1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
7 #include <linux/skbuff.h>
8 #include <linux/if_arp.h>
9 #include <linux/netdevice.h>
11 #include <linux/if_vlan.h>
12 #include <net/udp_tunnel.h>
13 #include <net/sch_generic.h>
14 #include <linux/netfilter.h>
15 #include <rdma/ib_addr.h>
21 static struct rxe_recv_sockets recv_sockets;
23 static struct dst_entry *rxe_find_route4(struct rxe_qp *qp,
24 struct net_device *ndev,
25 struct in_addr *saddr,
26 struct in_addr *daddr)
29 struct flowi4 fl = { { 0 } };
31 memset(&fl, 0, sizeof(fl));
32 fl.flowi4_oif = ndev->ifindex;
33 memcpy(&fl.saddr, saddr, sizeof(*saddr));
34 memcpy(&fl.daddr, daddr, sizeof(*daddr));
35 fl.flowi4_proto = IPPROTO_UDP;
37 rt = ip_route_output_key(&init_net, &fl);
39 rxe_dbg_qp(qp, "no route to %pI4\n", &daddr->s_addr);
46 #if IS_ENABLED(CONFIG_IPV6)
47 static struct dst_entry *rxe_find_route6(struct rxe_qp *qp,
48 struct net_device *ndev,
49 struct in6_addr *saddr,
50 struct in6_addr *daddr)
52 struct dst_entry *ndst;
53 struct flowi6 fl6 = { { 0 } };
55 memset(&fl6, 0, sizeof(fl6));
56 fl6.flowi6_oif = ndev->ifindex;
57 memcpy(&fl6.saddr, saddr, sizeof(*saddr));
58 memcpy(&fl6.daddr, daddr, sizeof(*daddr));
59 fl6.flowi6_proto = IPPROTO_UDP;
61 ndst = ipv6_stub->ipv6_dst_lookup_flow(sock_net(recv_sockets.sk6->sk),
62 recv_sockets.sk6->sk, &fl6,
65 rxe_dbg_qp(qp, "no route to %pI6\n", daddr);
69 if (unlikely(ndst->error)) {
70 rxe_dbg_qp(qp, "no route to %pI6\n", daddr);
82 static struct dst_entry *rxe_find_route6(struct rxe_qp *qp,
83 struct net_device *ndev,
84 struct in6_addr *saddr,
85 struct in6_addr *daddr)
92 static struct dst_entry *rxe_find_route(struct net_device *ndev,
96 struct dst_entry *dst = NULL;
98 if (qp_type(qp) == IB_QPT_RC)
99 dst = sk_dst_get(qp->sk->sk);
101 if (!dst || !dst_check(dst, qp->dst_cookie)) {
105 if (av->network_type == RXE_NETWORK_TYPE_IPV4) {
106 struct in_addr *saddr;
107 struct in_addr *daddr;
109 saddr = &av->sgid_addr._sockaddr_in.sin_addr;
110 daddr = &av->dgid_addr._sockaddr_in.sin_addr;
111 dst = rxe_find_route4(qp, ndev, saddr, daddr);
112 } else if (av->network_type == RXE_NETWORK_TYPE_IPV6) {
113 struct in6_addr *saddr6;
114 struct in6_addr *daddr6;
116 saddr6 = &av->sgid_addr._sockaddr_in6.sin6_addr;
117 daddr6 = &av->dgid_addr._sockaddr_in6.sin6_addr;
118 dst = rxe_find_route6(qp, ndev, saddr6, daddr6);
119 #if IS_ENABLED(CONFIG_IPV6)
122 rt6_get_cookie((struct rt6_info *)dst);
126 if (dst && (qp_type(qp) == IB_QPT_RC)) {
128 sk_dst_set(qp->sk->sk, dst);
134 static int rxe_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
138 struct net_device *ndev = skb->dev;
139 struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
141 /* takes a reference on rxe->ib_dev
142 * drop when skb is freed
144 rxe = rxe_get_dev_from_net(ndev);
145 if (!rxe && is_vlan_dev(ndev))
146 rxe = rxe_get_dev_from_net(vlan_dev_real_dev(ndev));
150 if (skb_linearize(skb)) {
151 ib_device_put(&rxe->ib_dev);
158 pkt->hdr = (u8 *)(udph + 1);
159 pkt->mask = RXE_GRH_MASK;
160 pkt->paylen = be16_to_cpu(udph->len) - sizeof(*udph);
162 /* remove udp header */
163 skb_pull(skb, sizeof(struct udphdr));
174 static struct socket *rxe_setup_udp_tunnel(struct net *net, __be16 port,
179 struct udp_port_cfg udp_cfg = { };
