Eric Lee / smarc-ti-linux-kernel

Download as
Browse Code ยป

Commit e08e4c85d4735a3296266cccc4594c23c14cd333

Authored by Santosh Shilimkar
Committed by Kishon Vijay Abraham I
1 parent 2001f67e47
Exists in ti-linux-3.12.y and in 3 other branches newt-ti-linux-3.12.y, smarc-ti-linux-3.12.10, smarc-ti-linux-3.12.y

drivers: net: cpsw: convert tx completion to NAPI 

CPSW driver TX and RX interrupt handling is not optimal. The driver handles
both irq's together and relies on RX NAPI to proces the TX packet. Lets
seperate the interrupt handling and convert the TX completion to NAPI as
well.

With these changes the network perfromance shoots up by almost 40 %
for UDP. For TCP we also we get pretty good boost.

Signed-off-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Signed-off-by: Mugunthan V N <mugunthanvnm@ti.com>

Showing 1 changed file with 87 additions and 20 deletions Side-by-side Diff

drivers/net/ethernet/ti/cpsw.c
Diff comments   View file @ e08e4c8
... ... @@ -368,6 +368,9 @@
368 368 struct platform_device *pdev;
369 369 struct net_device *ndev;
370 370 struct napi_struct napi;
  371 + struct napi_struct napi_tx;
  372 + bool irq_enabled;
  373 + bool irq_tx_enabled;
371 374 struct device *dev;
372 375 struct cpsw_platform_data data;
373 376 struct cpsw_ss_regs __iomem *regs;
... ... @@ -390,7 +393,6 @@
390 393 /* snapshot of IRQ numbers */
391 394 u32 irqs_table[4];
392 395 u32 num_irqs;
393   - bool irq_enabled;
394 396 struct cpts *cpts;
395 397 u32 emac_port;
396 398 };
... ... @@ -484,6 +486,7 @@
484 486 #define CPSW_STATS_LEN ARRAY_SIZE(cpsw_gstrings_stats)
485 487  
486 488 #define napi_to_priv(napi) container_of(napi, struct cpsw_priv, napi)
  489 +#define napi_tx_to_priv(napi) container_of(napi, struct cpsw_priv, napi_tx)
487 490 #define for_each_slave(priv, func, arg...) \
488 491 do { \
489 492 struct cpsw_slave *slave; \
... ... @@ -713,9 +716,11 @@
713 716 {
714 717 struct cpsw_priv *priv = dev_id;
715 718  
716   - cpsw_intr_disable(priv);
717   - if (priv->irq_enabled == true) {
718   - cpsw_disable_irq(priv);
  719 + __raw_writel(0, &priv->wr_regs->rx_en);
  720 + if (priv->irq_enabled) {
  721 + disable_irq_nosync(priv->irqs_table[0]);
  722 + disable_irq_nosync(priv->irqs_table[1]);
  723 + disable_irq_nosync(priv->irqs_table[3]);
719 724 priv->irq_enabled = false;
720 725 }
721 726  
722 727  
723 728  
724 729  
725 730  
726 731  
727 732  
728 733  
... ... @@ -735,36 +740,84 @@
735 740 return IRQ_NONE;
736 741 }
737 742  
  743 +static irqreturn_t cpsw_tx_interrupt(int irq, void *dev_id)
  744 +{
  745 + struct cpsw_priv *priv = dev_id;
  746 +
  747 + __raw_writel(0, &priv->wr_regs->tx_en);
  748 + if (priv->irq_tx_enabled) {
  749 + disable_irq_nosync(priv->irqs_table[2]);
  750 + priv->irq_tx_enabled = false;
  751 + }
  752 +
  753 + if (netif_running(priv->ndev)) {
