mv643xx_eth: implement Large Receive Offload

Controlled by a compile-time (Kconfig) option for now, since it
isn't a win in all cases.

Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/drivers/net/Kconfig b/drivers/net/Kconfig
index 62bc022..3fed334 100644 (file)
--- a/drivers/net/Kconfig
+++ b/drivers/net/Kconfig
@@ -2318,6 +2318,16 @@ config MV643XX_ETH
          Some boards that use the Discovery chipset are the Momenco
          Ocelot C and Jaguar ATX and Pegasos II.
 
+config MV643XX_ETH_LRO
+       tristate "Marvell 643XX ethernet driver LRO support"
+       depends on MV643XX_ETH
+       select INET_LRO
+       help
+         Say y here if you want to use Large Receive Offload for the
+         mv643xx_eth driver.
+
+         If in doubt, say N.
+
 config QLA3XXX
        tristate "QLogic QLA3XXX Network Driver Support"
        depends on PCI

diff --git a/drivers/net/mv643xx_eth.c b/drivers/net/mv643xx_eth.c
index a99e5d3..bb96931 100644 (file)
--- a/drivers/net/mv643xx_eth.c
+++ b/drivers/net/mv643xx_eth.c
@@ -53,6 +53,7 @@
 #include <linux/mv643xx_eth.h>
 #include <linux/io.h>
 #include <linux/types.h>
+#include <linux/inet_lro.h>
 #include <asm/system.h>
 
 static char mv643xx_eth_driver_name[] = "mv643xx_eth";
@@ -227,6 +228,12 @@ struct tx_desc {
 #define RX_ENABLE_INTERRUPT            0x20000000
 #define RX_FIRST_DESC                  0x08000000
 #define RX_LAST_DESC                   0x04000000
+#define RX_IP_HDR_OK                   0x02000000
+#define RX_PKT_IS_IPV4                 0x01000000
+#define RX_PKT_IS_ETHERNETV2           0x00800000
+#define RX_PKT_LAYER4_TYPE_MASK                0x00600000
+#define RX_PKT_LAYER4_TYPE_TCP_IPV4    0x00000000
+#define RX_PKT_IS_VLAN_TAGGED          0x00080000
 
 /* TX descriptor command */
 #define TX_ENABLE_INTERRUPT            0x00800000
@@ -324,6 +331,12 @@ struct mib_counters {
        u32 late_collision;
 };
 
+struct lro_counters {
+       u32 lro_aggregated;
+       u32 lro_flushed;
+       u32 lro_no_desc;
+};
+
 struct rx_queue {
        int index;
 
@@ -337,6 +350,11 @@ struct rx_queue {
        dma_addr_t rx_desc_dma;
        int rx_desc_area_size;
        struct sk_buff **rx_skb;
+
+#ifdef CONFIG_MV643XX_ETH_LRO
+       struct net_lro_mgr lro_mgr;
+       struct net_lro_desc lro_arr[8];
+#endif
 };
 
 struct tx_queue {
@@ -372,6 +390,8 @@ struct mv643xx_eth_private {
        spinlock_t mib_counters_lock;
        struct mib_counters mib_counters;
 
+       struct lro_counters lro_counters;
+
        struct work_struct tx_timeout_task;
 
        struct napi_struct napi;
@@ -496,12 +516,42 @@ static void txq_maybe_wake(struct tx_queue *txq)
 
 
 /* rx napi ******************************************************************/
+#ifdef CONFIG_MV643XX_ETH_LRO
+static int
+mv643xx_get_skb_header(struct sk_buff *skb, void **iphdr, void **tcph,
+                      u64 *hdr_flags, void *priv)
+{
+       unsigned long cmd_sts = (unsigned long)priv;
+
+       /*
+        * Make sure that this packet is Ethernet II, is not VLAN
+        * tagged, is IPv4, has a valid IP header, and is TCP.
+        */
+       if ((cmd_sts & (RX_IP_HDR_OK | RX_PKT_IS_IPV4 |
+                      RX_PKT_IS_ETHERNETV2 | RX_PKT_LAYER4_TYPE_MASK |
+                      RX_PKT_IS_VLAN_TAGGED)) !=
+           (RX_IP_HDR_OK | RX_PKT_IS_IPV4 |
+            RX_PKT_IS_ETHERNETV2 | RX_PKT_LAYER4_TYPE_TCP_IPV4))
+               return -1;
+
+       skb_reset_network_header(skb);
+       skb_set_transport_header(skb, ip_hdrlen(skb));
+       *iphdr = ip_hdr(skb);
+       *tcph = tcp_hdr(skb);
+       *hdr_flags = LRO_IPV4 | LRO_TCP;
+
+       return 0;
+}
+#endif
+
 static int rxq_process(struct rx_queue *rxq, int budget)
 {
        struct mv643xx_eth_private *mp = rxq_to_mp(rxq);
        struct net_device_stats *stats = &mp->dev->stats;
+       int lro_flush_needed;
        int rx;
 