180 struct udp_tunnel_sock_cfg tnl_cfg = { };
183 udp_cfg.family = AF_INET6;
184 udp_cfg.ipv6_v6only = 1;
186 udp_cfg.family = AF_INET;
189 udp_cfg.local_udp_port = port;
191 /* Create UDP socket */
192 err = udp_sock_create(net, &udp_cfg, &sock);
196 tnl_cfg.encap_type = 1;
197 tnl_cfg.encap_rcv = rxe_udp_encap_recv;
199 /* Setup UDP tunnel */
200 setup_udp_tunnel_sock(net, sock, &tnl_cfg);
205 static void rxe_release_udp_tunnel(struct socket *sk)
208 udp_tunnel_sock_release(sk);
211 static void prepare_udp_hdr(struct sk_buff *skb, __be16 src_port,
216 __skb_push(skb, sizeof(*udph));
217 skb_reset_transport_header(skb);
220 udph->dest = dst_port;
221 udph->source = src_port;
222 udph->len = htons(skb->len);
226 static void prepare_ipv4_hdr(struct dst_entry *dst, struct sk_buff *skb,
227 __be32 saddr, __be32 daddr, __u8 proto,
228 __u8 tos, __u8 ttl, __be16 df, bool xnet)
232 skb_scrub_packet(skb, xnet);
235 skb_dst_set(skb, dst_clone(dst));
236 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
238 skb_push(skb, sizeof(struct iphdr));
239 skb_reset_network_header(skb);
243 iph->version = IPVERSION;
244 iph->ihl = sizeof(struct iphdr) >> 2;
245 iph->tot_len = htons(skb->len);
247 iph->protocol = proto;
252 __ip_select_ident(dev_net(dst->dev), iph,
253 skb_shinfo(skb)->gso_segs ?: 1);
256 static void prepare_ipv6_hdr(struct dst_entry *dst, struct sk_buff *skb,
257 struct in6_addr *saddr, struct in6_addr *daddr,
258 __u8 proto, __u8 prio, __u8 ttl)
260 struct ipv6hdr *ip6h;
262 memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
263 IPCB(skb)->flags &= ~(IPSKB_XFRM_TUNNEL_SIZE | IPSKB_XFRM_TRANSFORMED
265 skb_dst_set(skb, dst_clone(dst));
267 __skb_push(skb, sizeof(*ip6h));
268 skb_reset_network_header(skb);
269 ip6h = ipv6_hdr(skb);
270 ip6_flow_hdr(ip6h, prio, htonl(0));
271 ip6h->payload_len = htons(skb->len);
272 ip6h->nexthdr = proto;
273 ip6h->hop_limit = ttl;
274 ip6h->daddr = *daddr;
275 ip6h->saddr = *saddr;
276 ip6h->payload_len = htons(skb->len - sizeof(*ip6h));
279 static int prepare4(struct rxe_av *av, struct rxe_pkt_info *pkt,
282 struct rxe_qp *qp = pkt->qp;
283 struct dst_entry *dst;
285 __be16 df = htons(IP_DF);
286 struct in_addr *saddr = &av->sgid_addr._sockaddr_in.sin_addr;
287 struct in_addr *daddr = &av->dgid_addr._sockaddr_in.sin_addr;
289 dst = rxe_find_route(skb->dev, qp, av);
291 rxe_dbg_qp(qp, "Host not reachable\n");
292 return -EHOSTUNREACH;
295 prepare_udp_hdr(skb, cpu_to_be16(qp->src_port),
296 cpu_to_be16(ROCE_V2_UDP_DPORT));
298 prepare_ipv4_hdr(dst, skb, saddr->s_addr, daddr->s_addr, IPPROTO_UDP,
299 av->grh.traffic_class, av->grh.hop_limit, df, xnet);
305 static int prepare6(struct rxe_av *av, struct rxe_pkt_info *pkt,
308 struct rxe_qp *qp = pkt->qp;
309 struct dst_entry *dst;
310 struct in6_addr *saddr = &av->sgid_addr._sockaddr_in6.sin6_addr;
311 struct in6_addr *daddr = &av->dgid_addr._sockaddr_in6.sin6_addr;
313 dst = rxe_find_route(skb->dev, qp, av);
315 rxe_dbg_qp(qp, "Host not reachable\n");
316 return -EHOSTUNREACH;
319 prepare_udp_hdr(skb, cpu_to_be16(qp->src_port),
320 cpu_to_be16(ROCE_V2_UDP_DPORT));
322 prepare_ipv6_hdr(dst, skb, saddr, daddr, IPPROTO_UDP,
323 av->grh.traffic_class,
330 int rxe_prepare(struct rxe_av *av, struct rxe_pkt_info *pkt,
335 if (skb->protocol == htons(ETH_P_IP))
