Hi
Here is a patch to drop the expensive memcpy of received ethernet frames in
interrupt context. I have not done any bench marking, but mounting a NFS rootfs
feels faster.
I am using a heavily modified enet.c in my system, but I think I got the patch
correct.
Also fixed a bug in set_multicast_list(), move the dmi list forward when
walking it(dmi = dmi->next;)
Comments? Anyone care to do some benchmarking?
Jocke
--- arch/ppc/8xx_io/enet.c.org Mon Oct 21 14:35:59 2002
+++ arch/ppc/8xx_io/enet.c Mon Oct 21 15:06:04 2002
@@ -96,18 +96,17 @@
* We don't need to allocate pages for the transmitter. We just use
* the skbuffer directly.
*/
+#define CPM_ENET_RX_FRSIZE 1600 /* 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
+
#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 RX_RING_SIZE 64
#define TX_RING_SIZE 64 /* Must be power of two */
#define TX_RING_MOD_MASK 63 /* 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 RX_RING_SIZE 8
#define TX_RING_SIZE 8 /* Must be power of two */
#define TX_RING_MOD_MASK 7 /* for this to work */
#endif
@@ -143,7 +142,7 @@
/* 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;
@@ -449,6 +448,7 @@
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;
@@ -497,25 +497,26 @@
pkt_len = bdp->cbd_datlen;
cep->stats.rx_bytes += pkt_len;
- /* This does 16 byte alignment, much more than we need.
+ /* 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) {
+ skb_tmp = dev_alloc_skb(CPM_ENET_RX_FRSIZE);
+ if (skb_tmp == NULL) {
printk("%s: Memory squeeze, dropping packet.\n",
dev->name);
cep->stats.rx_dropped++;
- }
- else {
+ } else {
+ skb = cep->rx_vaddr[bdp - cep->rx_bd_base];
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);
+
+ invalidate_dcache_range((unsigned long) skb_tmp->data,
+ (unsigned long) skb_tmp->data +
CPM_ENET_RX_FRSIZE);
+ bdp->cbd_bufaddr = __pa(skb_tmp->data);
+ cep->rx_vaddr[bdp - cep->rx_bd_base] = skb_tmp;
}
}
@@ -631,6 +632,7 @@
/* this delay is necessary here -- Cort */
udelay(10);
while (cpmp->cp_cpcr & CPM_CR_FLG);
+ dmi = dmi->next;
}
}
}
@@ -647,8 +649,7 @@
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 +840,15 @@
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);
+ bdp->cbd_sc = BD_ENET_RX_EMPTY | BD_ENET_RX_INTR;
+ bdp->cbd_bufaddr = __pa(skb->data);
+ invalidate_dcache_range((unsigned long) skb->data,
+ (unsigned long) skb->data +
CPM_ENET_RX_FRSIZE);
+ cep->rx_vaddr[k++] = skb;
+ bdp++;
}
/* Set the last buffer to wrap.
** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/
