[PATCH] S2io: Performance improvements
authorraghavendra.koushik@neterion.com <raghavendra.koushik@neterion.com>
Wed, 3 Aug 2005 19:32:00 +0000 (12:32 -0700)
committerJeff Garzik <jgarzik@pobox.com>
Thu, 11 Aug 2005 04:10:44 +0000 (00:10 -0400)
Hi,
This patch relates to mostly performance related changes.
1. Fixed incorrect computation of PANIC level in rx_buffer_level().
2. Removed unnecessary PIOs(read/write of tx_traffic_int and
   rx_traffic_int) from interrupt handler and removed read of
   general_int_status register from xmit routine.
3. Enable two-buffer mode(for Rx path) automatically for SGI
   systems. This improves Rx performance dramatically on
   SGI systems.

Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com>
Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
drivers/net/s2io.c
drivers/net/s2io.h

index 6668b99025c844b61e3b0580ce27ddf3b95838bf..28d6d3746c8073de9387754eb437929082cd1437 100644 (file)
@@ -100,8 +100,7 @@ static inline int rx_buffer_level(nic_t * sp, int rxb_size, int ring)
        mac_control = &sp->mac_control;
        if ((mac_control->rings[ring].pkt_cnt - rxb_size) > 16) {
                level = LOW;
-               if ((mac_control->rings[ring].pkt_cnt - rxb_size) <
-                               MAX_RXDS_PER_BLOCK) {
+               if (rxb_size <= MAX_RXDS_PER_BLOCK) {
                        level = PANIC;
                }
        }
@@ -2193,7 +2192,6 @@ static void rx_intr_handler(ring_info_t *ring_data)
 {
        nic_t *nic = ring_data->nic;
        struct net_device *dev = (struct net_device *) nic->dev;
-       XENA_dev_config_t __iomem *bar0 = nic->bar0;
        int get_block, get_offset, put_block, put_offset, ring_bufs;
        rx_curr_get_info_t get_info, put_info;
        RxD_t *rxdp;
@@ -2201,8 +2199,6 @@ static void rx_intr_handler(ring_info_t *ring_data)
 #ifndef CONFIG_S2IO_NAPI
        int pkt_cnt = 0;
 #endif
-       register u64 val64;
-
        spin_lock(&nic->rx_lock);
        if (atomic_read(&nic->card_state) == CARD_DOWN) {
                DBG_PRINT(ERR_DBG, "%s: %s going down for reset\n",
@@ -2210,13 +2206,6 @@ static void rx_intr_handler(ring_info_t *ring_data)
                spin_unlock(&nic->rx_lock);
        }
 
-       /*
-        * rx_traffic_int reg is an R1 register, hence we read and write
-        * back the same value in the register to clear it
-        */
-       val64 = readq(&bar0->tx_traffic_int);
-       writeq(val64, &bar0->tx_traffic_int);
-
        get_info = ring_data->rx_curr_get_info;
        get_block = get_info.block_index;
        put_info = ring_data->rx_curr_put_info;
@@ -2312,20 +2301,11 @@ static void rx_intr_handler(ring_info_t *ring_data)
 static void tx_intr_handler(fifo_info_t *fifo_data)
 {
        nic_t *nic = fifo_data->nic;
-       XENA_dev_config_t __iomem *bar0 = nic->bar0;
        struct net_device *dev = (struct net_device *) nic->dev;
        tx_curr_get_info_t get_info, put_info;
        struct sk_buff *skb;
        TxD_t *txdlp;
        u16 j, frg_cnt;
-       register u64 val64 = 0;
-
-       /*
-        * tx_traffic_int reg is an R1 register, hence we read and write
-        * back the same value in the register to clear it
-        */
-       val64 = readq(&bar0->tx_traffic_int);
-       writeq(val64, &bar0->tx_traffic_int);
 
        get_info = fifo_data->tx_curr_get_info;
        put_info = fifo_data->tx_curr_put_info;
@@ -2818,7 +2798,6 @@ int s2io_xmit(struct sk_buff *skb, struct net_device *dev)
 #endif
        mac_info_t *mac_control;
        struct config_param *config;
-       XENA_dev_config_t __iomem *bar0 = sp->bar0;
 
