[PATCH] s2io: performance improvements
authorAnanda Raju <Ananda.Raju@neterion.com>
Fri, 21 Apr 2006 23:03:13 +0000 (19:03 -0400)
committerJeff Garzik <jeff@garzik.org>
Tue, 2 May 2006 19:16:35 +0000 (15:16 -0400)
hi,
This patch contains all the changes that were done to improve
performance of s2io driver. one line  description of the changes
are

1. For Non NAPI case the rx interrupt handler is being called
   unconditionally
2. code optimization and adding prefetch skb->data
3. Remove modulo operations in fast path
4. Enable Group Reads and set backoff interval to 0x1000
5. correct PIC_CNTL_SHARED_SPLITS  macro definition, and reduce
   pause parameter
6. Corrected logic of identifying rx buffer level in rx_buffer_level()
7. fix DMA map and unmap done with different sizes in 1-buf mode
8. Removed forcible disabling of ERO
9. Send up the packets with transfer code = 0x5

Signed-off-by: Ananda Raju <ananda.raju@neterion.com>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
drivers/net/s2io-regs.h
drivers/net/s2io.c

index 00179bc3437fccea29fa780d9df5da4233d1f886..542ae1574801094c69f2f9a6c7a63dc6464c6298 100644 (file)
@@ -187,7 +187,7 @@ typedef struct _XENA_dev_config {
 /* PIC Control registers */
        u64 pic_control;
 #define PIC_CNTL_RX_ALARM_MAP_1                BIT(0)
-#define PIC_CNTL_SHARED_SPLITS(n)              vBIT(n,11,4)
+#define PIC_CNTL_SHARED_SPLITS(n)              vBIT(n,11,5)
 
        u64 swapper_ctrl;
 #define SWAPPER_CTRL_PIF_R_FE                  BIT(0)
@@ -284,9 +284,13 @@ typedef struct _XENA_dev_config {
        u64 gpio_control;
 #define GPIO_CTRL_GPIO_0               BIT(8)
        u64 misc_control;
+#define EXT_REQ_EN                     BIT(1)
 #define MISC_LINK_STABILITY_PRD(val)   vBIT(val,29,3)
 
-       u8 unused7_1[0x240 - 0x208];
+       u8 unused7_1[0x230 - 0x208];
+
+       u64 pic_control2;
+       u64 ini_dperr_ctrl;
 
        u64 wreq_split_mask;
 #define        WREQ_SPLIT_MASK_SET_MASK(val)   vBIT(val, 52, 12)
@@ -493,6 +497,7 @@ typedef struct _XENA_dev_config {
 #define PRC_CTRL_NO_SNOOP_DESC                 BIT(22)
 #define PRC_CTRL_NO_SNOOP_BUFF                 BIT(23)
 #define PRC_CTRL_BIMODAL_INTERRUPT             BIT(37)
+#define PRC_CTRL_GROUP_READS                   BIT(38)
 #define PRC_CTRL_RXD_BACKOFF_INTERVAL(val)     vBIT(val,40,24)
 
        u64 prc_alarm_action;
index 79208f434ac17773123c9cb92364131687c3fc64..846873159662b5509dd65c16827135d3107f1582 100644 (file)
@@ -106,18 +106,14 @@ static inline int RXD_IS_UP2DT(RxD_t *rxdp)
 #define LOW    2
 static inline int rx_buffer_level(nic_t * sp, int rxb_size, int ring)
 {
-       int level = 0;
        mac_info_t *mac_control;
 
        mac_control = &sp->mac_control;
-       if ((mac_control->rings[ring].pkt_cnt - rxb_size) > 16) {
-               level = LOW;
-               if (rxb_size <= rxd_count[sp->rxd_mode]) {
-                       level = PANIC;
-               }
-       }
-
-       return level;
+       if (rxb_size <= rxd_count[sp->rxd_mode])
+               return PANIC;
+       else if ((mac_control->rings[ring].pkt_cnt - rxb_size) > 16)
+               return  LOW;
+       return 0;
 }
 