+       lro_flush_needed = 0;
        rx = 0;
        while (rx < budget && rxq->rx_desc_count) {
                struct rx_desc *rx_desc;
@@ -561,7 +611,15 @@ static int rxq_process(struct rx_queue *rxq, int budget)
                if (cmd_sts & LAYER_4_CHECKSUM_OK)
                        skb->ip_summed = CHECKSUM_UNNECESSARY;
                skb->protocol = eth_type_trans(skb, mp->dev);
-               netif_receive_skb(skb);
+
+#ifdef CONFIG_MV643XX_ETH_LRO
+               if (skb->dev->features & NETIF_F_LRO &&
+                   skb->ip_summed == CHECKSUM_UNNECESSARY) {
+                       lro_receive_skb(&rxq->lro_mgr, skb, (void *)cmd_sts);
+                       lro_flush_needed = 1;
+               } else
+#endif
+                       netif_receive_skb(skb);
 
                continue;
 
@@ -582,6 +640,11 @@ err:
                dev_kfree_skb(skb);
        }
 
+#ifdef CONFIG_MV643XX_ETH_LRO
+       if (lro_flush_needed)
+               lro_flush_all(&rxq->lro_mgr);
+#endif
+
        if (rx < budget)
                mp->work_rx &= ~(1 << rxq->index);
 
@@ -1161,6 +1224,28 @@ static struct net_device_stats *mv643xx_eth_get_stats(struct net_device *dev)
        return stats;
 }
 
+static void mv643xx_eth_grab_lro_stats(struct mv643xx_eth_private *mp)
+{
+       u32 lro_aggregated = 0;
+       u32 lro_flushed = 0;
+       u32 lro_no_desc = 0;
+       int i;
+
+#ifdef CONFIG_MV643XX_ETH_LRO
+       for (i = 0; i < mp->rxq_count; i++) {
+               struct rx_queue *rxq = mp->rxq + i;
+
+               lro_aggregated += rxq->lro_mgr.stats.aggregated;
+               lro_flushed += rxq->lro_mgr.stats.flushed;
+               lro_no_desc += rxq->lro_mgr.stats.no_desc;
+       }
+#endif
+
+       mp->lro_counters.lro_aggregated = lro_aggregated;
+       mp->lro_counters.lro_flushed = lro_flushed;
+       mp->lro_counters.lro_no_desc = lro_no_desc;
+}
+
 static inline u32 mib_read(struct mv643xx_eth_private *mp, int offset)
 {
        return rdl(mp, MIB_COUNTERS(mp->port_num) + offset);
@@ -1319,6 +1404,10 @@ struct mv643xx_eth_stats {
        { #m, FIELD_SIZEOF(struct mib_counters, m),             \
          -1, offsetof(struct mv643xx_eth_private, mib_counters.m) }
 
+#define LROSTAT(m)                                             \
+       { #m, FIELD_SIZEOF(struct lro_counters, m),             \
+         -1, offsetof(struct mv643xx_eth_private, lro_counters.m) }
+
 static const struct mv643xx_eth_stats mv643xx_eth_stats[] = {
        SSTAT(rx_packets),
        SSTAT(tx_packets),
@@ -1358,6 +1447,9 @@ static const struct mv643xx_eth_stats mv643xx_eth_stats[] = {
        MIBSTAT(bad_crc_event),
        MIBSTAT(collision),
        MIBSTAT(late_collision),
+       LROSTAT(lro_aggregated),
+       LROSTAT(lro_flushed),
+       LROSTAT(lro_no_desc),
 };
 
 static int
@@ -1569,6 +1661,7 @@ static void mv643xx_eth_get_ethtool_stats(struct net_device *dev,
 
        mv643xx_eth_get_stats(dev);
        mib_counters_update(mp);
+       mv643xx_eth_grab_lro_stats(mp);
 
        for (i = 0; i < ARRAY_SIZE(mv643xx_eth_stats); i++) {
                const struct mv643xx_eth_stats *stat;
@@ -1610,6 +1703,8 @@ static const struct ethtool_ops mv643xx_eth_ethtool_ops = {
        .set_sg                 = ethtool_op_set_sg,
        .get_strings            = mv643xx_eth_get_strings,
        .get_ethtool_stats      = mv643xx_eth_get_ethtool_stats,
+       .get_flags              = ethtool_op_get_flags,
+       .set_flags              = ethtool_op_set_flags,
        .get_sset_count         = mv643xx_eth_get_sset_count,
 };
 
@@ -1844,6 +1939,21 @@ static int rxq_init(struct mv643xx_eth_private *mp, int index)
                                        nexti * sizeof(struct rx_desc);
        }
 
+#ifdef CONFIG_MV643XX_ETH_LRO
+       rxq->lro_mgr.dev = mp->dev;
+       memset(&rxq->lro_mgr.stats, 0, sizeof(rxq->lro_mgr.stats));
+       rxq->lro_mgr.features = LRO_F_NAPI;
+       rxq->lro_mgr.ip_summed = CHECKSUM_UNNECESSARY;
+       rxq->lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
+       rxq->lro_mgr.max_desc = ARRAY_SIZE(rxq->lro_arr);
+       rxq->lro_mgr.max_aggr = 32;
+       rxq->lro_mgr.frag_align_pad = 0;
+       rxq->lro_mgr.lro_arr = rxq->lro_arr;
+       rxq->lro_mgr.get_skb_header = mv643xx_get_skb_header;
+
+       memset(&rxq->lro_arr, 0, sizeof(rxq->lro_arr));
+#endif
+
        return 0;