X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=datapath%2Fdatapath.c;h=1ae1d771a31a5a3f2128ee75f474cf9e2c645f1d;hb=3f355f47f8e7343e909ccfa854454d667baf3c38;hp=aea8cc95c8bd8ac0f2a36f6143a319655ee29e51;hpb=2416b8eceae7b2508fe72efbc17d9cb71b69d330;p=sliver-openvswitch.git diff --git a/datapath/datapath.c b/datapath/datapath.c index aea8cc95c..1ae1d771a 100644 --- a/datapath/datapath.c +++ b/datapath/datapath.c @@ -56,9 +56,7 @@ int (*dp_ioctl_hook)(struct net_device *dev, struct ifreq *rq, int cmd); EXPORT_SYMBOL(dp_ioctl_hook); /* Datapaths. Protected on the read side by rcu_read_lock, on the write side - * by dp_mutex. dp_mutex is almost completely redundant with genl_mutex - * maintained by the Generic Netlink code, but the timeout path needs mutual - * exclusion too. + * by dp_mutex. * * dp_mutex nests inside the RTNL lock: if you need both you must take the RTNL * lock first. @@ -225,9 +223,7 @@ static int create_dp(int dp_idx, const char __user *devnamep) /* Initialize kobject for bridge. This will be added as * /sys/class/net//brif later, if sysfs is enabled. */ - kobject_set_name(&dp->ifobj, SYSFS_BRIDGE_PORT_SUBDIR); /* "brif" */ dp->ifobj.kset = NULL; - dp->ifobj.parent = NULL; kobject_init(&dp->ifobj, &dp_ktype); /* Allocate table. */ @@ -257,9 +253,7 @@ static int create_dp(int dp_idx, const char __user *devnamep) mutex_unlock(&dp_mutex); rtnl_unlock(); -#ifdef SUPPORT_SYSFS dp_sysfs_add_dp(dp); -#endif return 0; @@ -287,9 +281,7 @@ static void do_destroy_dp(struct datapath *dp) if (p->port_no != ODPP_LOCAL) dp_del_port(p); -#ifdef SUPPORT_SYSFS dp_sysfs_del_dp(dp); -#endif rcu_assign_pointer(dps[dp->dp_idx], NULL); @@ -334,7 +326,7 @@ static void release_nbp(struct kobject *kobj) } struct kobj_type brport_ktype = { -#ifdef SUPPORT_SYSFS +#ifdef CONFIG_SYSFS .sysfs_ops = &brport_sysfs_ops, #endif .release = release_nbp @@ -371,9 +363,7 @@ static int new_nbp(struct datapath *dp, struct net_device *dev, int port_no) /* Initialize kobject for bridge. This will be added as * /sys/class/net//brport later, if sysfs is enabled. */ - kobject_set_name(&p->kobj, SYSFS_BRIDGE_PORT_ATTR); /* "brport" */ p->kobj.kset = NULL; - p->kobj.parent = &p->dev->NETDEV_DEV_MEMBER.kobj; kobject_init(&p->kobj, &brport_ktype); dp_ifinfo_notify(RTM_NEWLINK, p); @@ -393,11 +383,6 @@ static int add_port(int dp_idx, struct odp_port __user *portp) if (copy_from_user(&port, portp, sizeof port)) goto out; port.devname[IFNAMSIZ - 1] = '\0'; - port_no = port.port; - - err = -EINVAL; - if (port_no < 0 || port_no >= DP_MAX_PORTS) - goto out; rtnl_lock(); dp = get_dp_locked(dp_idx); @@ -405,10 +390,13 @@ static int add_port(int dp_idx, struct odp_port __user *portp) if (!dp) goto out_unlock_rtnl; - err = -EEXIST; - if (dp->ports[port_no]) - goto out_unlock_dp; + for (port_no = 1; port_no < DP_MAX_PORTS; port_no++) + if (!dp->ports[port_no]) + goto got_port_no; + err = -EFBIG; + goto out_unlock_dp; +got_port_no: if (!(port.flags & ODP_PORT_INTERNAL)) { err = -ENODEV; dev = dev_get_by_name(&init_net, port.devname); @@ -431,9 +419,9 @@ static int add_port(int dp_idx, struct odp_port __user *portp) if (err) goto out_put; -#ifdef SUPPORT_SYSFS dp_sysfs_add_if(dp->ports[port_no]); -#endif + + err = __put_user(port_no, &port.port); out_put: dev_put(dev); @@ -449,10 +437,8 @@ int dp_del_port(struct net_bridge_port *p) { ASSERT_RTNL(); -#ifdef SUPPORT_SYSFS if (p->port_no != ODPP_LOCAL) dp_sysfs_del_if(p); -#endif dp_ifinfo_notify(RTM_DELLINK, p); p->dp->n_ports--; @@ -588,7 +574,7 @@ static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb) #error #endif -#ifdef CONFIG_XEN +#if defined(CONFIG_XEN) && LINUX_VERSION_CODE == KERNEL_VERSION(2,6,18) /* This code is copied verbatim from net/dev/core.c in Xen's * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call those functions * directly because they aren't exported. */ @@ -604,7 +590,7 @@ static int skb_pull_up_to(struct sk_buff *skb, void *ptr) } } -int skb_checksum_setup(struct sk_buff *skb) +int vswitch_skb_checksum_setup(struct sk_buff *skb) { if (skb->proto_csum_blank) { if (skb->protocol != htons(ETH_P_IP)) @@ -635,15 +621,97 @@ int skb_checksum_setup(struct sk_buff *skb) out: return -EPROTO; } +#else +int vswitch_skb_checksum_setup(struct sk_buff *skb) { return 0; } +#endif /* CONFIG_XEN && linux == 2.6.18 */ + +/* Append each packet in 'skb' list to 'queue'. There will be only one packet + * unless we broke up a GSO packet. */ +static int +queue_control_packets(struct sk_buff *skb, struct sk_buff_head *queue, + int queue_no, u32 arg) +{ + struct sk_buff *nskb; + int port_no; + int err; + + port_no = ODPP_LOCAL; + if (skb->dev) { + if (skb->dev->br_port) + port_no = skb->dev->br_port->port_no; + else if (is_dp_dev(skb->dev)) + port_no = dp_dev_priv(skb->dev)->port_no; + } + + do { + struct odp_msg *header; + + nskb = skb->next; + skb->next = NULL; + + /* If a checksum-deferred packet is forwarded to the + * controller, correct the pointers and checksum. This happens + * on a regular basis only on Xen, on which VMs can pass up + * packets that do not have their checksum computed. + */ + err = vswitch_skb_checksum_setup(skb); + if (err) + goto err_kfree_skbs; +#ifndef CHECKSUM_HW + if (skb->ip_summed == CHECKSUM_PARTIAL) { +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22) + /* Until 2.6.22, the start of the transport header was + * also the start of data to be checksummed. Linux + * 2.6.22 introduced the csum_start field for this + * purpose, but we should point the transport header to + * it anyway for backward compatibility, as + * dev_queue_xmit() does even in 2.6.28. */ + skb_set_transport_header(skb, skb->csum_start - + skb_headroom(skb)); +#endif + err = skb_checksum_help(skb); + if (err) + goto err_kfree_skbs; + } +#else + if (skb->ip_summed == CHECKSUM_HW) { + err = skb_checksum_help(skb, 0); + if (err) + goto err_kfree_skbs; + } #endif + err = skb_cow(skb, sizeof *header); + if (err) + goto err_kfree_skbs; + + header = (struct odp_msg*)__skb_push(skb, sizeof *header); + header->type = queue_no; + header->length = skb->len; + header->port = port_no; + header->reserved = 0; + header->arg = arg; + skb_queue_tail(queue, skb); + + skb = nskb; + } while (skb); + return 0; + +err_kfree_skbs: + kfree_skb(skb); + while ((skb = nskb) != NULL) { + nskb = skb->next; + kfree_skb(skb); + } + return err; +} + int dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no, u32 arg) { struct dp_stats_percpu *stats; struct sk_buff_head *queue; - int port_no; int err; WARN_ON_ONCE(skb_shared(skb)); @@ -654,40 +722,6 @@ dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no, if (skb_queue_len(queue) >= DP_MAX_QUEUE_LEN) goto err_kfree_skb; - /* If a checksum-deferred packet is forwarded to the controller, - * correct the pointers and checksum. This happens on a regular basis - * only on Xen (the CHECKSUM_HW case), on which VMs can pass up packets - * that do not have their checksum computed. We also implement it for - * the non-Xen case, but it is difficult to trigger or test this case - * there, hence the WARN_ON_ONCE(). - */ - err = skb_checksum_setup(skb); - if (err) - goto err_kfree_skb; -#ifndef CHECKSUM_HW - if (skb->ip_summed == CHECKSUM_PARTIAL) { - WARN_ON_ONCE(1); -#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22) - /* Until 2.6.22, the start of the transport header was also the - * start of data to be checksummed. Linux 2.6.22 introduced - * the csum_start field for this purpose, but we should point - * the transport header to it anyway for backward - * compatibility, as dev_queue_xmit() does even in 2.6.28. */ - skb_set_transport_header(skb, skb->csum_start - - skb_headroom(skb)); -#endif - err = skb_checksum_help(skb); - if (err) - goto