RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Stephan Linke]

Hi,

I am on an 862. Anyway I can't find another definition of dma_cache_inv() but the NO OP in asm-ppc/io.h. Could you give me e hint
where it is defined in your kernel?

Thanks, Stephan

> -----Original Message-----
> From: Joakim Tjernlund [mailto:joakim.tjernlund@lumentis.se]
> Sent: Montag, 3. Februar 2003 18:23
> To: Stephan Linke
> Subject: RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3
>
>
> > Hi Jocke,
> >
> > in your latest patch you are using dma_cache_inv() instead of invalidate_dcache_range().
> > The only dma_cache_inv() I can find is in include/asm-ppc/io.h. and it's a "do{}while(0)".
> > Are you shure that this was your intention? It seames to me like you could remove that call as well.
>
> I guess you are on 8260? On 8260 there is no need for invalidate_dcache_range() since
> it's the CPM is cache coherent. On 8xx it is not cache coherent. I switched
> to dma_cache_inv() because it's a no op on 8260 and a invalidate_dcache_range() on 8xx so
> it would be easy to adapt the patch to both CPU's.
>
>  Jocke
>
>


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Joakim Tjernlund]

>
> Hi,
>
> I am on an 862. Anyway I can't find another definition of dma_cache_inv() but the NO OP in asm-ppc/io.h. Could you give me e hint
> where it is defined in your kernel?

It's in asm-ppc/io.h (2.4.20) and there are 2 definitions of dma_cache_inv() which depends on
CONFIG_NOT_COHERENT_CACHE(should be defined for 8xx). What kernel version are you running?

 Jocke
>
> Thanks, Stephan
>
> > -----Original Message-----
> > From: Joakim Tjernlund [mailto:joakim.tjernlund@lumentis.se]
> > Sent: Montag, 3. Februar 2003 18:23
> > To: Stephan Linke
> > Subject: RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3
> >
> >
> > > Hi Jocke,
> > >
> > > in your latest patch you are using dma_cache_inv() instead of invalidate_dcache_range().
> > > The only dma_cache_inv() I can find is in include/asm-ppc/io.h. and it's a "do{}while(0)".
> > > Are you shure that this was your intention? It seames to me like you could remove that call as well.
> >
> > I guess you are on 8260? On 8260 there is no need for invalidate_dcache_range() since
> > it's the CPM is cache coherent. On 8xx it is not cache coherent. I switched
> > to dma_cache_inv() because it's a no op on 8260 and a invalidate_dcache_range() on 8xx so
> > it would be easy to adapt the patch to both CPU's.
> >
> >  Jocke
> >
> >
>
>


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Pantelis Antoniou]

[-- Attachment #1: Type: text/plain, Size: 1473 bytes --]

Joakim Tjernlund wrote:

>>Hi,
>>
>>I am on an 862. Anyway I can't find another definition of dma_cache_inv() but the NO OP in asm-ppc/io.h. Could you give me e hint
>>where it is defined in your kernel?
>>
>
>It's in asm-ppc/io.h (2.4.20) and there are 2 definitions of dma_cache_inv() which depends on
>CONFIG_NOT_COHERENT_CACHE(should be defined for 8xx). What kernel version are you running?
>
> Jocke
>
>>Thanks, Stephan
>>
>>
>>>-----Original Message-----
>>>From: Joakim Tjernlund [mailto:joakim.tjernlund@lumentis.se]
>>>Sent: Montag, 3. Februar 2003 18:23
>>>To: Stephan Linke
>>>Subject: RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3
>>>
>>>
>>>
>>>>Hi Jocke,
>>>>
>>>>in your latest patch you are using dma_cache_inv() instead of invalidate_dcache_range().
>>>>The only dma_cache_inv() I can find is in include/asm-ppc/io.h. and it's a "do{}while(0)".
>>>>Are you shure that this was your intention? It seames to me like you could remove that call as well.
>>>>
>>>I guess you are on 8260? On 8260 there is no need for invalidate_dcache_range() since
>>>it's the CPM is cache coherent. On 8xx it is not cache coherent. I switched
>>>to dma_cache_inv() because it's a no op on 8260 and a invalidate_dcache_range() on 8xx so
>>>it would be easy to adapt the patch to both CPU's.
>>>
>>> Jocke
>>>
>>>
>>>
>>
>
>
>
>
>
>
Hi guys,

I have created a patch that applies cleanly to the head of
linuxppc_2_4_devel
and it works great.

Keep up the good work!

Pantelis



[-- Attachment #2: joakim-enet-8xx-panto.patch --]
[-- Type: text/plain, Size: 10627 bytes --]

# This is a BitKeeper generated patch for the following project:
# Project Name: Linux 2.4 for PowerPC development tree
# This patch format is intended for GNU patch command version 2.5 or higher.
# This patch includes the following deltas:
#	           ChangeSet	1.1207  -> 1.1209
#	arch/ppc/8xx_io/enet.c	1.25    -> 1.27
#
# The following is the BitKeeper ChangeSet Log
# --------------------------------------------
# 03/02/10	panto@macpanto.intranet.gr	1.1208
# Import of Joakim Tjernlund enet patch.
# --------------------------------------------
# 03/02/10	panto@macpanto.intranet.gr	1.1209
# Removed private declaration of dma_cache_inv, include header.
# --------------------------------------------
#
diff -Nru a/arch/ppc/8xx_io/enet.c b/arch/ppc/8xx_io/enet.c
--- a/arch/ppc/8xx_io/enet.c	Mon Feb 10 13:06:35 2003
+++ b/arch/ppc/8xx_io/enet.c	Mon Feb 10 13:06:35 2003
@@ -34,7 +34,6 @@
 #include <linux/ioport.h>
 #include <linux/slab.h>
 #include <linux/interrupt.h>
-#include <linux/pci.h>
 #include <linux/init.h>
 #include <linux/delay.h>
 #include <linux/netdevice.h>
@@ -42,6 +41,7 @@
 #include <linux/skbuff.h>
 #include <linux/spinlock.h>

+#include <asm/io.h>
 #include <asm/8xx_immap.h>
 #include <asm/pgtable.h>
 #include <asm/mpc8xx.h>
@@ -86,6 +86,14 @@
  * All functions are directly controlled using I/O pins.  See <asm/commproc.h>.
  */

+/* Define COPY_SMALL_FRAMES if you want to save buffer memory for small packets
+ * at a small performance hit. Note performance testing needed */
+/* #define COPY_SMALL_FRAMES 1  */
+
+#ifdef COPY_SMALL_FRAMES
+  #define RX_COPYBREAK (256-16) /* dev_alloc_skb() adds 16 bytes for internal use */
+#endif
+
 /* The transmitter timeout
  */
 #define TX_TIMEOUT	(2*HZ)
@@ -97,19 +105,17 @@
  * the skbuffer directly.
  */
 #ifdef CONFIG_ENET_BIG_BUFFERS
-#define CPM_ENET_RX_PAGES	32
-#define CPM_ENET_RX_FRSIZE	2048
-#define CPM_ENET_RX_FRPPG	(PAGE_SIZE / CPM_ENET_RX_FRSIZE)
-#define RX_RING_SIZE		(CPM_ENET_RX_FRPPG * CPM_ENET_RX_PAGES)
-#define TX_RING_SIZE		64	/* Must be power of two */
-#define TX_RING_MOD_MASK	63	/*   for this to work */
+  #define RX_RING_SIZE		64
+  #define TX_RING_SIZE		64	/* Must be power of two for this to work */
 #else
-#define CPM_ENET_RX_PAGES	4
-#define CPM_ENET_RX_FRSIZE	2048
-#define CPM_ENET_RX_FRPPG	(PAGE_SIZE / CPM_ENET_RX_FRSIZE)
-#define RX_RING_SIZE		(CPM_ENET_RX_FRPPG * CPM_ENET_RX_PAGES)
-#define TX_RING_SIZE		8	/* Must be power of two */
-#define TX_RING_MOD_MASK	7	/*   for this to work */
+  #define RX_RING_SIZE		8
+  #define TX_RING_SIZE		8	/* Must be power of two for this to work */
+#endif
+#define TX_RING_MOD_MASK	(TX_RING_SIZE-1)
+
+#define CPM_ENET_RX_FRSIZE	1552 /* must be a multiple of cache line */
+#if CPM_ENET_RX_FRSIZE % L1_CACHE_LINE_SIZE != 0
+    #error CPM_ENET_RX_FRSIZE must be a multiple of L1 cache size
 #endif