        mac_control = &sp->mac_control;
        config = &sp->config;
@@ -2870,7 +2849,6 @@ int s2io_xmit(struct sk_buff *skb, struct net_device *dev)
        }
 
        txdp->Control_2 |= config->tx_intr_type;
-
        txdp->Control_1 |= (TXD_BUFFER0_SIZE(frg_len) |
                            TXD_GATHER_CODE_FIRST);
        txdp->Control_1 |= TXD_LIST_OWN_XENA;
@@ -2890,6 +2868,8 @@ int s2io_xmit(struct sk_buff *skb, struct net_device *dev)
        val64 = mac_control->fifos[queue].list_info[put_off].list_phy_addr;
        writeq(val64, &tx_fifo->TxDL_Pointer);
 
+       wmb();
+
        val64 = (TX_FIFO_LAST_TXD_NUM(frg_cnt) | TX_FIFO_FIRST_LIST |
                 TX_FIFO_LAST_LIST);
 
@@ -2899,9 +2879,6 @@ int s2io_xmit(struct sk_buff *skb, struct net_device *dev)
 #endif
        writeq(val64, &tx_fifo->List_Control);
 
-       /* Perform a PCI read to flush previous writes */
-       val64 = readq(&bar0->general_int_status);
-
        put_off++;
        put_off %= mac_control->fifos[queue].tx_curr_put_info.fifo_len + 1;
        mac_control->fifos[queue].tx_curr_put_info.offset = put_off;
@@ -2940,7 +2917,7 @@ static irqreturn_t s2io_isr(int irq, void *dev_id, struct pt_regs *regs)
        nic_t *sp = dev->priv;
        XENA_dev_config_t __iomem *bar0 = sp->bar0;
        int i;
-       u64 reason = 0;
+       u64 reason = 0, val64;
        mac_info_t *mac_control;
        struct config_param *config;
 
@@ -2978,6 +2955,13 @@ static irqreturn_t s2io_isr(int irq, void *dev_id, struct pt_regs *regs)
 #else
        /* If Intr is because of Rx Traffic */
        if (reason & GEN_INTR_RXTRAFFIC) {
+               /*
+                * rx_traffic_int reg is an R1 register, writing all 1's
+                * will ensure that the actual interrupt causing bit get's
+                * cleared and hence a read can be avoided.
+                */
+               val64 = 0xFFFFFFFFFFFFFFFFULL;
+               writeq(val64, &bar0->rx_traffic_int);
                for (i = 0; i < config->rx_ring_num; i++) {
                        rx_intr_handler(&mac_control->rings[i]);
                }
@@ -2986,6 +2970,14 @@ static irqreturn_t s2io_isr(int irq, void *dev_id, struct pt_regs *regs)
 
        /* If Intr is because of Tx Traffic */
        if (reason & GEN_INTR_TXTRAFFIC) {
+               /*
+                * tx_traffic_int reg is an R1 register, writing all 1's
+                * will ensure that the actual interrupt causing bit get's
+                * cleared and hence a read can be avoided.
+                */
+               val64 = 0xFFFFFFFFFFFFFFFFULL;
+               writeq(val64, &bar0->tx_traffic_int);
+
                for (i = 0; i < config->tx_fifo_num; i++)
                        tx_intr_handler(&mac_control->fifos[i]);
        }
index 69dd0e51dda02400a1a79f185ec41329628ae389..ce9bf6d5ee0065ef9fa9bf6dbe7a2869a4d7bea3 100644 (file)
 #ifndef _S2IO_H
 #define _S2IO_H
 
+/* Enable 2 buffer mode by default for SGI system */
+#ifdef CONFIG_IA64_SGI_SN2
+#define CONFIG_2BUFF_MODE
+#endif
+
 #define TBD 0
 #define BIT(loc)               (0x8000000000000000ULL >> (loc))
 #define vBIT(val, loc, sz)     (((u64)val) << (64-loc-sz))