 /* Ethtool related variables and Macros. */
@@ -311,7 +307,7 @@ static unsigned int rts_frm_len[MAX_RX_RINGS] =
     {[0 ...(MAX_RX_RINGS - 1)] = 0 };
 static unsigned int rx_ring_mode = 1;
 static unsigned int use_continuous_tx_intrs = 1;
-static unsigned int rmac_pause_time = 65535;
+static unsigned int rmac_pause_time = 0x100;
 static unsigned int mc_pause_threshold_q0q3 = 187;
 static unsigned int mc_pause_threshold_q4q7 = 187;
 static unsigned int shared_splits;
@@ -1545,13 +1541,22 @@ static int init_nic(struct s2io_nic *nic)
        val64 |= PIC_CNTL_SHARED_SPLITS(shared_splits);
        writeq(val64, &bar0->pic_control);
 
+       if (nic->config.bus_speed == 266) {
+               writeq(TXREQTO_VAL(0x7f) | TXREQTO_EN, &bar0->txreqtimeout);
+               writeq(0x0, &bar0->read_retry_delay);
+               writeq(0x0, &bar0->write_retry_delay);
+       }
+
        /*
         * Programming the Herc to split every write transaction
         * that does not start on an ADB to reduce disconnects.
         */
        if (nic->device_type == XFRAME_II_DEVICE) {
-               val64 = WREQ_SPLIT_MASK_SET_MASK(255);
-               writeq(val64, &bar0->wreq_split_mask);
+               val64 = EXT_REQ_EN | MISC_LINK_STABILITY_PRD(3);
+               writeq(val64, &bar0->misc_control);
+               val64 = readq(&bar0->pic_control2);
+               val64 &= ~(BIT(13)|BIT(14)|BIT(15));
+               writeq(val64, &bar0->pic_control2);
        }
 
        /* Setting Link stability period to 64 ms */ 
@@ -1948,6 +1953,10 @@ static int start_nic(struct s2io_nic *nic)
                        val64 |= PRC_CTRL_RC_ENABLED;
                else
                        val64 |= PRC_CTRL_RC_ENABLED | PRC_CTRL_RING_MODE_3;
+               if (nic->device_type == XFRAME_II_DEVICE)
+                       val64 |= PRC_CTRL_GROUP_READS;
+               val64 &= ~PRC_CTRL_RXD_BACKOFF_INTERVAL(0xFFFFFF);
+               val64 |= PRC_CTRL_RXD_BACKOFF_INTERVAL(0x1000);
                writeq(val64, &bar0->prc_ctrl_n[i]);
        }
 
@@ -2231,13 +2240,12 @@ static int fill_rx_buffers(struct s2io_nic *nic, int ring_no)
        alloc_cnt = mac_control->rings[ring_no].pkt_cnt -
            atomic_read(&nic->rx_bufs_left[ring_no]);
 
+       block_no1 = mac_control->rings[ring_no].rx_curr_get_info.block_index;
+       off1 = mac_control->rings[ring_no].rx_curr_get_info.offset;
        while (alloc_tab < alloc_cnt) {
                block_no = mac_control->rings[ring_no].rx_curr_put_info.
                    block_index;
-               block_no1 = mac_control->rings[ring_no].rx_curr_get_info.
-                   block_index;
                off = mac_control->rings[ring_no].rx_curr_put_info.offset;
-               off1 = mac_control->rings[ring_no].rx_curr_get_info.offset;
 
                rxdp = mac_control->rings[ring_no].
                                rx_blocks[block_no].rxds[off].virt_addr;
@@ -2307,9 +2315,9 @@ static int fill_rx_buffers(struct s2io_nic *nic, int ring_no)
                        memset(rxdp, 0, sizeof(RxD1_t));
                        skb_reserve(skb, NET_IP_ALIGN);
                        ((RxD1_t*)rxdp)->Buffer0_ptr = pci_map_single
-                           (nic->pdev, skb->data, size, PCI_DMA_FROMDEVICE);
-                       rxdp->Control_2 &= (~MASK_BUFFER0_SIZE_1);
-                       rxdp->Control_2 |= SET_BUFFER0_SIZE_1(size);
+                           (nic->pdev, skb->data, size - NET_IP_ALIGN,
+                               PCI_DMA_FROMDEVICE);
+                       rxdp->Control_2 = SET_BUFFER0_SIZE_1(size - NET_IP_ALIGN);
 
