X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=datapath%2Fdatapath.c;h=539859a40c908a2698d12fa9970e70f484809ac6;hb=8cdaca99b5742489c35b1cac6e39791a47451952;hp=9d163ea5f07c179fd73a537490e785b0b3e4919f;hpb=3c71830aef987c86338d2d0223d4e6913aea25a4;p=openvswitch diff --git a/datapath/datapath.c b/datapath/datapath.c index 9d163ea5..539859a4 100644 --- a/datapath/datapath.c +++ b/datapath/datapath.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2007, 2008, 2009 Nicira Networks. + * Copyright (c) 2007, 2008, 2009, 2010 Nicira Networks. * Distributed under the terms of the GNU GPL version 2. * * Significant portions of this file may be copied from parts of the Linux @@ -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. @@ -234,7 +232,7 @@ static int create_dp(int dp_idx, const char __user *devnamep) if (!dp->table) goto err_free_dp; - /* Setup our datapath device */ + /* Set up our datapath device. */ dp_dev = dp_dev_create(dp, devname, ODPP_LOCAL); err = PTR_ERR(dp_dev); if (IS_ERR(dp_dev)) @@ -576,10 +574,11 @@ static int dp_frame_hook(struct net_bridge_port *p, struct sk_buff **pskb) #error #endif -#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. */ +#if defined(CONFIG_XEN) && defined(HAVE_PROTO_DATA_VALID) +/* This code is based on a skb_checksum_setup from net/dev/core.c from a + * combination of Lenny's 2.6.26 Xen kernel and Xen's + * linux-2.6.18-92.1.10.el5.xs5.0.0.394.644. We can't call this function + * directly because it isn't exported in all versions. */ static int skb_pull_up_to(struct sk_buff *skb, void *ptr) { if (ptr < (void *)skb->tail) @@ -594,46 +593,205 @@ static int skb_pull_up_to(struct sk_buff *skb, void *ptr) int vswitch_skb_checksum_setup(struct sk_buff *skb) { - if (skb->proto_csum_blank) { - if (skb->protocol != htons(ETH_P_IP)) - goto out; - if (!skb_pull_up_to(skb, skb->nh.iph + 1)) - goto out; - skb->h.raw = (unsigned char *)skb->nh.iph + 4*skb->nh.iph->ihl; - switch (skb->nh.iph->protocol) { - case IPPROTO_TCP: - skb->csum = offsetof(struct tcphdr, check); - break; - case IPPROTO_UDP: - skb->csum = offsetof(struct udphdr, check); - break; - default: - if (net_ratelimit()) - printk(KERN_ERR "Attempting to checksum a non-" - "TCP/UDP packet, dropping a protocol" - " %d packet", skb->nh.iph->protocol); - goto out; - } - if (!skb_pull_up_to(skb, skb->h.raw + skb->csum + 2)) - goto out; - skb->ip_summed = CHECKSUM_HW; - skb->proto_csum_blank = 0; + struct iphdr *iph; + unsigned char *th; + int err = -EPROTO; + __u16 csum_start, csum_offset; + + if (!skb->proto_csum_blank) + return 0; + + if (skb->protocol != htons(ETH_P_IP)) + goto out; + + if (!skb_pull_up_to(skb, skb_network_header(skb) + 1)) + goto out; + + iph = ip_hdr(skb); + th = skb_network_header(skb) + 4 * iph->ihl; + + csum_start = th - skb->head; + switch (iph->protocol) { + case IPPROTO_TCP: + csum_offset = offsetof(struct tcphdr, check); + break; + case IPPROTO_UDP: + csum_offset = offsetof(struct udphdr, check); + break; + default: + if (net_ratelimit()) + printk(KERN_ERR "Attempting to checksum a non-" + "TCP/UDP packet, dropping a protocol" + " %d packet", iph->protocol); + goto out; } - return 0; + + if (!skb_pull_up_to(skb, th + csum_offset + 2)) + goto out; + + skb->ip_summed = CHECKSUM_PARTIAL; + skb->proto_csum_blank = 0; + +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,22) + skb->csum_start = csum_start; + skb->csum_offset = csum_offset; +#else + skb_set_transport_header(skb, csum_start - skb_headroom(skb)); + skb->csum = csum_offset; +#endif + + err = 0; + out: - return -EPROTO; + return err; } #else int vswitch_skb_checksum_setup(struct sk_buff *skb) { return 