rxsize;
+	unsigned int entry = priv->cur_rx;
 	unsigned int next_entry;
 	unsigned int count = 0;
 	int coe = priv->hw->rx_csum;
@@ -2246,7 +2263,10 @@  static int stmmac_rx(struct stmmac_priv *priv, int limit)
 
 		count++;
 
-		next_entry = (++priv->cur_rx) % rxsize;
+		if (unlikely(++priv->cur_rx >= rxsize))
+			priv->cur_rx = 0;
+		next_entry = priv->cur_rx;
+
 		if (priv->extend_desc)
 			prefetch(priv->dma_erx + next_entry);
 		else