 /* The CPM stores dest/src/type, data, and checksum for receive packets.
@@ -143,7 +149,7 @@
 	/* Virtual addresses for the receive buffers because we can't
 	 * do a __va() on them anymore.
 	 */
-	unsigned char *rx_vaddr[RX_RING_SIZE];
+	void    *rx_vaddr[RX_RING_SIZE];
 	struct	net_device_stats stats;
 	uint	tx_full;
 	spinlock_t lock;
@@ -205,7 +211,6 @@
 	/* Fill in a Tx ring entry */
 	bdp = cep->cur_tx;

-#ifndef final_version
 	if (bdp->cbd_sc & BD_ENET_TX_READY) {
 		/* Ooops.  All transmit buffers are full.  Bail out.
 		 * This should not happen, since cep->tx_busy should be set.
@@ -213,7 +218,6 @@
 		printk("%s: tx queue full!.\n", dev->name);
 		return 1;
 	}
-#endif

 	/* Clear all of the status flags.
 	 */
@@ -279,7 +283,6 @@

 	printk("%s: transmit timed out.\n", dev->name);
 	cep->stats.tx_errors++;
-#ifndef final_version
 	{
 		int	i;
 		cbd_t	*bdp;
@@ -299,7 +302,6 @@
 			       bdp->cbd_datlen,
 			       bdp->cbd_bufaddr);
 	}
-#endif
 	if (!cep->tx_full)
 		netif_wake_queue(dev);
 }
@@ -370,11 +372,11 @@

 		cep->stats.tx_packets++;

-		/* Deferred means some collisions occurred during transmit,
-		 * but we eventually sent the packet OK.
-		 */
-		if (bdp->cbd_sc & BD_ENET_TX_DEF)
-			cep->stats.collisions++;
+		/* Check retry counter, i.e. collision counter */
+		if (bdp->cbd_sc & BD_ENET_TX_RCMASK){
+			/* Note that counter cannot go higher than 15 */
+			cep->stats.collisions+=(bdp->cbd_sc & BD_ENET_TX_RCMASK)>>2;
+		}

 		/* Free the sk buffer associated with this last transmit.
 		*/
@@ -449,6 +451,7 @@
 	struct	scc_enet_private *cep;
 	volatile cbd_t	*bdp;
 	struct	sk_buff *skb;
+	struct	sk_buff *skb_tmp;
 	ushort	pkt_len;

 	cep = (struct scc_enet_private *)dev->priv;
@@ -458,83 +461,93 @@
 	 */
 	bdp = cep->cur_rx;

-for (;;) {
-	if (bdp->cbd_sc & BD_ENET_RX_EMPTY)
-		break;
-
+	for (;;) {
+		if (bdp->cbd_sc & BD_ENET_RX_EMPTY)
+			break;
+
+#define RX_BD_ERRORS (BD_ENET_RX_LG | BD_ENET_RX_SH | BD_ENET_RX_NO | BD_ENET_RX_CR | BD_ENET_RX_OV | BD_ENET_RX_CL)
 #ifndef final_version
-	/* Since we have allocated space to hold a complete frame, both
-	 * the first and last indicators should be set.
-	 */
-	if ((bdp->cbd_sc & (BD_ENET_RX_FIRST | BD_ENET_RX_LAST)) !=
-		(BD_ENET_RX_FIRST | BD_ENET_RX_LAST))
+		/* Since we have allocated space to hold a complete frame, both
+		 * the first and last indicators should be set.
+		 */
+		if ((bdp->cbd_sc & (BD_ENET_RX_FIRST | BD_ENET_RX_LAST)) !=
+		    (BD_ENET_RX_FIRST | BD_ENET_RX_LAST))
 			printk("CPM ENET: rcv is not first+last\n");
 #endif
+		if(bdp->cbd_sc & RX_BD_ERRORS){ /* Receive errors ? */
+			cep->stats.rx_errors++;
+			if (bdp->cbd_sc & (BD_ENET_RX_LG | BD_ENET_RX_SH)) /* Frame too long or too short. */
+				cep->stats.rx_length_errors++;
+			if (bdp->cbd_sc & BD_ENET_RX_NO)	/* Frame alignment */
+				cep->stats.rx_frame_errors++;
+			if (bdp->cbd_sc & BD_ENET_RX_CR)	/* CRC Error */
+				cep->stats.rx_crc_errors++;
+			if (bdp->cbd_sc & BD_ENET_RX_OV)	/* FIFO overrun */
+				cep->stats.rx_fifo_errors++;
+			if (bdp->cbd_sc & BD_ENET_RX_CL)        /* Late collision */
+				cep->stats.collisions++;
+		} else {
+			/* Process the incoming frame.
+			 */
+			cep->stats.rx_packets++;
+			pkt_len = bdp->cbd_datlen;
+			cep->stats.rx_bytes += pkt_len;
+			pkt_len -= 4; /* The packet length includes FCS, but we don't want to
+				       * include that when passing upstream as it messes up
+				       * bridging applications. Is this still true ???? */
+#ifdef COPY_SMALL_FRAMES
+			/* Allocate the next buffer now so we are sure to have one when needed
+			 * This does 16 byte alignment, exactly what we need(L1_CACHE aligned). */
+			if(pkt_len < RX_COPYBREAK)
+				skb_tmp = __dev_alloc_skb(pkt_len, GFP_ATOMIC | GFP_DMA);
+			else
+#endif
+				skb_tmp = __dev_alloc_skb(CPM_ENET_RX_FRSIZE, GFP_ATOMIC | GFP_DMA);

-	/* Frame too long or too short.
-	*/
-	if (bdp->cbd_sc & (BD_ENET_RX_LG | BD_ENET_RX_SH))
-		cep->stats.rx_length_errors++;
-	if (bdp->cbd_sc & BD_ENET_RX_NO)	/* Frame alignment */
-		cep->stats.rx_frame_errors++;
-	if (bdp->cbd_sc & BD_ENET_RX_CR)	/* CRC Error */
-		cep->stats.rx_crc_errors++;
-	if (bdp->cbd_sc & BD_ENET_RX_OV)	/* FIFO overrun */
-		cep->stats.rx_crc_errors++;
-
-	/* Report late collisions as a frame error.
-	 * On this error, the BD is closed, but we don't know what we
-	 * have in the buffer.  So, just drop this frame on the floor.
-	 */
-	if (bdp->cbd_sc & BD_ENET_RX_CL) {
-		cep->stats.rx_frame_errors++;
-	}
-	else {
-
-		/* Process the incoming frame.
-		*/
-		cep->stats.rx_packets++;
-		pkt_len = bdp->cbd_datlen;
-		cep->stats.rx_bytes += pkt_len;
-
-		/* This does 16 byte alignment, much more than we need.
-		 * The packet length includes FCS, but we don't want to
-		 * include that when passing upstream as it messes up
-		 * bridging applications.
-		 */
-		skb = dev_alloc_skb(pkt_len-4);
-
-		if (skb == NULL) {
-			printk("%s: Memory squeeze, dropping packet.\n", dev->name);
-			cep->stats.rx_dropped++;
-		}
-		else {
-			skb->dev = dev;
-			skb_put(skb,pkt_len-4);	/* Make room */
-			eth_copy_and_sum(skb,
-				cep->rx_vaddr[bdp - cep->rx_bd_base],
-				pkt_len-4, 0);
-			skb->protocol=eth_type_trans(skb,dev);
-			netif_rx(skb);
+			if (skb_tmp == NULL) {
+				printk("%s: Memory squeeze, dropping packet.\n", dev->name);
+				cep->stats.rx_dropped++;
+
+			} else {
+				skb = cep->rx_vaddr[bdp - cep->rx_bd_base];
+#ifdef COPY_SMALL_FRAMES
+				if(pkt_len < RX_COPYBREAK) {
+					typeof(skb) skb_swap = skb;
+					memcpy(skb_put(skb_tmp, pkt_len), skb->data, pkt_len);
+					/* swap the skb and skb_tmp */
+					skb = skb_tmp;
+					skb_tmp = skb_swap;
+				}
+				else
+#endif
+				{
+					skb_put(skb, pkt_len);  /* Make room */
+					bdp->cbd_bufaddr = __pa(skb_tmp->data);
+					cep->rx_vaddr[bdp - cep->rx_bd_base] = skb_tmp;
+				}
+				dma_cache_inv((unsigned long) skb_tmp->data, CPM_ENET_RX_FRSIZE);
+				skb->dev = dev;
+				skb->protocol=eth_type_trans(skb, dev);
+				netif_rx(skb);
+			}
 		}
-	}