  754 + napi_schedule(&priv->napi_tx);
  755 + return IRQ_HANDLED;
  756 + }
  757 +
  758 + priv = cpsw_get_slave_priv(priv, 1);
  759 + if (!priv)
  760 + return IRQ_NONE;
  761 +
  762 + if (netif_running(priv->ndev)) {
  763 + napi_schedule(&priv->napi_tx);
  764 + return IRQ_HANDLED;
  765 + }
  766 + return IRQ_NONE;
  767 +}
  768 +
738 769 static int cpsw_poll(struct napi_struct *napi, int budget)
739 770 {
740 771 struct cpsw_priv *priv = napi_to_priv(napi);
741   - int num_tx, num_rx;
  772 + int num_rx;
742 773  
743   - num_tx = cpdma_chan_process(priv->txch, 128);
744   - if (num_tx)
745   - cpdma_ctlr_eoi(priv->dma, CPDMA_EOI_TX);
746   -
747 774 num_rx = cpdma_chan_process(priv->rxch, budget);
748 775 if (num_rx < budget) {
749 776 struct cpsw_priv *prim_cpsw;
750 777  
751 778 napi_complete(napi);
752   - cpsw_intr_enable(priv);
  779 + __raw_writel(0xFF, &priv->wr_regs->rx_en);
753 780 cpdma_ctlr_eoi(priv->dma, CPDMA_EOI_RX);
754 781 prim_cpsw = cpsw_get_slave_priv(priv, 0);
755   - if (prim_cpsw->irq_enabled == false) {
  782 + if (!prim_cpsw->irq_enabled) {
756 783 prim_cpsw->irq_enabled = true;
757   - cpsw_enable_irq(priv);
  784 + enable_irq(priv->irqs_table[0]);
  785 + enable_irq(priv->irqs_table[1]);
  786 + enable_irq(priv->irqs_table[3]);
758 787 }
759 788 }
760 789  
761   - if (num_rx || num_tx)
762   - cpsw_dbg(priv, intr, "poll %d rx, %d tx pkts\n",
763   - num_rx, num_tx);
  790 + if (num_rx)
  791 + cpsw_dbg(priv, intr, "poll %d rx pkts\n", num_rx);
764 792  
765 793 return num_rx;
766 794 }
767 795  
  796 +static int cpsw_tx_poll(struct napi_struct *napi, int budget)
  797 +{
  798 + struct cpsw_priv *priv = napi_tx_to_priv(napi);
  799 + int num_tx;
  800 +
  801 + num_tx = cpdma_chan_process(priv->txch, budget);
  802 + if (num_tx < budget) {
  803 + struct cpsw_priv *prim_cpsw;
  804 +
  805 + napi_complete(napi);
  806 + __raw_writel(0xFF, &priv->wr_regs->tx_en);
  807 + cpdma_ctlr_eoi(priv->dma, CPDMA_EOI_TX);
  808 + prim_cpsw = cpsw_get_slave_priv(priv, 0);
  809 + if (!prim_cpsw->irq_tx_enabled) {
  810 + prim_cpsw->irq_tx_enabled = true;
  811 + enable_irq(priv->irqs_table[2]);
  812 + }
  813 + }
  814 +
  815 + if (num_tx)
  816 + cpsw_dbg(priv, intr, "poll %d tx pkts\n", num_tx);
  817 +
  818 + return num_tx;
  819 +}
  820 +
768 821 static inline void soft_reset(const char *module, void __iomem *reg)
769 822 {
770 823 unsigned long timeout = jiffies + HZ;
771 824  
772 825  
... ... @@ -1245,15 +1298,17 @@
1245 1298 }
1246 1299  
1247 1300 napi_enable(&priv->napi);
  1301 + napi_enable(&priv->napi_tx);
1248 1302 cpdma_ctlr_start(priv->dma);
1249 1303 cpsw_intr_enable(priv);
1250 1304 cpdma_ctlr_eoi(priv->dma, CPDMA_EOI_RX);
1251 1305 cpdma_ctlr_eoi(priv->dma, CPDMA_EOI_TX);
1252 1306  
1253 1307 prim_cpsw = cpsw_get_slave_priv(priv, 0);