                } else if (nic->rxd_mode >= RXD_MODE_3A) {
                        /*
@@ -2516,7 +2524,7 @@ static int s2io_poll(struct net_device *dev, int *budget)
        mac_info_t *mac_control;
        struct config_param *config;
        XENA_dev_config_t __iomem *bar0 = nic->bar0;
-       u64 val64;
+       u64 val64 = 0xFFFFFFFFFFFFFFFFULL;
        int i;
 
        atomic_inc(&nic->isr_cnt);
@@ -2528,8 +2536,8 @@ static int s2io_poll(struct net_device *dev, int *budget)
                nic->pkts_to_process = dev->quota;
        org_pkts_to_process = nic->pkts_to_process;
 
-       val64 = readq(&bar0->rx_traffic_int);
        writeq(val64, &bar0->rx_traffic_int);
+       val64 = readl(&bar0->rx_traffic_int);
 
        for (i = 0; i < config->rx_ring_num; i++) {
                rx_intr_handler(&mac_control->rings[i]);
@@ -2666,6 +2674,7 @@ static void rx_intr_handler(ring_info_t *ring_data)
                                         ((RxD3_t*)rxdp)->Buffer2_ptr,
                                         dev->mtu, PCI_DMA_FROMDEVICE);
                }
+               prefetch(skb->data);
                rx_osm_handler(ring_data, rxdp);
                get_info.offset++;
                ring_data->rx_curr_get_info.offset = get_info.offset;
@@ -2760,7 +2769,8 @@ to loss of link\n");
                dev_kfree_skb_irq(skb);
 
                get_info.offset++;
-               get_info.offset %= get_info.fifo_len + 1;
+               if (get_info.offset == get_info.fifo_len + 1)
+                       get_info.offset = 0;
                txdlp = (TxD_t *) fifo_data->list_info
                    [get_info.offset].list_virt_addr;
                fifo_data->tx_curr_get_info.offset =
@@ -3545,7 +3555,8 @@ static int s2io_xmit(struct sk_buff *skb, struct net_device *dev)
 
        queue_len = mac_control->fifos[queue].tx_curr_put_info.fifo_len + 1;
        /* Avoid "put" pointer going beyond "get" pointer */
-       if (txdp->Host_Control || (((put_off + 1) % queue_len) == get_off)) {
+       if (txdp->Host_Control ||
+                  ((put_off+1) == queue_len ? 0 : (put_off+1)) == get_off) {
                DBG_PRINT(TX_DBG, "Error in xmit, No free TXDs.\n");
                netif_stop_queue(dev);
                dev_kfree_skb(skb);
@@ -3655,11 +3666,12 @@ static int s2io_xmit(struct sk_buff *skb, struct net_device *dev)
        mmiowb();
 
        put_off++;
-       put_off %= mac_control->fifos[queue].tx_curr_put_info.fifo_len + 1;
+       if (put_off == mac_control->fifos[queue].tx_curr_put_info.fifo_len + 1)
+               put_off = 0;
        mac_control->fifos[queue].tx_curr_put_info.offset = put_off;
 
        /* Avoid "put" pointer going beyond "get" pointer */
-       if (((put_off + 1) % queue_len) == get_off) {
+       if (((put_off+1) == queue_len ? 0 : (put_off+1)) == get_off) {
                DBG_PRINT(TX_DBG,
                          "No free TxDs for xmit, Put: 0x%x Get:0x%x\n",
                          put_off, get_off);
@@ -3887,43 +3899,37 @@ static irqreturn_t s2io_isr(int irq, void *dev_id, struct pt_regs *regs)
                return IRQ_NONE;
        }
 