err_kfree_skb; - } -#else - if (skb->ip_summed == CHECKSUM_HW) { - err = skb_checksum_help(skb, 0); - if (err) - goto err_kfree_skb; - } -#endif - /* Break apart GSO packets into their component pieces. Otherwise * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */ if (skb_is_gso(skb)) { @@ -705,45 +739,9 @@ dp_output_control(struct datapath *dp, struct sk_buff *skb, int queue_no, } } - /* Figure out port number. */ - port_no = ODPP_LOCAL; - if (skb->dev) { - if (skb->dev->br_port) - port_no = skb->dev->br_port->port_no; - else if (is_dp_dev(skb->dev)) - port_no = dp_dev_priv(skb->dev)->port_no; - } - - /* Append each packet to queue. There will be only one packet unless - * we broke up a GSO packet above. */ - do { - struct odp_msg *header; - struct sk_buff *nskb = skb->next; - skb->next = NULL; - - err = skb_cow(skb, sizeof *header); - if (err) { - while (nskb) { - kfree_skb(skb); - skb = nskb; - nskb = skb->next; - } - goto err_kfree_skb; - } - - header = (struct odp_msg*)__skb_push(skb, sizeof *header); - header->type = queue_no; - header->length = skb->len; - header->port = port_no; - header->reserved = 0; - header->arg = arg; - skb_queue_tail(queue, skb); - - skb = nskb; - } while (skb); - + err = queue_control_packets(skb, queue, queue_no, arg); wake_up_interruptible(&dp->waitqueue); - return 0; + return err; err_kfree_skb: kfree_skb(skb); @@ -838,6 +836,7 @@ static void get_stats(struct sw_flow *flow, struct odp_flow_stats *stats) stats->n_bytes = flow->byte_count; stats->ip_tos = flow->ip_tos; stats->tcp_flags = flow->tcp_flags; + stats->error = 0; } static void clear_stats(struct sw_flow *flow) @@ -965,8 +964,6 @@ static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp) if (!n_actions) return 0; - if (ufp->n_actions > INT_MAX / sizeof(union odp_action)) - return -EINVAL; sf_acts = rcu_dereference(flow->sf_acts); if (__put_user(sf_acts->n_actions, &ufp->n_actions) || @@ -992,9 +989,7 @@ static int answer_query(struct sw_flow *flow, struct odp_flow __user *ufp) return put_actions(flow, ufp); } -static int del_or_query_flow(struct datapath *dp, - struct odp_flow __user *ufp, - unsigned int cmd) +static int del_flow(struct datapath *dp, struct odp_flow __user *ufp) { struct dp_table *table = rcu_dereference(dp->table); struct odp_flow uf; @@ -1011,29 +1006,24 @@ static int del_or_query_flow(struct datapath *dp, if (!flow) goto error; - if (cmd == ODP_FLOW_DEL) { - /* XXX redundant lookup */ - error = dp_table_delete(table, flow); - if (error) - goto error; + /* XXX redundant lookup */ + error = dp_table_delete(table, flow); + if (error) + goto error; - /* XXX These statistics might lose a few packets, since other - * CPUs can be using this flow. We used to synchronize_rcu() - * to make sure that we get completely accurate stats, but that - * blows our performance, badly. */ - dp->n_flows--; - error = answer_query(flow, ufp); - flow_deferred_free(flow); - } else { - error = answer_query(flow, ufp); - } + /* XXX These statistics might lose a few packets, since other CPUs can + * be using this flow. We used to synchronize_rcu() to make sure that + * we get completely accurate stats, but that blows our performance, + * badly. */ + dp->n_flows--; + error = answer_query(flow, ufp); + flow_deferred_free(flow); error: return error; } -static int query_multiple_flows(struct datapath *dp, - const struct odp_flowvec *flowvec) +static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec) { struct dp_table *table = rcu_dereference(dp->table); int i; @@ -1049,7 +1039,7 @@ static int query_multiple_flows(struct datapath *dp, flow = dp_table_lookup(table, &uf.key); if (!flow) - error = __clear_user(&ufp->stats, sizeof ufp->stats); + error = __put_user(ENOENT, &ufp->stats.error); else error = answer_query(flow, ufp); if (error) @@ -1194,8 +1184,7 @@ error: return err; } -static int -get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp) +static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp) { struct odp_stats stats; int i; @@ -1310,7 +1299,7 @@ list_ports(struct datapath *dp, struct odp_portvec __user *pvp) break; } } - return put_user(idx, &pvp->n_ports); + return put_user(dp->n_ports, &pvp->n_ports); } /* RCU callback for freeing a dp_port_group */ @@ -1404,24 +1393,28 @@ static long openvswitch_ioctl(struct file *f, unsigned int cmd, /* Handle commands with special locking requirements up front. */ switch (cmd) { case ODP_DP_CREATE: - return create_dp(dp_idx, (char __user *)argp); + err = create_dp(dp_idx, (char __user *)argp); + goto exit; case ODP_DP_DESTROY: - return destroy_dp(dp_idx); + err = destroy_dp(dp_idx); + goto exit; case ODP_PORT_ADD: - return add_port(dp_idx, (struct odp_port __user *)argp); + err = add_port(dp_idx, (struct odp_port __user *)argp); + goto exit; case ODP_PORT_DEL: err = get_user(port_no, (int __user *)argp); - if (err) - break; - return del_port(dp_idx, port_no); + if (!err) + err = del_port(dp_idx, port_no); + goto exit; } dp = get_dp_locked(dp_idx); + err = -ENODEV; if (!dp) - return -ENODEV; + goto exit; switch (cmd) { case ODP_DP_STATS: @@ -1483,13 +1476,11 @@ static long openvswitch_ioctl(struct file *f, unsigned int cmd, break; case ODP_FLOW_DEL: - case ODP_FLOW_GET: - err = del_or_query_flow(dp, (struct odp_flow __user *)argp, - cmd); + err = del_flow(dp, (struct odp_flow __user *)argp); break; - case ODP_FLOW_GET_MULTIPLE: - err = do_flowvec_ioctl(dp, argp, query_multiple_flows); + case ODP_FLOW_GET: + err = do_flowvec_ioctl(dp, argp, query_flows); break; case ODP_FLOW_LIST: @@ -1505,6 +1496,7 @@ static long openvswitch_ioctl(struct file *f, unsigned int cmd, break; } mutex_unlock(&dp->mutex); +exit: return err; } @@ -1601,6 +1593,8 @@ struct file_operations openvswitch_fops = { }; static int major; + +#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27) static struct llc_sap *dp_stp_sap; static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev, @@ -1613,12 +1607,8 @@ static int dp_stp_rcv(struct sk_buff *skb, struct net_device *dev, return 0; } -static int __init dp_init(void) +static int dp_avoid_bridge_init(void) { - int err; - - printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR); - /* Register to receive STP packets because the bridge module also * attempts to do so. Since there can only be a single listener for a * given protocol, this provides mutual exclusion against the bridge @@ -1629,6 +1619,43 @@ static int __init dp_init(void) printk(KERN_ERR "openvswitch: can't register sap for STP (probably the bridge module is loaded)\n"); return -EADDRINUSE; } + return 0; +} + +static void dp_avoid_bridge_exit(void) +{ + llc_sap_put(dp_stp_sap); +} +#else /* Linux 2.6.27 or later. */ +static int dp_avoid_bridge_init(void) +{ + /* Linux 2.6.27 introduces a way for multiple clients to register for + * STP packets, which interferes with what we try to do above. + * Instead, just check whether there's a bridge hook defined. This is + * not as safe--the bridge module is willing to load over the top of + * us--but it provides a little bit of protection. */ + if (br_handle_frame_hook) { + printk(KERN_ERR "openvswitch: bridge module is loaded, cannot load over it\n"); + return -EADDRINUSE; + } + return 0; +} + +static void dp_avoid_bridge_exit(void) +{ + /* Nothing to do. */ +} +#endif /* Linux 2.6.27 or later */ + +static int __init dp_init(void) +{ + int err; + + printk("Open vSwitch %s, built "__DATE__" "__TIME__"\n", VERSION BUILDNR); + + err = dp_avoid_bridge_init(); + if (err) + return err; err = flow_init(); if (err) @@ -1663,7 +1690,7 @@ static void dp_cleanup(void) unregister_netdevice_notifier(&dp_device_notifier); flow_exit(); br_handle_frame_hook = NULL; - llc_sap_put(dp_stp_sap); + dp_avoid_bridge_exit(); } module_init(dp_init);