-	/* Clear the status flags for this buffer.
-	*/
-	bdp->cbd_sc &= ~BD_ENET_RX_STATS;
+		/* Clear the status flags for this buffer.
+		 */
+		bdp->cbd_sc &= ~BD_ENET_RX_STATS;

-	/* Mark the buffer empty.
-	*/
-	bdp->cbd_sc |= BD_ENET_RX_EMPTY;
+		/* Mark the buffer empty.
+		 */
+		bdp->cbd_sc |= BD_ENET_RX_EMPTY;

-	/* Update BD pointer to next entry.
-	*/
-	if (bdp->cbd_sc & BD_ENET_RX_WRAP)
-		bdp = cep->rx_bd_base;
-	else
-		bdp++;
+		/* Update BD pointer to next entry.
+		 */
+		if (bdp->cbd_sc & BD_ENET_RX_WRAP)
+			bdp = cep->rx_bd_base;
+		else
+			bdp++;

-   }
+  	}
 	cep->cur_rx = (cbd_t *)bdp;

 	return 0;
@@ -647,8 +660,7 @@
 	struct net_device *dev;
 	struct scc_enet_private *cep;
 	int i, j, k;
-	unsigned char	*eap, *ba;
-	dma_addr_t	mem_addr;
+	unsigned char	*eap;
 	bd_t		*bd;
 	volatile	cbd_t		*bdp;
 	volatile	cpm8xx_t	*cp;
@@ -839,22 +851,14 @@

 	bdp = cep->rx_bd_base;
 	k = 0;
-	for (i=0; i<CPM_ENET_RX_PAGES; i++) {
-
-		/* Allocate a page.
-		*/
-		ba = (unsigned char *)consistent_alloc(GFP_KERNEL, PAGE_SIZE, &mem_addr);
-
-		/* Initialize the BD for every fragment in the page.
-		*/
-		for (j=0; j<CPM_ENET_RX_FRPPG; j++) {
-			bdp->cbd_sc = BD_ENET_RX_EMPTY | BD_ENET_RX_INTR;
-			bdp->cbd_bufaddr = mem_addr;
-			cep->rx_vaddr[k++] = ba;
-			mem_addr += CPM_ENET_RX_FRSIZE;
-			ba += CPM_ENET_RX_FRSIZE;
-			bdp++;
-		}
+	/* Initialize the BDs. */
+	for (j=0; j < RX_RING_SIZE; j++) {
+		struct	sk_buff * skb = __dev_alloc_skb(CPM_ENET_RX_FRSIZE, GFP_ATOMIC | GFP_DMA);
+		dma_cache_inv((unsigned long) skb->data, CPM_ENET_RX_FRSIZE);
+		bdp->cbd_sc = BD_ENET_RX_EMPTY | BD_ENET_RX_INTR;
+		bdp->cbd_bufaddr = __pa(skb->data);
+		cep->rx_vaddr[k++] = skb;
+		bdp++;
 	}

 	/* Set the last buffer to wrap.

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Joakim Tjernlund]

> Hi guys,
>
> I have created a patch that applies cleanly to the head of
> linuxppc_2_4_devel
> and it works great.
>
> Keep up the good work!
>
> Pantelis

Hi Pantelis

I don't follow you here. Didn't my patch apply cleanly against linuxppc_2_4_devel?
I generated my patch against linuxppc_2_4_devel(I think).

    Jocke


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Pantelis Antoniou]

[-- Attachment #1: Type: text/plain, Size: 496 bytes --]

Joakim Tjernlund wrote:

>>Hi guys,
>>
>>I have created a patch that applies cleanly to the head of
>>linuxppc_2_4_devel
>>and it works great.
>>
>>Keep up the good work!
>>
>>Pantelis
>>
>>
>
>Hi Pantelis
>
>I don't follow you here. Didn't my patch apply cleanly against linuxppc_2_4_devel?
>I generated my patch against linuxppc_2_4_devel(I think).
>
>    Jocke
>
>
>
>
>
Nope.

Fails at hunks #7 and #8.
Don't worry, it's trivial stuff.

Here is enet.c.rej, if you want to check it.

Regards


[-- Attachment #2: enet.c.rej --]
[-- Type: text/plain, Size: 5586 bytes --]

***************
*** 458,540 ****
  	 */
  	bdp = cep->cur_rx;

- for (;;) {
- 	if (bdp->cbd_sc & BD_ENET_RX_EMPTY)
- 		break;
-
  #ifndef final_version
- 	/* Since we have allocated space to hold a complete frame, both
- 	 * the first and last indicators should be set.
- 	 */
- 	if ((bdp->cbd_sc & (BD_ENET_RX_FIRST | BD_ENET_RX_LAST)) !=
- 		(BD_ENET_RX_FIRST | BD_ENET_RX_LAST))
  			printk("CPM ENET: rcv is not first+last\n");
  #endif
-
- 	/* Frame too long or too short.
- 	*/
- 	if (bdp->cbd_sc & (BD_ENET_RX_LG | BD_ENET_RX_SH))
- 		cep->stats.rx_length_errors++;
- 	if (bdp->cbd_sc & BD_ENET_RX_NO)	/* Frame alignment */
- 		cep->stats.rx_frame_errors++;
- 	if (bdp->cbd_sc & BD_ENET_RX_CR)	/* CRC Error */
- 		cep->stats.rx_crc_errors++;
- 	if (bdp->cbd_sc & BD_ENET_RX_OV)	/* FIFO overrun */
- 		cep->stats.rx_crc_errors++;
-
- 	/* Report late collisions as a frame error.
- 	 * On this error, the BD is closed, but we don't know what we
- 	 * have in the buffer.  So, just drop this frame on the floor.
- 	 */
- 	if (bdp->cbd_sc & BD_ENET_RX_CL) {
- 		cep->stats.rx_frame_errors++;
- 	}
- 	else {
-
- 		/* Process the incoming frame.
- 		*/
- 		cep->stats.rx_packets++;
- 		pkt_len = bdp->cbd_datlen;
- 		cep->stats.rx_bytes += pkt_len;
-
- 		/* This does 16 byte alignment, much more than we need.
- 		 * The packet length includes FCS, but we don't want to
- 		 * include that when passing upstream as it messes up
- 		 * bridging applications.
- 		 */
- 		skb = dev_alloc_skb(pkt_len-4);
-
- 		if (skb == NULL) {
- 			printk("%s: Memory squeeze, dropping packet.\n", dev->name);
- 			cep->stats.rx_dropped++;
- 		}
- 		else {
- 			skb->dev = dev;
- 			skb_put(skb,pkt_len-4);	/* Make room */
- 			eth_copy_and_sum(skb,
- 				cep->rx_vaddr[bdp - cep->rx_bd_base],
- 				pkt_len-4, 0);
- 			skb->protocol=eth_type_trans(skb,dev);
- 			netif_rx(skb);
  		}
  	}
-
- 	/* Clear the status flags for this buffer.
- 	*/
- 	bdp->cbd_sc &= ~BD_ENET_RX_STATS;
-
- 	/* Mark the buffer empty.
- 	*/
- 	bdp->cbd_sc |= BD_ENET_RX_EMPTY;
-
- 	/* Update BD pointer to next entry.
- 	*/
- 	if (bdp->cbd_sc & BD_ENET_RX_WRAP)
- 		bdp = cep->rx_bd_base;
- 	else
- 		bdp++;
-
-    }
  	cep->cur_rx = (cbd_t *)bdp;

  	return 0;
--- 464,556 ----
  	 */
  	bdp = cep->cur_rx;