336 err = prepare4(av, pkt, skb);
337 else if (skb->protocol == htons(ETH_P_IPV6))
338 err = prepare6(av, pkt, skb);
340 if (ether_addr_equal(skb->dev->dev_addr, av->dmac))
341 pkt->mask |= RXE_LOOPBACK_MASK;
346 static void rxe_skb_tx_dtor(struct sk_buff *skb)
348 struct sock *sk = skb->sk;
349 struct rxe_qp *qp = sk->sk_user_data;
350 int skb_out = atomic_dec_return(&qp->skb_out);
352 if (unlikely(qp->need_req_skb &&
353 skb_out < RXE_INFLIGHT_SKBS_PER_QP_LOW))
354 rxe_sched_task(&qp->req.task);
359 static int rxe_send(struct sk_buff *skb, struct rxe_pkt_info *pkt)
363 skb->destructor = rxe_skb_tx_dtor;
364 skb->sk = pkt->qp->sk->sk;
367 atomic_inc(&pkt->qp->skb_out);
369 if (skb->protocol == htons(ETH_P_IP)) {
370 err = ip_local_out(dev_net(skb_dst(skb)->dev), skb->sk, skb);
371 } else if (skb->protocol == htons(ETH_P_IPV6)) {
372 err = ip6_local_out(dev_net(skb_dst(skb)->dev), skb->sk, skb);
374 rxe_dbg_qp(pkt->qp, "Unknown layer 3 protocol: %d\n",
376 atomic_dec(&pkt->qp->skb_out);
382 if (unlikely(net_xmit_eval(err))) {
383 rxe_dbg_qp(pkt->qp, "error sending packet: %d\n", err);
390 /* fix up a send packet to match the packets
391 * received from UDP before looping them back
393 static int rxe_loopback(struct sk_buff *skb, struct rxe_pkt_info *pkt)
395 memcpy(SKB_TO_PKT(skb), pkt, sizeof(*pkt));
397 if (skb->protocol == htons(ETH_P_IP))
398 skb_pull(skb, sizeof(struct iphdr));
400 skb_pull(skb, sizeof(struct ipv6hdr));
402 if (WARN_ON(!ib_device_try_get(&pkt->rxe->ib_dev))) {
407 /* remove udp header */
408 skb_pull(skb, sizeof(struct udphdr));
415 int rxe_xmit_packet(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
419 int is_request = pkt->mask & RXE_REQ_MASK;
420 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
423 spin_lock_irqsave(&qp->state_lock, flags);
424 if ((is_request && (qp_state(qp) < IB_QPS_RTS)) ||
425 (!is_request && (qp_state(qp) < IB_QPS_RTR))) {
426 spin_unlock_irqrestore(&qp->state_lock, flags);
427 rxe_dbg_qp(qp, "Packet dropped. QP is not in ready state\n");
430 spin_unlock_irqrestore(&qp->state_lock, flags);
432 rxe_icrc_generate(skb, pkt);
434 if (pkt->mask & RXE_LOOPBACK_MASK)
435 err = rxe_loopback(skb, pkt);
437 err = rxe_send(skb, pkt);
439 rxe_counter_inc(rxe, RXE_CNT_SEND_ERR);
443 if ((qp_type(qp) != IB_QPT_RC) &&
444 (pkt->mask & RXE_END_MASK)) {
445 pkt->wqe->state = wqe_state_done;
446 rxe_sched_task(&qp->comp.task);
449 rxe_counter_inc(rxe, RXE_CNT_SENT_PKTS);
459 struct sk_buff *rxe_init_packet(struct rxe_dev *rxe, struct rxe_av *av,
460 int paylen, struct rxe_pkt_info *pkt)
462 unsigned int hdr_len;
463 struct sk_buff *skb = NULL;
464 struct net_device *ndev;
465 const struct ib_gid_attr *attr;
466 const int port_num = 1;
468 attr = rdma_get_gid_attr(&rxe->ib_dev, port_num, av->grh.sgid_index);
472 if (av->network_type == RXE_NETWORK_TYPE_IPV4)
473 hdr_len = ETH_HLEN + sizeof(struct udphdr) +
474 sizeof(struct iphdr);
476 hdr_len = ETH_HLEN + sizeof(struct udphdr) +
477 sizeof(struct ipv6hdr);
480 ndev = rdma_read_gid_attr_ndev_rcu(attr);
485 skb = alloc_skb(paylen + hdr_len + LL_RESERVED_SPACE(ndev),
488 if (unlikely(!skb)) {
493 skb_reserve(skb, hdr_len + LL_RESERVED_SPACE(ndev));
495 /* FIXME: hold reference to this netdev until life of this skb. */