0; } #endif /* CONFIG_XEN && linux == 2.6.18 */ + /* Types of checksums that we can receive (these all refer to L4 checksums): + * 1. CHECKSUM_NONE: Device that did not compute checksum, contains full + * (though not verified) checksum in packet but not in skb->csum. Packets + * from the bridge local port will also have this type. + * 2. CHECKSUM_COMPLETE (CHECKSUM_HW): Good device that computes checksums, + * also the GRE module. This is the same as CHECKSUM_NONE, except it has + * a valid skb->csum. Importantly, both contain a full checksum (not + * verified) in the packet itself. The only difference is that if the + * packet gets to L4 processing on this machine (not in DomU) we won't + * have to recompute the checksum to verify. Most hardware devices do not + * produce packets with this type, even if they support receive checksum + * offloading (they produce type #5). + * 3. CHECKSUM_PARTIAL (CHECKSUM_HW): Packet without full checksum and needs to + * be computed if it is sent off box. Unfortunately on earlier kernels, + * this case is impossible to distinguish from #2, despite having opposite + * meanings. Xen adds an extra field on earlier kernels (see #4) in order + * to distinguish the different states. The only real user of this type + * with bridging is Xen (on later kernels). + * 4. CHECKSUM_UNNECESSARY (with proto_csum_blank true): This packet was + * generated locally by a Xen DomU and has a partial checksum. If it is + * handled on this machine (Dom0 or DomU), then the checksum will not be + * computed. If it goes off box, the checksum in the packet needs to + * completed. Calling skb_checksum_setup converts this to CHECKSUM_HW + * (CHECKSUM_PARTIAL) so that the checksum can be completed. In later + * kernels, this combination is replaced with CHECKSUM_PARTIAL. + * 5. CHECKSUM_UNNECESSARY (with proto_csum_blank false): Packet with a correct + * full checksum or using a protocol without a checksum. skb->csum is + * undefined. This is common from devices with receive checksum + * offloading. This is somewhat similar to CHECKSUM_NONE, except that + * nobody will try to verify the checksum with CHECKSUM_UNNECESSARY. + * + * Note that on earlier kernels, CHECKSUM_COMPLETE and CHECKSUM_PARTIAL are + * both defined as CHECKSUM_HW. Normally the meaning of CHECKSUM_HW is clear + * based on whether it is on the transmit or receive path. After the datapath + * it will be intepreted as CHECKSUM_PARTIAL. If the packet already has a + * checksum, we will panic. Since we can receive packets with checksums, we + * assume that all CHECKSUM_HW packets have checksums and map them to + * CHECKSUM_NONE, which has a similar meaning (the it is only different if the + * packet is processed by the local IP stack, in which case it will need to + * be reverified). If we receive a packet with CHECKSUM_HW that really means + * CHECKSUM_PARTIAL, it will be sent with the wrong checksum. However, there + * shouldn't be any devices that do this with bridging. + * + * The bridge has similar behavior and this function closely resembles + * skb_forward_csum(). It is slightly different because we are only concerned + * with bridging and not other types of forwarding and can get away with + * slightly more optimal behavior.