+ 	for (;;) {
+ 		if (bdp->cbd_sc & BD_ENET_RX_EMPTY)
+ 			break;
+
+ #define RX_BD_ERRORS (BD_ENET_RX_LG | BD_ENET_RX_SH | BD_ENET_RX_NO | BD_ENET_RX_CR | BD_ENET_RX_OV | BD_ENET_RX_CL)
  #ifndef final_version
+ 		/* Since we have allocated space to hold a complete frame, both
+ 		 * the first and last indicators should be set.
+ 		 */
+ 		if ((bdp->cbd_sc & (BD_ENET_RX_FIRST | BD_ENET_RX_LAST)) !=
+ 		    (BD_ENET_RX_FIRST | BD_ENET_RX_LAST))
  			printk("CPM ENET: rcv is not first+last\n");
  #endif
+ 		if(bdp->cbd_sc & RX_BD_ERRORS){ /* Receive errors ? */
+ 			cep->stats.rx_errors++;
+ 			if (bdp->cbd_sc & (BD_ENET_RX_LG | BD_ENET_RX_SH)) /* Frame too long or too short. */
+ 				cep->stats.rx_length_errors++;
+ 			if (bdp->cbd_sc & BD_ENET_RX_NO)	/* Frame alignment */
+ 				cep->stats.rx_frame_errors++;
+ 			if (bdp->cbd_sc & BD_ENET_RX_CR)	/* CRC Error */
+ 				cep->stats.rx_crc_errors++;
+ 			if (bdp->cbd_sc & BD_ENET_RX_OV)	/* FIFO overrun */
+ 				cep->stats.rx_fifo_errors++;
+ 			if (bdp->cbd_sc & BD_ENET_RX_CL)        /* Late collision */
+ 				cep->stats.collisions++;
+ 		} else {
+ 			/* Process the incoming frame.
+ 			 */
+ 			cep->stats.rx_packets++;
+ 			pkt_len = bdp->cbd_datlen;
+ 			cep->stats.rx_bytes += pkt_len;
+ 			pkt_len -= 4; /* The packet length includes FCS, but we don't want to
+ 				       * include that when passing upstream as it messes up
+ 				       * bridging applications. Is this still true ???? */
+ #ifdef COPY_SMALL_FRAMES
+ 			/* Allocate the next buffer now so we are sure to have one when needed
+ 			 * This does 16 byte alignment, exactly what we need(L1_CACHE aligned). */
+ 			if(pkt_len < RX_COPYBREAK)
+ 				skb_tmp = __dev_alloc_skb(pkt_len, GFP_ATOMIC | GFP_DMA);
+ 			else
+ #endif
+ 				skb_tmp = __dev_alloc_skb(CPM_ENET_RX_FRSIZE, GFP_ATOMIC | GFP_DMA);
+
+ 			if (skb_tmp == NULL) {
+ 				printk("%s: Memory squeeze, dropping packet.\n", dev->name);
+ 				cep->stats.rx_dropped++;
+
+ 			} else {
+ 				skb = cep->rx_vaddr[bdp - cep->rx_bd_base];
+ #ifdef COPY_SMALL_FRAMES
+ 				if(pkt_len < RX_COPYBREAK) {
+ 					typeof(skb) skb_swap = skb;
+ 					memcpy(skb_put(skb_tmp, pkt_len), skb->data, pkt_len);
+ 					/* swap the skb and skb_tmp */
+ 					skb = skb_tmp;
+ 					skb_tmp = skb_swap;
+ 				}
+ 				else
+ #endif
+ 				{
+ 					skb_put(skb, pkt_len);  /* Make room */
+ 					bdp->cbd_bufaddr = __pa(skb_tmp->data);
+ 					cep->rx_vaddr[bdp - cep->rx_bd_base] = skb_tmp;
+ 				}
+ 				dma_cache_inv((unsigned long) skb_tmp->data, CPM_ENET_RX_FRSIZE);
+ 				skb->dev = dev;
+ 				skb->protocol=eth_type_trans(skb, dev);
+ 				netif_rx(skb);
+ 			}
  		}
+
+ 		/* Clear the status flags for this buffer.
+ 		 */
+ 		bdp->cbd_sc &= ~BD_ENET_RX_STATS;
+
+ 		/* Mark the buffer empty.
+ 		 */
+ 		bdp->cbd_sc |= BD_ENET_RX_EMPTY;
+
+ 		/* Update BD pointer to next entry.
+ 		 */
+ 		if (bdp->cbd_sc & BD_ENET_RX_WRAP)
+ 			bdp = cep->rx_bd_base;
+ 		else
+ 			bdp++;
+
  	}
  	cep->cur_rx = (cbd_t *)bdp;

  	return 0;
***************
*** 608,614 ****

  			dmi = dev->mc_list;

- 			for (i=0; i<dev->mc_count; i++) {

  				/* Only support group multicast for now.
  				*/
--- 624,630 ----

  			dmi = dev->mc_list;

+ 			for (i=0; i<dev->mc_count; i++, dmi = dmi->next) {

  				/* Only support group multicast for now.
  				*/

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Joakim Tjernlund]

> >Hi Pantelis
> >
> >I don't follow you here. Didn't my patch apply cleanly against linuxppc_2_4_devel?
> >I generated my patch against linuxppc_2_4_devel(I think).
> >
> >    Jocke
> Nope.
>
> Fails at hunks #7 and #8.
> Don't worry, it's trivial stuff.
>
> Here is enet.c.rej, if you want to check it.

hmm, something is broken here. I can apply my patch to my local copy of
8xx_io/enet.c, BK version 1.24 just fine. Whats your version of enet.c?

   Jocke


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Pantelis Antoniou]

Joakim Tjernlund wrote:

>>>Hi Pantelis
>>>
>>>I don't follow you here. Didn't my patch apply cleanly against linuxppc_2_4_devel?
>>>I generated my patch against linuxppc_2_4_devel(I think).
>>>
>>>   Jocke
>>>
>>Nope.
>>
>>Fails at hunks #7 and #8.
>>Don't worry, it's trivial stuff.
>>
>>Here is enet.c.rej, if you want to check it.
>>
>
>hmm, something is broken here. I can apply my patch to my local copy of
>8xx_io/enet.c, BK version 1.24 just fine. Whats your version of enet.c?
>
>   Jocke
>
>
>
>
My version is the head of the BK tree, updated today.

Regards

Pantelis


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Joakim Tjernlund]

> >hmm, something is broken here. I can apply my patch to my local copy of
> >8xx_io/enet.c, BK version 1.24 just fine. Whats your version of enet.c?
> >
> >   Jocke
> >
> >
> >
> >
> My version is the head of the BK tree, updated today.

What version is that?

 Jocke


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Pantelis Antoniou]

Joakim Tjernlund wrote:

>>>hmm, something is broken here. I can apply my patch to my local copy of
>>>8xx_io/enet.c, BK version 1.24 just fine. Whats your version of enet.c?
>>>
>>>  Jocke
>>>
>>>
>>>
>>>
>>>
>>>
>>My version is the head of the BK tree, updated today.
>>
>>
>
>What version is that?
>
> Jocke
>
>
>
>
>
>
>
2.4.21-pre4

Regards

Pantelis


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Steven Scholz]

Hi there,

has someone actually ported Jocke's patch to 8xx FEC !?

Any benchmarks?

Thanks,

Steven


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Joakim Tjernlund]

> has someone actually ported Jocke's patch to 8xx FEC !?

I just did a port, but the patch is on my home computer. I can send
it later ...

It's not tested at all since I don't have the FEC connected on my boards.