1254   - if (prim_cpsw->irq_enabled == false) {
  1308 + if (!prim_cpsw->irq_enabled && !prim_cpsw->irq_tx_enabled) {
1255 1309 if ((priv == prim_cpsw) || !netif_running(prim_cpsw->ndev)) {
1256 1310 prim_cpsw->irq_enabled = true;
  1311 + prim_cpsw->irq_tx_enabled = true;
1257 1312 cpsw_enable_irq(prim_cpsw);
1258 1313 }
1259 1314 }
... ... @@ -1277,6 +1332,7 @@
1277 1332 cpsw_info(priv, ifdown, "shutting down cpsw device\n");
1278 1333 netif_stop_queue(priv->ndev);
1279 1334 napi_disable(&priv->napi);
  1335 + napi_disable(&priv->napi_tx);
1280 1336 netif_carrier_off(priv->ndev);
1281 1337  
1282 1338 if (cpsw_common_res_usage_state(priv) <= 1) {
... ... @@ -1948,6 +2004,8 @@
1948 2004 ndev->netdev_ops = &cpsw_netdev_ops;
1949 2005 SET_ETHTOOL_OPS(ndev, &cpsw_ethtool_ops);
1950 2006 netif_napi_add(ndev, &priv_sl2->napi, cpsw_poll, CPSW_POLL_WEIGHT);
  2007 + netif_napi_add(ndev, &priv_sl2->napi_tx, cpsw_tx_poll,
  2008 + CPSW_POLL_WEIGHT);
1951 2009  
1952 2010 /* register the network device */
1953 2011 SET_NETDEV_DEV(ndev, &pdev->dev);
... ... @@ -1971,7 +2029,7 @@
1971 2029 void __iomem *ss_regs;
1972 2030 struct resource *res, *ss_res;
1973 2031 u32 slave_offset, sliver_offset, slave_size;
1974   - int ret = 0, i, k = 0;
  2032 + int ret = 0, i, j = 0, k = 0;
1975 2033  
1976 2034 ndev = alloc_etherdev(sizeof(struct cpsw_priv));
1977 2035 if (!ndev) {
... ... @@ -1989,6 +2047,7 @@
1989 2047 priv->rx_packet_max = max(rx_packet_max, 128);
1990 2048 priv->cpts = devm_kzalloc(&pdev->dev, sizeof(struct cpts), GFP_KERNEL);
1991 2049 priv->irq_enabled = true;
  2050 + priv->irq_tx_enabled = true;
1992 2051 if (!priv->cpts) {
1993 2052 pr_err("error allocating cpts\n");
1994 2053 goto clean_ndev_ret;
... ... @@ -2164,6 +2223,7 @@
2164 2223 ndev->netdev_ops = &cpsw_netdev_ops;
2165 2224 SET_ETHTOOL_OPS(ndev, &cpsw_ethtool_ops);
2166 2225 netif_napi_add(ndev, &priv->napi, cpsw_poll, CPSW_POLL_WEIGHT);
  2226 + netif_napi_add(ndev, &priv->napi_tx, cpsw_tx_poll, CPSW_POLL_WEIGHT);
2167 2227  
2168 2228 /* register the network device */
2169 2229 SET_NETDEV_DEV(ndev, &pdev->dev);
... ... @@ -2175,9 +2235,16 @@
2175 2235 }
2176 2236  
2177 2237 while ((res = platform_get_resource(priv->pdev, IORESOURCE_IRQ, k))) {
2178   - for (i = res->start; i <= res->end; i++) {
2179   - if (devm_request_irq(&pdev->dev, i, cpsw_interrupt, 0,
2180   - dev_name(priv->dev), priv)) {
  2238 + for (i = res->start; i <= res->end; i++, j++) {
  2239 + if (j == 2)
  2240 + ret = devm_request_irq(&pdev->dev, i,
  2241 + cpsw_tx_interrupt, 0,
  2242 + "eth-tx", priv);
  2243 + else
  2244 + ret = devm_request_irq(&pdev->dev, i,
  2245 + cpsw_interrupt, 0,
  2246 + dev_name(priv->dev), priv);
  2247 + if (ret) {
2181 2248 dev_err(priv->dev, "error attaching irq\n");
2182 2249 goto clean_ale_ret;
2183 2250 }