From: Felix Fietkau Date: Sat, 22 Sep 2012 00:34:49 +0000 (+0000) Subject: cns3xxx: implement ethernet fraglist tx support for improved bridging performance... X-Git-Url: http://git.lede-project.org./?a=commitdiff_plain;h=fb71465a4ea5f88af59be208b38d7dc004f82c25;p=openwrt%2Fstaging%2Fnbd.git cns3xxx: implement ethernet fraglist tx support for improved bridging performance with jumbo frames SVN-Revision: 33503 --- diff --git a/target/linux/cns3xxx/patches-3.3/410-ethernet_fix_jumbo_frame.patch b/target/linux/cns3xxx/patches-3.3/410-ethernet_fix_jumbo_frame.patch index 733af98da5..0500fce8fa 100644 --- a/target/linux/cns3xxx/patches-3.3/410-ethernet_fix_jumbo_frame.patch +++ b/target/linux/cns3xxx/patches-3.3/410-ethernet_fix_jumbo_frame.patch @@ -120,12 +120,12 @@ + /* process received frame */ + dma_unmap_single(NULL, rx_ring->phys_tab[i], + RX_SEGMENT_MRU, DMA_FROM_DEVICE); - -- dev = switch_port_tab[desc->sp]->netdev; ++ + skb = build_skb(rx_ring->buff_tab[i]); + if (!skb) + break; -+ + +- dev = switch_port_tab[desc->sp]->netdev; + skb->dev = switch_port_tab[desc->sp]->netdev; length = desc->sdl; @@ -158,12 +158,12 @@ + sw->frag_first->truesize += skb->truesize; + } + sw->frag_last = skb; -+ -+ if (desc->lsd) { -+ struct net_device *dev; - dev->stats.rx_packets++; - dev->stats.rx_bytes += length; ++ if (desc->lsd) { ++ struct net_device *dev; ++ + skb = sw->frag_first; + dev = skb->dev; + skb->protocol = eth_type_trans(skb, dev); @@ -217,7 +217,51 @@ if (++i == RX_DESCS) { i = 0; desc = &(rx_ring)->desc[i]; -@@ -671,12 +709,6 @@ static int eth_xmit(struct sk_buff *skb, +@@ -653,42 +691,60 @@ static int eth_poll(struct napi_struct * + return received; + } + ++static void eth_set_desc(struct _tx_ring *tx_ring, int index, int index_last, ++ void *data, int len, u32 config0, u32 pmap) ++{ ++ struct tx_desc *tx_desc = &(tx_ring)->desc[index]; ++ unsigned int phys; ++ ++ phys = dma_map_single(NULL, data, len, DMA_TO_DEVICE); ++ tx_desc->sdp = phys; ++ tx_desc->pmap = pmap; ++ tx_ring->phys_tab[index] = phys; ++ ++ config0 |= len; ++ if (index == TX_DESCS - 1) ++ config0 |= END_OF_RING; ++ if (index == index_last) ++ config0 |= LAST_SEGMENT; ++ ++ mb(); ++ tx_desc->config0 = config0; ++} ++ + static int eth_xmit(struct sk_buff *skb, struct net_device *dev) + { + struct port *port = netdev_priv(dev); + struct sw *sw = port->sw; + struct _tx_ring *tx_ring = sw->tx_ring; +- struct tx_desc *tx_desc; +- int index; +- int len; ++ struct sk_buff *skb1; + char pmap = (1 << port->id); +- unsigned int phys; + int nr_frags = skb_shinfo(skb)->nr_frags; +- struct skb_frag_struct *frag; ++ int nr_desc = nr_frags; ++ int index0, index, index_last; ++ int len0; + unsigned int i; +- u32 config0 = 0; ++ u32 config0; + if (pmap == 8) pmap = (1 << 4); @@ -226,21 +270,132 @@ - dev->stats.tx_errors++; - return NETDEV_TX_OK; - } -- ++ skb_walk_frags(skb, skb1) ++ nr_desc++; + spin_lock(&tx_lock); - if ((tx_ring->num_used + nr_frags) >= TX_DESCS) { -@@ -701,8 +733,7 @@ static int eth_xmit(struct sk_buff *skb, +- if ((tx_ring->num_used + nr_frags) >= TX_DESCS) { ++ if ((tx_ring->num_used + nr_desc + 1) >= TX_DESCS) { + clear_tx_desc(sw); +- if ((tx_ring->num_used + nr_frags) >= TX_DESCS) { ++ if ((tx_ring->num_used + nr_desc + 1) >= TX_DESCS) { + spin_unlock(&tx_lock); + return NETDEV_TX_BUSY; + } + } + +- index = tx_ring->cur_index; +- tx_ring->cur_index = ((tx_ring->cur_index + nr_frags + 1) % TX_DESCS); ++ index = index0 = tx_ring->cur_index; ++ index_last = (index0 + nr_desc) % TX_DESCS; ++ tx_ring->cur_index = (index_last + 1) % TX_DESCS; - len = skb->len; + spin_unlock(&tx_lock); + +@@ -696,79 +752,41 @@ static int eth_xmit(struct sk_buff *skb, + if (skb->ip_summed == CHECKSUM_PARTIAL) + config0 |= UDP_CHECKSUM | TCP_CHECKSUM; + +- if (!nr_frags) { +- tx_desc = &(tx_ring)->desc[index]; +- +- len = skb->len; ++ len0 = skb->len; - phys = dma_map_single(NULL, skb->data, len, - DMA_TO_DEVICE); -+ phys = dma_map_single(NULL, skb->data, len, DMA_TO_DEVICE); ++ /* fragments */ ++ for (i = 0; i < nr_frags; i++) { ++ struct skb_frag_struct *frag; ++ void *addr; + +- tx_desc->sdp = phys; +- tx_desc->pmap = pmap; +- tx_ring->phys_tab[index] = phys; ++ index = (index + 1) % TX_DESCS; + +- tx_ring->buff_tab[index] = skb; +- config0 |= FIRST_SEGMENT | LAST_SEGMENT; +- } else { +- index = ((index + nr_frags) % TX_DESCS); +- tx_desc = &(tx_ring)->desc[index]; ++ frag = &skb_shinfo(skb)->frags[i]; ++ addr = page_address(skb_frag_page(frag)) + frag->page_offset; + +- /* fragments */ +- for (i = nr_frags; i > 0; i--) { +- u32 config; +- void *addr; +- +- frag = &skb_shinfo(skb)->frags[i-1]; +- len = frag->size; +- +- addr = page_address(skb_frag_page(frag)) + +- frag->page_offset; +- phys = dma_map_single(NULL, addr, len, DMA_TO_DEVICE); +- +- tx_desc->sdp = phys; +- +- tx_desc->pmap = pmap; +- tx_ring->phys_tab[index] = phys; +- +- config = config0 | len; +- if (i == nr_frags) { +- config |= LAST_SEGMENT; +- tx_ring->buff_tab[index] = skb; +- } +- if (index == TX_DESCS - 1) +- config |= END_OF_RING; +- tx_desc->config0 = config; +- +- if (index == 0) { +- index = TX_DESCS - 1; +- tx_desc = &(tx_ring)->desc[index]; +- } else { +- index--; +- tx_desc--; +- } +- } ++ eth_set_desc(tx_ring, index, index_last, addr, frag->size, ++ config0, pmap); ++ } + +- /* header */ +- len = skb->len - skb->data_len; ++ if (nr_frags) ++ len0 = skb->len - skb->data_len; + +- phys = dma_map_single(NULL, skb->data, len, DMA_TO_DEVICE); ++ skb_walk_frags(skb, skb1) { ++ index = (index + 1) % TX_DESCS; ++ len0 -= skb1->len; - tx_desc->sdp = phys; - tx_desc->pmap = pmap; -@@ -849,24 +880,24 @@ static int init_rings(struct sw *sw) +- tx_desc->sdp = phys; +- tx_desc->pmap = pmap; +- tx_ring->phys_tab[index] = phys; +- config0 |= FIRST_SEGMENT; ++ eth_set_desc(tx_ring, index, index_last, skb1->data, skb1->len, ++ config0, pmap); + } + +- if (index == TX_DESCS - 1) +- config0 |= END_OF_RING; +- +- tx_desc->config0 = config0 | len; ++ tx_ring->buff_tab[index0] = skb; ++ eth_set_desc(tx_ring, index0, index_last, skb->data, len0, ++ config0 | FIRST_SEGMENT, pmap); + + mb(); + + spin_lock(&tx_lock); +- tx_ring->num_used += nr_frags + 1; ++ tx_ring->num_used += nr_desc + 1; + spin_unlock(&tx_lock); + + dev->stats.tx_packets++; +@@ -849,24 +867,24 @@ static int init_rings(struct sw *sw) /* Setup RX buffers */ for (i = 0; i < RX_DESCS; i++) { struct rx_desc *desc = &(rx_ring)->desc[i]; @@ -276,7 +431,7 @@ rx_ring->phys_tab[i] = desc->sdp; desc->cown = 0; } -@@ -905,12 +936,13 @@ static void destroy_rings(struct sw *sw) +@@ -905,12 +923,13 @@ static void destroy_rings(struct sw *sw) struct _rx_ring *rx_ring = sw->rx_ring; struct rx_desc *desc = &(rx_ring)->desc[i]; struct sk_buff *skb = sw->rx_ring->buff_tab[i]; @@ -296,7 +451,7 @@ } dma_pool_free(rx_dma_pool, sw->rx_ring->desc, sw->rx_ring->phys_addr); dma_pool_destroy(rx_dma_pool); -@@ -1085,13 +1117,22 @@ static int eth_set_mac(struct net_device +@@ -1085,13 +1104,22 @@ static int eth_set_mac(struct net_device return 0; } @@ -320,7 +475,16 @@ .ndo_set_mac_address = eth_set_mac, .ndo_validate_addr = eth_validate_addr, }; -@@ -1124,6 +1165,10 @@ static int __devinit eth_init_one(struct +@@ -1111,7 +1139,7 @@ static int __devinit eth_init_one(struct + if (!(napi_dev = alloc_etherdev(sizeof(struct sw)))) + return -ENOMEM; + strcpy(napi_dev->name, "switch%d"); +- napi_dev->features = NETIF_F_IP_CSUM | NETIF_F_SG; ++ napi_dev->features = NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_FRAGLIST; + + SET_NETDEV_DEV(napi_dev, &pdev->dev); + sw = netdev_priv(napi_dev); +@@ -1124,6 +1152,10 @@ static int __devinit eth_init_one(struct goto err_free; } @@ -331,3 +495,12 @@ for (i = 0; i < 4; i++) { temp = __raw_readl(&sw->regs->mac_cfg[i]); temp |= (PORT_DISABLE); +@@ -1185,7 +1217,7 @@ static int __devinit eth_init_one(struct + dev->netdev_ops = &cns3xxx_netdev_ops; + dev->ethtool_ops = &cns3xxx_ethtool_ops; + dev->tx_queue_len = 1000; +- dev->features = NETIF_F_IP_CSUM | NETIF_F_SG; ++ dev->features = NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_FRAGLIST; + + switch_port_tab[port->id] = port; + memcpy(dev->dev_addr, &plat->hwaddr[i], ETH_ALEN);