 Jocke


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Joakim Tjernlund]

> Hi there,
>
> has someone actually ported Jocke's patch to 8xx FEC !?
>
> Any benchmarks?
>
> Thanks,
>
> Steven

Actually I had a copy at work so here it goes. Not tested(not even compiled)

 Jocke

--- arch/ppc/8xx_io/fec.c	Fri Nov  1 14:44:05 2002
+++ arch/ppc/8xx_io/new_fec.c	Sun Feb  9 14:24:45 2003
@@ -83,26 +83,34 @@
 } phy_info_t;
 #endif	/* CONFIG_USE_MDIO */

+/* Define COPY_SMALL_FRAMES if you want to save buffer memory for small packets
+ * at a small performance hit. Note performance testing needed */
+/* #define COPY_SMALL_FRAMES 1  */
+
+#ifdef COPY_SMALL_FRAMES
+  #define RX_COPYBREAK (256-16) /* dev_alloc_skb() adds 16 bytes for internal use */
+#endif
+
+
 /* The number of Tx and Rx buffers.  These are allocated from the page
  * pool.  The code may assume these are power of two, so it is best
  * to keep them that size.
  * We don't need to allocate pages for the transmitter.  We just use
  * the skbuffer directly.
  */
+
 #ifdef CONFIG_ENET_BIG_BUFFERS
-#define FEC_ENET_RX_PAGES	16
-#define FEC_ENET_RX_FRSIZE	2048
-#define FEC_ENET_RX_FRPPG	(PAGE_SIZE / FEC_ENET_RX_FRSIZE)
-#define RX_RING_SIZE		(FEC_ENET_RX_FRPPG * FEC_ENET_RX_PAGES)
-#define TX_RING_SIZE		16	/* Must be power of two */
-#define TX_RING_MOD_MASK	15	/*   for this to work */
+  #define RX_RING_SIZE		32
+  #define TX_RING_SIZE		16	/* Must be power of two for this to work */
 #else
-#define FEC_ENET_RX_PAGES	4
-#define FEC_ENET_RX_FRSIZE	2048
-#define FEC_ENET_RX_FRPPG	(PAGE_SIZE / FEC_ENET_RX_FRSIZE)
-#define RX_RING_SIZE		(FEC_ENET_RX_FRPPG * FEC_ENET_RX_PAGES)
-#define TX_RING_SIZE		8	/* Must be power of two */
-#define TX_RING_MOD_MASK	7	/*   for this to work */
+  #define RX_RING_SIZE		8
+  #define TX_RING_SIZE		8	/* Must be power of two for this to work */
+#endif
+#define TX_RING_MOD_MASK	(TX_RING_SIZE-1)
+
+#define CPM_ENET_RX_FRSIZE	1552 /* must be a multiple of cache line */
+#if CPM_ENET_RX_FRSIZE % L1_CACHE_LINE_SIZE != 0
+    #error CPM_ENET_RX_FRSIZE must be a multiple of L1 cache size
 #endif

 /* Interrupt events/masks.
@@ -573,7 +581,7 @@

 	/* Check for errors. */
 	if (bdp->cbd_sc & (BD_ENET_RX_LG | BD_ENET_RX_SH | BD_ENET_RX_NO |
-			   BD_ENET_RX_CR | BD_ENET_RX_OV)) {
+			   BD_ENET_RX_CR | BD_ENET_RX_OV | BD_ENET_RX_CL)) {
 		fep->stats.rx_errors++;
 		if (bdp->cbd_sc & (BD_ENET_RX_LG | BD_ENET_RX_SH)) {
 		/* Frame too long or too short. */
@@ -585,41 +593,53 @@
 			fep->stats.rx_crc_errors++;
 		if (bdp->cbd_sc & BD_ENET_RX_OV)	/* FIFO overrun */
 			fep->stats.rx_crc_errors++;
-	}
-
-	/* Report late collisions as a frame error.
-	 * On this error, the BD is closed, but we don't know what we
-	 * have in the buffer.  So, just drop this frame on the floor.
-	 */
-	if (bdp->cbd_sc & BD_ENET_RX_CL) {
-		fep->stats.rx_errors++;
-		fep->stats.rx_frame_errors++;
-		goto rx_processing_done;
-	}
-
-	/* Process the incoming frame.
-	 */
-	fep->stats.rx_packets++;
-	pkt_len = bdp->cbd_datlen;
-	fep->stats.rx_bytes += pkt_len;
-	data = fep->rx_vaddr[bdp - fep->rx_bd_base];
-
-	/* This does 16 byte alignment, exactly what we need.
-	 * The packet length includes FCS, but we don't want to
-	 * include that when passing upstream as it messes up
-	 * bridging applications.
-	 */
-	skb = dev_alloc_skb(pkt_len-4);
-
-	if (skb == NULL) {
-		printk("%s: Memory squeeze, dropping packet.\n", dev->name);
-		fep->stats.rx_dropped++;
+		if (bdp->cbd_sc & BD_ENET_RX_CL)	/* Late collision */
+			fep->stats.rx_frame_errors++;	/* Report as a frame error. */
 	} else {
-		skb->dev = dev;
-		skb_put(skb,pkt_len-4);	/* Make room */
-		eth_copy_and_sum(skb, data, pkt_len-4, 0);
-		skb->protocol=eth_type_trans(skb,dev);
-		netif_rx(skb);
+
+		/* Process the incoming frame.
+		 */
+		fep->stats.rx_packets++;
+		pkt_len = bdp->cbd_datlen;
+		fep->stats.rx_bytes += pkt_len;
+		pkt_len -= 4; /* The packet length includes FCS, but we don't want to
+			       * include that when passing upstream as it messes up
+			       * bridging applications. Is this still true ???? */
+#ifdef COPY_SMALL_FRAMES
+		/* Allocate the next buffer now so we are sure to have one when needed
+		 * This does 16 byte alignment, exactly what we need(L1_CACHE aligned). */
+		if(pkt_len < RX_COPYBREAK)
+			skb_tmp = __dev_alloc_skb(pkt_len, GFP_ATOMIC | GFP_DMA);
+		else
+#endif
+			skb_tmp = __dev_alloc_skb(CPM_ENET_RX_FRSIZE, GFP_ATOMIC | GFP_DMA);
+
+		if (skb_tmp == NULL) {
+			printk("%s: Memory squeeze, dropping packet.\n", dev->name);
+			fep->stats.rx_dropped++;
+
+		} else {
+			skb = fep->rx_vaddr[bdp - fep->rx_bd_base];
+#ifdef COPY_SMALL_FRAMES
+			if(pkt_len < RX_COPYBREAK) {
+				typeof(skb) skb_swap = skb;
+				memcpy(skb_put(skb_tmp, pkt_len), skb->data, pkt_len);
+				/* swap the skb and skb_tmp */
+				skb = skb_tmp;
+				skb_tmp = skb_swap;
+			}
+			else
+#endif
+			{
+				skb_put(skb, pkt_len);  /* Make room */
+				bdp->cbd_bufaddr = __pa(skb_tmp->data);
+				fep->rx_vaddr[bdp - fep->rx_bd_base] = skb_tmp;
+			}
+			dma_cache_inv((unsigned long) skb_tmp->data, CPM_ENET_RX_FRSIZE);
+			skb->dev = dev;
+			skb->protocol=eth_type_trans(skb, dev);
+			netif_rx(skb);
+		}
 	}
   rx_processing_done:

@@ -1662,7 +1682,7 @@

 			dmi = dev->mc_list;

-			for (i=0; i<dev->mc_count; i++) {
+			for (i=0; i<dev->mc_count; i++, dmi = dmi->next) {

 				/* Only support group multicast for now.
 				*/
@@ -1698,8 +1718,7 @@
 	struct net_device *dev;
 	struct fec_enet_private *fep;
 	int i, j, k;
-	unsigned char	*eap, *iap, *ba;
-	unsigned long	mem_addr;
+	unsigned char	*eap, *iap;
 	volatile	cbd_t	*bdp;
 	cbd_t		*cbd_base;
 	volatile	immap_t	*immap;
@@ -1783,24 +1802,15 @@
 	*/
 	bdp = fep->rx_bd_base;
 	k = 0;
-	for (i=0; i<FEC_ENET_RX_PAGES; i++) {
-
-		/* Allocate a page.
-		*/
-		ba = (unsigned char *)consistent_alloc(GFP_KERNEL, PAGE_SIZE, &mem_addr);
-
-		/* Initialize the BD for every fragment in the page.
-		*/
-		for (j=0; j<FEC_ENET_RX_FRPPG; j++) {
-			bdp->cbd_sc = BD_ENET_RX_EMPTY;
-			bdp->cbd_bufaddr = mem_addr;
-			fep->rx_vaddr[k++] = ba;
-			mem_addr += FEC_ENET_RX_FRSIZE;
-			ba += FEC_ENET_RX_FRSIZE;
-			bdp++;
-		}
+	/* Initialize the BDs. */
+	for (j=0; j < RX_RING_SIZE; j++) {
+		struct	sk_buff * skb = __dev_alloc_skb(CPM_ENET_RX_FRSIZE, GFP_ATOMIC | GFP_DMA);
+		dma_cache_inv((unsigned long) skb->data, CPM_ENET_RX_FRSIZE);
+		bdp->cbd_sc = BD_ENET_RX_EMPTY | BD_ENET_RX_INTR;
+		bdp->cbd_bufaddr = __pa(skb->data);
+		fep->rx_vaddr[k++] = skb;
+		bdp++;
 	}
-
 	/* Set the last buffer to wrap.
 	*/
 	bdp--;


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Stephan Linke]

Hi Jocke,

I am running a 2.4.18 kernel. Maybe that's the problem?
But let me guess: keeping the invalidate_dcache_range() would be fine as well for our 8xx board?