499 if (av->network_type == RXE_NETWORK_TYPE_IPV4)
500 skb->protocol = htons(ETH_P_IP);
502 skb->protocol = htons(ETH_P_IPV6);
505 pkt->port_num = port_num;
506 pkt->hdr = skb_put(skb, paylen);
507 pkt->mask |= RXE_GRH_MASK;
510 rdma_put_gid_attr(attr);
515 * this is required by rxe_cfg to match rxe devices in
516 * /sys/class/infiniband up with their underlying ethernet devices
518 const char *rxe_parent_name(struct rxe_dev *rxe, unsigned int port_num)
520 return rxe->ndev->name;
523 int rxe_net_add(const char *ibdev_name, struct net_device *ndev)
526 struct rxe_dev *rxe = NULL;
528 rxe = ib_alloc_device(rxe_dev, ib_dev);
534 err = rxe_add(rxe, ndev->mtu, ibdev_name);
536 ib_dealloc_device(&rxe->ib_dev);
543 static void rxe_port_event(struct rxe_dev *rxe,
544 enum ib_event_type event)
548 ev.device = &rxe->ib_dev;
549 ev.element.port_num = 1;
552 ib_dispatch_event(&ev);
555 /* Caller must hold net_info_lock */
556 void rxe_port_up(struct rxe_dev *rxe)
558 struct rxe_port *port;
561 port->attr.state = IB_PORT_ACTIVE;
563 rxe_port_event(rxe, IB_EVENT_PORT_ACTIVE);
564 dev_info(&rxe->ib_dev.dev, "set active\n");
567 /* Caller must hold net_info_lock */
568 void rxe_port_down(struct rxe_dev *rxe)
570 struct rxe_port *port;
573 port->attr.state = IB_PORT_DOWN;
575 rxe_port_event(rxe, IB_EVENT_PORT_ERR);
576 rxe_counter_inc(rxe, RXE_CNT_LINK_DOWNED);
577 dev_info(&rxe->ib_dev.dev, "set down\n");
580 void rxe_set_port_state(struct rxe_dev *rxe)
582 if (netif_running(rxe->ndev) && netif_carrier_ok(rxe->ndev))
588 static int rxe_notify(struct notifier_block *not_blk,
592 struct net_device *ndev = netdev_notifier_info_to_dev(arg);
593 struct rxe_dev *rxe = rxe_get_dev_from_net(ndev);
599 case NETDEV_UNREGISTER:
600 ib_unregister_device_queued(&rxe->ib_dev);
608 case NETDEV_CHANGEMTU:
609 rxe_dbg_dev(rxe, "%s changed mtu to %d\n", ndev->name, ndev->mtu);
610 rxe_set_mtu(rxe, ndev->mtu);
613 rxe_set_port_state(rxe);
616 case NETDEV_GOING_DOWN:
617 case NETDEV_CHANGEADDR:
618 case NETDEV_CHANGENAME:
619 case NETDEV_FEAT_CHANGE:
621 rxe_dbg_dev(rxe, "ignoring netdev event = %ld for %s\n",
626 ib_device_put(&rxe->ib_dev);
630 static struct notifier_block rxe_net_notifier = {
631 .notifier_call = rxe_notify,
634 static int rxe_net_ipv4_init(void)
636 recv_sockets.sk4 = rxe_setup_udp_tunnel(&init_net,
637 htons(ROCE_V2_UDP_DPORT), false);
638 if (IS_ERR(recv_sockets.sk4)) {
639 recv_sockets.sk4 = NULL;
640 pr_err("Failed to create IPv4 UDP tunnel\n");
647 static int rxe_net_ipv6_init(void)
649 #if IS_ENABLED(CONFIG_IPV6)
651 recv_sockets.sk6 = rxe_setup_udp_tunnel(&init_net,
652 htons(ROCE_V2_UDP_DPORT), true);
653 if (PTR_ERR(recv_sockets.sk6) == -EAFNOSUPPORT) {
654 recv_sockets.sk6 = NULL;
655 pr_warn("IPv6 is not supported, can not create a UDPv6 socket\n");
659 if (IS_ERR(recv_sockets.sk6)) {
660 recv_sockets.sk6 = NULL;
661 pr_err("Failed to create IPv6 UDP tunnel\n");
668 void rxe_net_exit(void)
670 rxe_release_udp_tunnel(recv_sockets.sk6);
671 rxe_release_udp_tunnel(recv_sockets.sk4);
672 unregister_netdevice_notifier(&rxe_net_notifier);
675 int rxe_net_init(void)
679 recv_sockets.sk6 = NULL;
681 err = rxe_net_ipv4_init();
684 err = rxe_net_ipv6_init();
687 err = register_netdevice_notifier(&rxe_net_notifier);
689 pr_err("Failed to register netdev notifier\n");