+       val64 = 0xFFFFFFFFFFFFFFFFULL;
 #ifdef CONFIG_S2IO_NAPI
        if (reason & GEN_INTR_RXTRAFFIC) {
                if (netif_rx_schedule_prep(dev)) {
-                       en_dis_able_nic_intrs(sp, RX_TRAFFIC_INTR,
-                                             DISABLE_INTRS);
+                       writeq(val64, &bar0->rx_traffic_mask);
                        __netif_rx_schedule(dev);
                }
        }
 #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]);
-               }
+       /*
+        * Rx handler is called by default, without checking for the
+        * cause of interrupt.
+        * 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.
+        */
+       writeq(val64, &bar0->rx_traffic_int);
+       for (i = 0; i < config->rx_ring_num; i++) {
+               rx_intr_handler(&mac_control->rings[i]);
        }
 #endif
 
-       /* 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);
+       /*
+        * 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.
+        */
+       writeq(val64, &bar0->tx_traffic_int);
 
-               for (i = 0; i < config->tx_fifo_num; i++)
-                       tx_intr_handler(&mac_control->fifos[i]);
-       }
+       for (i = 0; i < config->tx_fifo_num; i++)
+               tx_intr_handler(&mac_control->fifos[i]);
 
        if (reason & GEN_INTR_TXPIC)
                s2io_txpic_intr_handle(sp);
@@ -5695,18 +5701,27 @@ static int rx_osm_handler(ring_info_t *ring_data, RxD_t * rxdp)
                ((unsigned long) rxdp->Host_Control);
        int ring_no = ring_data->ring_no;
        u16 l3_csum, l4_csum;
+       unsigned long long err = rxdp->Control_1 & RXD_T_CODE;
        lro_t *lro;
 
        skb->dev = dev;
-       if (rxdp->Control_1 & RXD_T_CODE) {
-               unsigned long long err = rxdp->Control_1 & RXD_T_CODE;
-               DBG_PRINT(ERR_DBG, "%s: Rx error Value: 0x%llx\n",
-                         dev->name, err);
-               dev_kfree_skb(skb);
-               sp->stats.rx_crc_errors++;
-               atomic_dec(&sp->rx_bufs_left[ring_no]);
-               rxdp->Host_Control = 0;
-               return 0;
+       if (err) {
+               /*
+               * Drop the packet if bad transfer code. Exception being
+               * 0x5, which could be due to unsupported IPv6 extension header.
+               * In this case, we let stack handle the packet.
+               * Note that in this case, since checksum will be incorrect,
+               * stack will validate the same.
+               */
+               if (err && ((err >> 48) != 0x5)) {
+                       DBG_PRINT(ERR_DBG, "%s: Rx error Value: 0x%llx\n",
+                               dev->name, err);
+                       sp->stats.rx_crc_errors++;
+                       dev_kfree_skb(skb);
+                       atomic_dec(&sp->rx_bufs_left[ring_no]);
+                       rxdp->Host_Control = 0;
+                       return 0;
+               }
        }
 
        /* Updating statistics */
@@ -5918,13 +5933,6 @@ static void s2io_init_pci(nic_t * sp)
        pci_write_config_word(sp->pdev, PCI_COMMAND,
                              (pci_cmd | PCI_COMMAND_PARITY));
        pci_read_config_word(sp->pdev, PCI_COMMAND, &pci_cmd);
-
-       /* Forcibly disabling relaxed ordering capability of the card. */
-       pcix_cmd &= 0xfffd;
-       pci_write_config_word(sp->pdev, PCIX_COMMAND_REGISTER,
-                             pcix_cmd);
-       pci_read_config_word(sp->pdev, PCIX_COMMAND_REGISTER,
-                            &(pcix_cmd));
 }
 
 MODULE_AUTHOR("Raghavendra Koushik <raghavendra.koushik@neterion.com>");