Stephan

> -----Original Message-----
> From: Joakim Tjernlund [mailto:joakim.tjernlund@lumentis.se]
> Sent: Dienstag, 4. Februar 2003 10:59
> To: Stephan Linke
> Cc: Linuxppc-Embedded
> Subject: RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3
>
>
> >
> > Hi,
> >
> > I am on an 862. Anyway I can't find another definition of dma_cache_inv() but the NO OP in asm-ppc/io.h. Could you give
> me e hint
> > where it is defined in your kernel?
>
> It's in asm-ppc/io.h (2.4.20) and there are 2 definitions of dma_cache_inv() which depends on
> CONFIG_NOT_COHERENT_CACHE(should be defined for 8xx). What kernel version are you running?
>
>  Jocke
> >
> > Thanks, Stephan
> >
> > > -----Original Message-----
> > > From: Joakim Tjernlund [mailto:joakim.tjernlund@lumentis.se]
> > > Sent: Montag, 3. Februar 2003 18:23
> > > To: Stephan Linke
> > > Subject: RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3
> > >
> > >
> > > > Hi Jocke,
> > > >
> > > > in your latest patch you are using dma_cache_inv() instead of invalidate_dcache_range().
> > > > The only dma_cache_inv() I can find is in include/asm-ppc/io.h. and it's a "do{}while (0)".
> > > > Are you shure that this was your intention? It seames to me like you could remove that call as well.
> > >
> > > I guess you are on 8260? On 8260 there is no need for invalidate_dcache_range() since
> > > it's the CPM is cache coherent. On 8xx it is not cache coherent. I switched
> > > to dma_cache_inv() because it's a no op on 8260 and a invalidate_dcache_range() on 8xx so
> > > it would be easy to adapt the patch to both CPU's.
> > >
> > >  Jocke
> > >
> > >
> >
> >
>
>

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 2

[Joakim Tjernlund]

>
>
> On Thu, Oct 24, 2002 at 04:23:31PM +0200, Joakim Tjernlund wrote:
>
> > This is the second version of my patch that removes the expensive memcpy of
> > received
> > ethernet frames in interrupt context.
> >
> > I have 1 report(from Ricardo Scop) of 20% increase in packets/second, packet
> > size 1500 when
> > applied to 8260 FEC(needs to be applied manually). But min packet size
> > decreased with 10 %.
> > This version should fix the 10% decrease case.
> >
> > This patch could be adapted 8xx_io/fec.c and 8260_io/enet.c and
> > 8260/fcc_enet.c with little effort.
> >
> > Better fix a bug in set_multicast_list(), move the dmi list forward when
> >     walking it(dmi = dmi->next;)
> >
> > New stuff:
> >    - Configrable: copy small packets or pass them directly, see
> > COPY_SMALL_FRAMES in code.
> >    - Collision reporting fix form Thomas Lange.
> >    - Don't pass receive frames which has error upwards.
> >    - Report RX_OV errors as fifo errors, not crc errors.
> >
> > Please test and report any problems and performace improvements.
>
> All of this sounds good, but can you please break this one patch up into
> seperate logical patches?

I rather not.

The logical pices in this patch are:
 1) remove the expensive memcpy.
   - This is the main part of the patch.

 2) dmi->next fix, which is a one liner.
    Don't pass receive frames which has error upwards.
   - Trival stuff.

 3) Collision reporting fix form Thomas Lange.
    Report RX_OV errors as fifo errors, not crc errors.
   - These are just minor accounting fixes.

So I don't think it's worth the effort. Let's just wait a few days and see
if it was success or not.

Folks, please report you findings to the list.

    Jocke


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

[PATCH] arch/ppc/8xx_io/enet.c, version 3

[Joakim Tjernlund]

> This is the second version of my patch that removes the expensive memcpy of
> received
> ethernet frames in interrupt context.
>
> I have 1 report(from Ricardo Scop) of 20% increase in packets/second, packet
> size 1500 when
> applied to 8260 FEC(needs to be applied manually). But min packet size
> decreased with 10 %.
> This version should fix the 10% decrease case.
>
> This patch could be adapted 8xx_io/fec.c and 8260_io/enet.c and
> 8260/fcc_enet.c with little effort.
>
> Better fix a bug in set_multicast_list(), move the dmi list forward when
>     walking it(dmi = dmi->next;)
>
> New stuff:
>    - Configrable: copy small packets or pass them directly, see
> COPY_SMALL_FRAMES in code.
>    - Collision reporting fix form Thomas Lange.
>    - Don't pass receive frames which has error upwards.
>    - Report RX_OV errors as fifo errors, not crc errors.
>
> Please test and report any problems and performace improvements.

Hi

This is the third version of my optimized enet.c patch.
Changes since version 2:
  1) invalidate the whole buffer BEFORE it is given to he CPM. Previously
     it was invalidated after the packet was received and that could lead to buffer
     corruption in some cases.

  2) use dma_cache_inv() instead of invalidate_dcache_range() since that will work
     for both 8xx and 82xx.

  3) decrease the allocated buffer length.

  4) disabled COPY_SMALL_FRAME. Define it somewhere if you want to save some memory.

  5) probably some white space changes got in too.

Any chance to see it the devel tree?

More than 3 months has passed since version 2 and the only problem reported was
1) and the fix has been know since mid November.

Dan, you said you would integrate this patch(or some version of it) in November. I
think I have waited long enough now. Please do ASAP.

 Jocke

--- arch/ppc/8xx_io/enet.c	Fri Nov  1 14:44:05 2002
+++ arch/ppc/8xx_io/new_enet.c	Sat Jan 25 19:57:50 2003
@@ -34,7 +34,6 @@
 #include <linux/ioport.h>
 #include <linux/slab.h>
 #include <linux/interrupt.h>
-#include <linux/pci.h>
 #include <linux/init.h>
 #include <linux/delay.h>
 #include <linux/netdevice.h>
@@ -86,6 +85,14 @@
  * All functions are directly controlled using I/O pins.  See <asm/commproc.h>.
  */

+/* Define COPY_SMALL_FRAMES if you want to save buffer memory for small packets
+ * at a small performance hit. Note performance testing needed */
+/* #define COPY_SMALL_FRAMES 1  */
+
+#ifdef COPY_SMALL_FRAMES
+  #define RX_COPYBREAK (256-16) /* dev_alloc_skb() adds 16 bytes for internal use */
+#endif
+
 /* The transmitter timeout
  */
 #define TX_TIMEOUT	(2*HZ)
@@ -97,19 +104,17 @@
  * the skbuffer directly.
  */
 #ifdef CONFIG_ENET_BIG_BUFFERS
-#define CPM_ENET_RX_PAGES	32
-#define CPM_ENET_RX_FRSIZE	2048
-#define CPM_ENET_RX_FRPPG	(PAGE_SIZE / CPM_ENET_RX_FRSIZE)
-#define RX_RING_SIZE		(CPM_ENET_RX_FRPPG * CPM_ENET_RX_PAGES)
-#define TX_RING_SIZE		64	/* Must be power of two */
-#define TX_RING_MOD_MASK	63	/*   for this to work */
+  #define RX_RING_SIZE		64
+  #define TX_RING_SIZE		64	/* Must be power of two for this to work */
 #else
-#define CPM_ENET_RX_PAGES	4
-#define CPM_ENET_RX_FRSIZE	2048
-#define CPM_ENET_RX_FRPPG	(PAGE_SIZE / CPM_ENET_RX_FRSIZE)
-#define RX_RING_SIZE		(CPM_ENET_RX_FRPPG * CPM_ENET_RX_PAGES)
-#define TX_RING_SIZE		8	/* Must be power of two */
-#define TX_RING_MOD_MASK	7	/*   for this to work */
+  #define RX_RING_SIZE		8
+  #define TX_RING_SIZE		8	/* Must be power of two for this to work */
+#endif
+#define TX_RING_MOD_MASK	(TX_RING_SIZE-1)
+
+#define CPM_ENET_RX_FRSIZE	1552 /* must be a multiple of cache line */
+#if CPM_ENET_RX_FRSIZE % L1_CACHE_LINE_SIZE != 0
+    #error CPM_ENET_RX_FRSIZE must be a multiple of L1 cache size
 #endif

