[PATCH] bonding: inherit zero-copy flags of slaves
authorArthur Kepner <akepner@sgi.com>
Tue, 23 Aug 2005 05:34:53 +0000 (01:34 -0400)
committerJeff Garzik <jgarzik@pobox.com>
Tue, 23 Aug 2005 05:34:53 +0000 (01:34 -0400)
This change allows a bonding device to inherit the "zero-copy"
features of its slave devices.

It was inspired by a couple of previous postings on this topic:
http://marc.theaimsgroup.com/?l=bonding-devel&m=111924607327794&w=2
http://marc.theaimsgroup.com/?l=bonding-devel&m=111925242706297&w=2
and it's largely a combination of the patches that appear in those
emails.

Signed-off-by: Arthur Kepner <akepner@sgi.com>
drivers/net/bonding/bond_main.c
drivers/net/bonding/bonding.h

index 2c930da90a854d957f881cff3b3032ce0b1ff921..94c9f68dd16bacf4686eac5cb7da66062bf8df33 100644 (file)
@@ -1604,6 +1604,44 @@ static int bond_sethwaddr(struct net_device *bond_dev, struct net_device *slave_
        return 0;
 }
 
+#define BOND_INTERSECT_FEATURES \
+       (NETIF_F_SG|NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM)
+
+/* 
+ * Compute the features available to the bonding device by 
+ * intersection of all of the slave devices' BOND_INTERSECT_FEATURES.
+ * Call this after attaching or detaching a slave to update the 
+ * bond's features.
+ */
+static int bond_compute_features(struct bonding *bond)
+{
+       int i;
+       struct slave *slave;
+       struct net_device *bond_dev = bond->dev;
+       int features = bond->bond_features;
+
+       bond_for_each_slave(bond, slave, i) {
+               struct net_device * slave_dev = slave->dev;
+               if (i == 0) {
+                       features |= BOND_INTERSECT_FEATURES;
+               }
+               features &=
+                       ~(~slave_dev->features & BOND_INTERSECT_FEATURES);
+       }
+
+       /* turn off NETIF_F_SG if we need a csum and h/w can't do it */
+       if ((features & NETIF_F_SG) && 
+               !(features & (NETIF_F_IP_CSUM |
+                             NETIF_F_NO_CSUM |
+                             NETIF_F_HW_CSUM))) {
+               features &= ~NETIF_F_SG;
+       }
+
+       bond_dev->features = features;
+
+       return 0;
+}
+
 /* enslave device <slave> to bond device <master> */
 static int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev)
 {
@@ -1811,6 +1849,8 @@ static int bond_enslave(struct net_device *bond_dev, struct net_device *slave_de
        new_slave->delay = 0;
        new_slave->link_failure_count = 0;
 
+       bond_compute_features(bond);
+
        if (bond->params.miimon && !bond->params.use_carrier) {
                link_reporting = bond_check_dev_link(bond, slave_dev, 1);
 
@@ -2015,7 +2055,7 @@ err_free:
 
 err_undo_flags:
        bond_dev->features = old_features;
-
        return res;
 }
 
@@ -2100,6 +2140,8 @@ static int bond_release(struct net_device *bond_dev, struct net_device *slave_de
        /* release the slave from its bond */
        bond_detach_slave(bond, slave);
 
+       bond_compute_features(bond);
+
        if (bond->primary_slave == slave) {
                bond->primary_slave = NULL;
        }
@@ -2243,6 +2285,8 @@ static int bond_release_all(struct net_device *bond_dev)
                        bond_alb_deinit_slave(bond, slave);
                }
 
+               bond_compute_features(bond);
+
                /* now that the slave is detached, unlock and perform
                 * all the undo steps that should not be called from
                 * within a lock.
@@ -3588,6 +3632,7 @@ static int bond_master_netdev_event(unsigned long event, struct net_device *bond
 static int bond_slave_netdev_event(unsigned long event, struct net_device *slave_dev)
 {
        struct net_device *bond_dev = slave_dev->master;
+       struct bonding *bond = bond_dev->priv;
 
        switch (event) {
        case NETDEV_UNREGISTER:
@@ -3626,6 +3671,9 @@ static int bond_slave_netdev_event(unsigned long event, struct net_device *slave
                 * TODO: handle changing the primary's name
                 */
                break;
+       case NETDEV_FEAT_CHANGE:
+               bond_compute_features(bond);
+               break;
        default:
                break;
        }
@@ -4526,6 +4574,11 @@ static inline void bond_set_mode_ops(struct bonding *bond, int mode)
        }
 }
 
+static struct ethtool_ops bond_ethtool_ops = {
+       .get_tx_csum            = ethtool_op_get_tx_csum,
+       .get_sg                 = ethtool_op_get_sg,
+};
+
 /*
  * Does not allocate but creates a /proc entry.
  * Allowed to fail.
@@ -4555,6 +4608,7 @@ static int __init bond_init(struct net_device *bond_dev, struct bond_params *par
        bond_dev->stop = bond_close;
        bond_dev->get_stats = bond_get_stats;
        bond_dev->do_ioctl = bond_do_ioctl;
+       bond_dev->ethtool_ops = &bond_ethtool_ops;
        bond_dev->set_multicast_list = bond_set_multicast_list;
        bond_dev->change_mtu = bond_change_mtu;
        bond_dev->set_mac_address = bond_set_mac_address;
@@ -4591,6 +4645,8 @@ static int __init bond_init(struct net_device *bond_dev, struct bond_params *par
                               NETIF_F_HW_VLAN_RX |
                               NETIF_F_HW_VLAN_FILTER);
 
+       bond->bond_features = bond_dev->features;
+
 #ifdef CONFIG_PROC_FS
        bond_create_proc_entry(bond);
 #endif
index d27f377b3eeb52bf2f8a47c3f3b6a5a61463faad..3881969808627cebcb7ed6a25f6b9b3557b6a1ec 100644 (file)
@@ -211,6 +211,9 @@ struct bonding {
        struct   bond_params params;
        struct   list_head vlan_list;
        struct   vlan_group *vlgrp;
+       /* the features the bonding device supports, independently 
+        * of any slaves */
+       int      bond_features; 
 };
 
 /**