*/ +void +forward_ip_summed(struct sk_buff *skb) +{ +#ifdef CHECKSUM_HW + if (skb->ip_summed == CHECKSUM_HW) + skb->ip_summed = CHECKSUM_NONE; +#endif +} + +/* 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)); @@ -644,39 +802,7 @@ 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 = vswitch_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 + forward_ip_summed(skb); /* Break apart GSO packets into their component pieces. Otherwise * userspace may try to stuff a 64kB packet into a 1500-byte MTU. */ @@ -695,45 +821,9 @@ 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); @@ -774,7 +864,8 @@ static int validate_actions(const struct sw_flow_actions *actions) break; case ODPAT_SET_VLAN_PCP: - if (a->vlan_pcp.vlan_pcp & ~VLAN_PCP_MASK) + if (a->vlan_pcp.vlan_pcp + & ~(VLAN_PCP_MASK >> VLAN_PCP_SHIFT)) return -EINVAL; break; @@ -842,7 +933,7 @@ static void clear_stats(struct sw_flow *flow) static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp) { struct odp_flow_put uf; - struct sw_flow *flow, **bucket; + struct sw_flow *flow; struct dp_table *table; struct odp_flow_stats stats; int error; @@ -852,15 +943,10 @@ static int put_flow(struct datapath *dp, struct odp_flow_put __user *ufp) goto error; uf.flow.key.reserved = 0; -retry: table = rcu_dereference(dp->table); - bucket = dp_table_lookup_for_insert(table, &uf.flow.key); - if (!bucket) { - /* No such flow, and the slots where it could go are full. */ - error = uf.flags & ODPPF_CREATE ? -EFBIG : -ENOENT; - goto error; - } else if (!*bucket) { - /* No such flow, but we found an available slot for it. */ + flow = dp_table_lookup(table, &uf.flow.key); + if (!flow) { + /* No such flow. */ struct sw_flow_actions *acts; error = -ENOENT; @@ -868,14 +954,15 @@ retry: goto error; /* Expand table, if necessary, to make room. */ - if (dp->n_flows * 4 >= table->n_buckets && - table->n_buckets < DP_MAX_BUCKETS) { + if (dp->n_flows >= table->n_buckets) { + error = -ENOSPC; + if (table->n_buckets >= DP_MAX_BUCKETS) + goto error; + error = dp_table_expand(dp); if (error) goto error; - - /* The bucket's location has changed. Try again. */ - goto retry; + table = rcu_dereference(dp->table); } /* Allocate flow. */ @@ -895,12 +982,13 @@ retry: rcu_assign_pointer(flow->sf_acts, acts); /* Put flow in bucket. */ - rcu_assign_pointer(*bucket, flow); + error = dp_table_insert(table, flow); + if (error) + goto error_free_flow_acts; dp->n_flows++; memset(&stats, 0, sizeof(struct odp_flow_stats)); } else { /* We found a matching flow. */ - struct sw_flow *flow = *rcu_dereference(bucket); struct sw_flow_actions *old_acts, *new_acts; unsigned long int flags; @@ -938,6 +1026,8 @@ retry: return -EFAULT; return 0; +error_free_flow_acts: + kfree(flow->sf_acts); error_free_flow: kmem_cache_free(flow_cache, flow); error: @@ -967,13 +1057,18 @@ static int put_actions(const struct sw_flow *flow, struct odp_flow __user *ufp) return 0; } -static int answer_query(struct sw_flow *flow, struct odp_flow __user *ufp) +static int answer_query(struct sw_flow *flow, u32 query_flags, + struct odp_flow __user *ufp) { struct odp_flow_stats stats; unsigned long int flags; spin_lock_irqsave(&flow->lock, flags); get_stats(flow, &stats); + + if (query_flags & ODPFF_ZERO_TCP_FLAGS) { + flow->tcp_flags = 0; + } spin_unlock_irqrestore(&flow->lock, flags); if (__copy_to_user(&ufp->stats, &stats, sizeof(struct odp_flow_stats))) @@ -1008,7 +1103,7 @@ static int del_flow(struct datapath *dp, struct odp_flow __user *ufp) * we get completely accurate stats, but that blows our performance, * badly. */ dp->n_flows--; - error = answer_query(flow, ufp); + error = answer_query(flow, 0, ufp); flow_deferred_free(flow); error: @@ -1033,7 +1128,7 @@ static int query_flows(struct datapath *dp, const struct odp_flowvec *flowvec) if (!flow) error = __put_user(ENOENT, &ufp->stats.error); else - error = answer_query(flow, ufp); + error = answer_query(flow, uf.flags, ufp); if (error) return -EFAULT; } @@ -1054,7 +1149,7 @@ static int list_flow(struct sw_flow *flow, void *cbdata_) if (__copy_to_user(&ufp->key, &flow->key, sizeof flow->key)) return -EFAULT; - error = answer_query(flow, ufp); + error = answer_query(flow, 0, ufp); if (error) return error; @@ -1111,6 +1206,7 @@ static int do_execute(struct datapath *dp, const struct odp_execute *executep) struct odp_flow_key key; struct sk_buff *skb; struct sw_flow_actions *actions; + struct ethhdr *eth; int err; err = -EFAULT; @@ -1150,6 +1246,17 @@ static int do_execute(struct datapath *dp, const struct odp_execute *executep) execute.length)) goto error_free_skb; + skb_reset_mac_header(skb); + eth = eth_hdr(skb); + + /* Normally, setting the skb 'protocol' field would be handled by a + * call to eth_type_trans(), but it assumes there's a sending + * device, which we may not have. */ + if (ntohs(eth->h_proto) >= 1536) + skb->protocol = eth->h_proto; + else + skb->protocol = htons(ETH_P_802_2); + flow_extract(skb, execute.in_port, &key); err = execute_actions(dp, skb, &key, actions->actions, actions->n_actions, GFP_KERNEL); @@ -1170,8 +1277,8 @@ static int get_dp_stats(struct datapath *dp, struct odp_stats __user *statsp) int i; stats.n_flows = dp->n_flows; - stats.cur_capacity = rcu_dereference(dp->table)->n_buckets * 2; - stats.max_capacity = DP_MAX_BUCKETS * 2; + stats.cur_capacity = rcu_dereference(dp->table)->n_buckets; + stats.max_capacity = DP_MAX_BUCKETS; stats.n_ports = dp->n_ports; stats.max_ports = DP_MAX_PORTS; stats.max_groups = DP_MAX_GROUPS; @@ -1352,6 +1459,16 @@ get_port_group(struct datapath *dp, struct odp_port_group *upg) return 0; } +static int get_listen_mask(const struct file *f) +{ + return (long)f->private_data; +} + +static void set_listen_mask(struct file *f, int listen_mask) +{ + f->private_data = (void*)(long)listen_mask; +} + static long openvswitch_ioctl(struct file *f, unsigned int cmd, unsigned long argp) { @@ -1407,7 +1524,7 @@ static long openvswitch_ioctl(struct file *f, unsigned int cmd, break; case ODP_GET_LISTEN_MASK: - err = put_user((int)f->private_data, (int __user *)argp); + err = put_user(get_listen_mask(f), (int __user *)argp); break; case ODP_SET_LISTEN_MASK: @@ -1418,7 +1535,7 @@ static long openvswitch_ioctl(struct file *f, unsigned int cmd, if (listeners & ~ODPL_ALL) break; err = 0; - f->private_data = (void*)listeners; + set_listen_mask(f, listeners); break; case ODP_PORT_QUERY: @@ -1484,7 +1601,7 @@ ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes, loff_t *ppos) { /* XXX is there sufficient synchronization here? */ - int listeners = (int) f->private_data; + int listeners = get_listen_mask(f); int dp_idx = iminor(f->f_dentry->d_inode); struct datapath *dp = get_dp(dp_idx); struct sk_buff *skb; @@ -1524,7 +1641,7 @@ ssize_t openvswitch_read(struct file *f, char __user *buf, size_t nbytes, } } success: - copy_bytes = min(skb->len, nbytes); + copy_bytes = min_t(size_t, skb->len, nbytes); iov.iov_base = buf; iov.iov_len = copy_bytes; retval = skb_copy_datagram_iovec(skb, 0, &iov, iov.iov_len); @@ -1546,7 +1663,7 @@ static unsigned int openvswitch_poll(struct file *file, poll_table *wait) if (dp) { mask = 0; poll_wait(file, &dp->waitqueue, wait); - if (dp_has_packet_of_interest(dp, (int)file->private_data)) + if (dp_has_packet_of_interest(dp, get_listen_mask(file))) mask |= POLLIN | POLLRDNORM; } else { mask = POLLIN | POLLRDNORM | POLLHUP; @@ -1563,6 +1680,8 @@ 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, @@ -1575,12 +1694,8 @@ 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 @@ -1591,6 +1706,43 @@ 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) @@ -1625,7 +1777,7 @@ 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);