 /* The CPM stores dest/src/type, data, and checksum for receive packets.
@@ -143,7 +148,7 @@
 	/* Virtual addresses for the receive buffers because we can't
 	 * do a __va() on them anymore.
 	 */
-	unsigned char *rx_vaddr[RX_RING_SIZE];
+	void    *rx_vaddr[RX_RING_SIZE];
 	struct	net_device_stats stats;
 	uint	tx_full;
 	spinlock_t lock;
@@ -370,11 +375,11 @@

 		cep->stats.tx_packets++;

-		/* Deferred means some collisions occurred during transmit,
-		 * but we eventually sent the packet OK.
-		 */
-		if (bdp->cbd_sc & BD_ENET_TX_DEF)
-			cep->stats.collisions++;
+		/* Check retry counter, i.e. collision counter */
+		if (bdp->cbd_sc & BD_ENET_TX_RCMASK){
+			/* Note that counter cannot go higher than 15 */
+			cep->stats.collisions+=(bdp->cbd_sc & BD_ENET_TX_RCMASK)>>2;
+		}

 		/* Free the sk buffer associated with this last transmit.
 		*/
@@ -449,6 +454,7 @@
 	struct	scc_enet_private *cep;
 	volatile cbd_t	*bdp;
 	struct	sk_buff *skb;
+	struct	sk_buff *skb_tmp;
 	ushort	pkt_len;

 	cep = (struct scc_enet_private *)dev->priv;
@@ -458,83 +464,93 @@
 	 */
 	bdp = cep->cur_rx;

-for (;;) {
-	if (bdp->cbd_sc & BD_ENET_RX_EMPTY)
-		break;
-
+	for (;;) {
+		if (bdp->cbd_sc & BD_ENET_RX_EMPTY)
+			break;
+
+#define RX_BD_ERRORS (BD_ENET_RX_LG | BD_ENET_RX_SH | BD_ENET_RX_NO | BD_ENET_RX_CR | BD_ENET_RX_OV | BD_ENET_RX_CL)
 #ifndef final_version
-	/* Since we have allocated space to hold a complete frame, both
-	 * the first and last indicators should be set.
-	 */
-	if ((bdp->cbd_sc & (BD_ENET_RX_FIRST | BD_ENET_RX_LAST)) !=
-		(BD_ENET_RX_FIRST | BD_ENET_RX_LAST))
+		/* Since we have allocated space to hold a complete frame, both
+		 * the first and last indicators should be set.
+		 */
+		if ((bdp->cbd_sc & (BD_ENET_RX_FIRST | BD_ENET_RX_LAST)) !=
+		    (BD_ENET_RX_FIRST | BD_ENET_RX_LAST))
 			printk("CPM ENET: rcv is not first+last\n");
 #endif
-
-	/* Frame too long or too short.
-	*/
-	if (bdp->cbd_sc & (BD_ENET_RX_LG | BD_ENET_RX_SH))
-		cep->stats.rx_length_errors++;
-	if (bdp->cbd_sc & BD_ENET_RX_NO)	/* Frame alignment */
-		cep->stats.rx_frame_errors++;
-	if (bdp->cbd_sc & BD_ENET_RX_CR)	/* CRC Error */
-		cep->stats.rx_crc_errors++;
-	if (bdp->cbd_sc & BD_ENET_RX_OV)	/* FIFO overrun */
-		cep->stats.rx_crc_errors++;
-
-	/* Report late collisions as a frame error.
-	 * On this error, the BD is closed, but we don't know what we
-	 * have in the buffer.  So, just drop this frame on the floor.
-	 */
-	if (bdp->cbd_sc & BD_ENET_RX_CL) {
-		cep->stats.rx_frame_errors++;
-	}
-	else {
-
-		/* Process the incoming frame.
-		*/
-		cep->stats.rx_packets++;
-		pkt_len = bdp->cbd_datlen;
-		cep->stats.rx_bytes += pkt_len;
-
-		/* This does 16 byte alignment, much more than we need.
-		 * The packet length includes FCS, but we don't want to
-		 * include that when passing upstream as it messes up
-		 * bridging applications.
-		 */
-		skb = dev_alloc_skb(pkt_len-4);
-
-		if (skb == NULL) {
-			printk("%s: Memory squeeze, dropping packet.\n", dev->name);
-			cep->stats.rx_dropped++;
-		}
-		else {
-			skb->dev = dev;
-			skb_put(skb,pkt_len-4);	/* Make room */
-			eth_copy_and_sum(skb,
-				cep->rx_vaddr[bdp - cep->rx_bd_base],
-				pkt_len-4, 0);
-			skb->protocol=eth_type_trans(skb,dev);
-			netif_rx(skb);
+		if(bdp->cbd_sc & RX_BD_ERRORS){ /* Receive errors ? */
+			cep->stats.rx_errors++;
+			if (bdp->cbd_sc & (BD_ENET_RX_LG | BD_ENET_RX_SH)) /* Frame too long or too short. */
+				cep->stats.rx_length_errors++;
+			if (bdp->cbd_sc & BD_ENET_RX_NO)	/* Frame alignment */
+				cep->stats.rx_frame_errors++;
+			if (bdp->cbd_sc & BD_ENET_RX_CR)	/* CRC Error */
+				cep->stats.rx_crc_errors++;
+			if (bdp->cbd_sc & BD_ENET_RX_OV)	/* FIFO overrun */
+				cep->stats.rx_fifo_errors++;
+			if (bdp->cbd_sc & BD_ENET_RX_CL)        /* Late collision */
+				cep->stats.collisions++;
+		} else {
+			/* Process the incoming frame.
+			 */
+			cep->stats.rx_packets++;
+			pkt_len = bdp->cbd_datlen;
+			cep->stats.rx_bytes += pkt_len;
+			pkt_len -= 4; /* The packet length includes FCS, but we don't want to
+				       * include that when passing upstream as it messes up
+				       * bridging applications. Is this still true ???? */
+#ifdef COPY_SMALL_FRAMES
+			/* Allocate the next buffer now so we are sure to have one when needed
+			 * This does 16 byte alignment, exactly what we need(L1_CACHE aligned). */
+			if(pkt_len < RX_COPYBREAK)
+				skb_tmp = __dev_alloc_skb(pkt_len, GFP_ATOMIC | GFP_DMA);
+			else
+#endif
+				skb_tmp = __dev_alloc_skb(CPM_ENET_RX_FRSIZE, GFP_ATOMIC | GFP_DMA);
+
+			if (skb_tmp == NULL) {
+				printk("%s: Memory squeeze, dropping packet.\n", dev->name);
+				cep->stats.rx_dropped++;
+
+			} else {
+				skb = cep->rx_vaddr[bdp - cep->rx_bd_base];
+#ifdef COPY_SMALL_FRAMES
+				if(pkt_len < RX_COPYBREAK) {
+					typeof(skb) skb_swap = skb;
+					memcpy(skb_put(skb_tmp, pkt_len), skb->data, pkt_len);
+					/* swap the skb and skb_tmp */
+					skb = skb_tmp;
+					skb_tmp = skb_swap;
+				}
+				else
+#endif
+				{
+					skb_put(skb, pkt_len);  /* Make room */
+					bdp->cbd_bufaddr = __pa(skb_tmp->data);
+					cep->rx_vaddr[bdp - cep->rx_bd_base] = skb_tmp;
+				}
+				dma_cache_inv((unsigned long) skb_tmp->data, CPM_ENET_RX_FRSIZE);
+				skb->dev = dev;
+				skb->protocol=eth_type_trans(skb, dev);
+				netif_rx(skb);
+			}
 		}
+
+		/* Clear the status flags for this buffer.
+		 */
+		bdp->cbd_sc &= ~BD_ENET_RX_STATS;
+
+		/* Mark the buffer empty.
+		 */
+		bdp->cbd_sc |= BD_ENET_RX_EMPTY;
+
+		/* Update BD pointer to next entry.
+		 */
+		if (bdp->cbd_sc & BD_ENET_RX_WRAP)
+			bdp = cep->rx_bd_base;
+		else
+			bdp++;
+
 	}
-
-	/* Clear the status flags for this buffer.
-	*/
-	bdp->cbd_sc &= ~BD_ENET_RX_STATS;
-
-	/* Mark the buffer empty.
-	*/
-	bdp->cbd_sc |= BD_ENET_RX_EMPTY;
-
-	/* Update BD pointer to next entry.
-	*/
-	if (bdp->cbd_sc & BD_ENET_RX_WRAP)
-		bdp = cep->rx_bd_base;
-	else
-		bdp++;
-
-   }
 	cep->cur_rx = (cbd_t *)bdp;

 	return 0;
@@ -608,7 +624,7 @@

 			dmi = dev->mc_list;

-			for (i=0; i<dev->mc_count; i++) {
+			for (i=0; i<dev->mc_count; i++, dmi = dmi->next) {

 				/* Only support group multicast for now.
 				*/
@@ -647,8 +663,7 @@
 	struct net_device *dev;
 	struct scc_enet_private *cep;
 	int i, j, k;
-	unsigned char	*eap, *ba;
-	dma_addr_t	mem_addr;
+	unsigned char	*eap;
 	bd_t		*bd;
 	volatile	cbd_t		*bdp;
 	volatile	cpm8xx_t	*cp;
@@ -839,22 +854,14 @@

 	bdp = cep->rx_bd_base;
 	k = 0;
-	for (i=0; i<CPM_ENET_RX_PAGES; i++) {
-
-		/* Allocate a page.
-		*/
-		ba = (unsigned char *)consistent_alloc(GFP_KERNEL, PAGE_SIZE, &mem_addr);
-
-		/* Initialize the BD for every fragment in the page.
-		*/
-		for (j=0; j<CPM_ENET_RX_FRPPG; j++) {
-			bdp->cbd_sc = BD_ENET_RX_EMPTY | BD_ENET_RX_INTR;
-			bdp->cbd_bufaddr = mem_addr;
-			cep->rx_vaddr[k++] = ba;
-			mem_addr += CPM_ENET_RX_FRSIZE;
-			ba += CPM_ENET_RX_FRSIZE;
-			bdp++;
-		}
+	/* Initialize the BDs. */
+	for (j=0; j < RX_RING_SIZE; j++) {
+		struct	sk_buff * skb = __dev_alloc_skb(CPM_ENET_RX_FRSIZE, GFP_ATOMIC | GFP_DMA);
+		dma_cache_inv((unsigned long) skb->data, CPM_ENET_RX_FRSIZE);
+		bdp->cbd_sc = BD_ENET_RX_EMPTY | BD_ENET_RX_INTR;
+		bdp->cbd_bufaddr = __pa(skb->data);
+		cep->rx_vaddr[k++] = skb;
+		bdp++;
 	}

 	/* Set the last buffer to wrap.


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

RE: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Joakim Tjernlund]

> > This is the second version of my patch that removes the expensive memcpy of
> > received
> > ethernet frames in interrupt context.
> >
> > I have 1 report(from Ricardo Scop) of 20% increase in packets/second, packet
> > size 1500 when
> > applied to 8260 FEC(needs to be applied manually). But min packet size
> > decreased with 10 %.
> > This version should fix the 10% decrease case.
> >
> > This patch could be adapted 8xx_io/fec.c and 8260_io/enet.c and
> > 8260/fcc_enet.c with little effort.
> >
> > Better fix a bug in set_multicast_list(), move the dmi list forward when
> >     walking it(dmi = dmi->next;)
> >
> > New stuff:
> >    - Configrable: copy small packets or pass them directly, see
> > COPY_SMALL_FRAMES in code.
> >    - Collision reporting fix form Thomas Lange.
> >    - Don't pass receive frames which has error upwards.
> >    - Report RX_OV errors as fifo errors, not crc errors.
> >
> > Please test and report any problems and performace improvements.
>
> Hi
>
> This is the third version of my optimized enet.c patch.
> Changes since version 2:
>   1) invalidate the whole buffer BEFORE it is given to he CPM. Previously
>      it was invalidated after the packet was received and that could lead to buffer
>      corruption in some cases.
>
>   2) use dma_cache_inv() instead of invalidate_dcache_range() since that will work
>      for both 8xx and 82xx.
>
>   3) decrease the allocated buffer length.
>
>   4) disabled COPY_SMALL_FRAME. Define it somewhere if you want to save some memory.
>
>   5) probably some white space changes got in too.
>
> Any chance to see it the devel tree?
>
> More than 3 months has passed since version 2 and the only problem reported was
> 1) and the fix has been know since mid November.
>
> Dan, you said you would integrate this patch(or some version of it) in November. I
> think I have waited long enough now. Please do ASAP.
>
>  Jocke
[SNIP]

Zero feedback so far. Is there zero interest to improve the mpc code in linuxppc?
Anyhow, I give up. Don't have the energy to resend over and over again.

   Jocke


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: [PATCH] arch/ppc/8xx_io/enet.c, version 3

[Pantelis Antoniou]

Joakim Tjernlund wrote:

>>>This is the second version of my patch that removes the expensive memcpy of
>>>received
>>>ethernet frames in interrupt context.
>>>
>>>I have 1 report(from Ricardo Scop) of 20% increase in packets/second, packet
>>>size 1500 when
>>>applied to 8260 FEC(needs to be applied manually). But min packet size
>>>decreased with 10 %.
>>>This version should fix the 10% decrease case.
>>>
>>>This patch could be adapted 8xx_io/fec.c and 8260_io/enet.c and
>>>8260/fcc_enet.c with little effort.
>>>
>>>Better fix a bug in set_multicast_list(), move the dmi list forward when
>>>    walking it(dmi = dmi->next;)
>>>
>>>New stuff:
>>>   - Configrable: copy small packets or pass them directly, see
>>>COPY_SMALL_FRAMES in code.
>>>   - Collision reporting fix form Thomas Lange.
>>>   - Don't pass receive frames which has error upwards.
>>>   - Report RX_OV errors as fifo errors, not crc errors.
>>>
>>>Please test and report any problems and performace improvements.
>>>
>>>
>>Hi
>>
>>This is the third version of my optimized enet.c patch.
>>Changes since version 2:
>>  1) invalidate the whole buffer BEFORE it is given to he CPM. Previously
>>     it was invalidated after the packet was received and that could lead to buffer
>>     corruption in some cases.
>>
>>  2) use dma_cache_inv() instead of invalidate_dcache_range() since that will work
>>     for both 8xx and 82xx.
>>
>>  3) decrease the allocated buffer length.
>>
>>  4) disabled COPY_SMALL_FRAME. Define it somewhere if you want to save some memory.
>>
>>  5) probably some white space changes got in too.
>>
>>Any chance to see it the devel tree?
>>
>>More than 3 months has passed since version 2 and the only problem reported was
>>1) and the fix has been know since mid November.
>>
>>Dan, you said you would integrate this patch(or some version of it) in November. I
>>think I have waited long enough now. Please do ASAP.
>>
>> Jocke
>>
>>
>[SNIP]
>
>Zero feedback so far. Is there zero interest to improve the mpc code in linuxppc?
>Anyhow, I give up. Don't have the energy to resend over and over again.
>
>   Jocke
>
>
>
>
>
>
Please don't give up.

I'll try it and you'll have my feedback within the week.

Anyhow, what is the deal with 8xx/82xx?
At least the maintainers should ack the patch, and if it is not suitable
for inclusion they should state for what reason.

We don't want to discourage people sending patches.
Right?

Pantelis


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/