Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2/*
3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5 */
6
7#include <linux/skbuff.h>
8#include <linux/if_arp.h>
9#include <linux/netdevice.h>
10#include <linux/if.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>
16
17#include "rxe.h"
18#include "rxe_net.h"
19#include "rxe_loc.h"
20
21static struct rxe_recv_sockets recv_sockets;
22
23static 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)
27{
28 struct rtable *rt;
29 struct flowi4 fl = { { 0 } };
30
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;
36
37 rt = ip_route_output_key(&init_net, &fl);
38 if (IS_ERR(rt)) {
39 rxe_dbg_qp(qp, "no route to %pI4\n", &daddr->s_addr);
40 return NULL;
41 }
42
43 return &rt->dst;
44}
45
46#if IS_ENABLED(CONFIG_IPV6)
47static 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)
51{
52 struct dst_entry *ndst;
53 struct flowi6 fl6 = { { 0 } };
54
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;
60
61 ndst = ipv6_stub->ipv6_dst_lookup_flow(sock_net(recv_sockets.sk6->sk),
62 recv_sockets.sk6->sk, &fl6,
63 NULL);
64 if (IS_ERR(ndst)) {
65 rxe_dbg_qp(qp, "no route to %pI6\n", daddr);
66 return NULL;
67 }
68
69 if (unlikely(ndst->error)) {
70 rxe_dbg_qp(qp, "no route to %pI6\n", daddr);
71 goto put;
72 }
73
74 return ndst;
75put:
76 dst_release(ndst);
77 return NULL;
78}
79
80#else
81
82static 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)
86{
87 return NULL;
88}
89
90#endif
91
92static struct dst_entry *rxe_find_route(struct net_device *ndev,
93 struct rxe_qp *qp,
94 struct rxe_av *av)
95{
96 struct dst_entry *dst = NULL;
97
98 if (qp_type(qp) == IB_QPT_RC)
99 dst = sk_dst_get(qp->sk->sk);
100
101 if (!dst || !dst_check(dst, qp->dst_cookie)) {
102 if (dst)
103 dst_release(dst);
104
105 if (av->network_type == RXE_NETWORK_TYPE_IPV4) {
106 struct in_addr *saddr;
107 struct in_addr *daddr;
108
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;
115
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)
120 if (dst)
121 qp->dst_cookie =
122 rt6_get_cookie((struct rt6_info *)dst);
123#endif
124 }
125
126 if (dst && (qp_type(qp) == IB_QPT_RC)) {
127 dst_hold(dst);
128 sk_dst_set(qp->sk->sk, dst);
129 }
130 }
131 return dst;
132}
133
134static int rxe_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
135{
136 struct udphdr *udph;
137 struct rxe_dev *rxe;
138 struct net_device *ndev = skb->dev;
139 struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
140
141 /* takes a reference on rxe->ib_dev
142 * drop when skb is freed
143 */
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));
147 if (!rxe)
148 goto drop;
149
150 if (skb_linearize(skb)) {
151 ib_device_put(&rxe->ib_dev);
152 goto drop;
153 }
154
155 udph = udp_hdr(skb);
156 pkt->rxe = rxe;
157 pkt->port_num = 1;
158 pkt->hdr = (u8 *)(udph + 1);
159 pkt->mask = RXE_GRH_MASK;
160 pkt->paylen = be16_to_cpu(udph->len) - sizeof(*udph);
161
162 rxe_rcv(skb);
163
164 return 0;
165drop:
166 kfree_skb(skb);
167
168 return 0;
169}
170
171static struct socket *rxe_setup_udp_tunnel(struct net *net, __be16 port,
172 bool ipv6)
173{
174 int err;
175 struct socket *sock;
176 struct udp_port_cfg udp_cfg = { };
177 struct udp_tunnel_sock_cfg tnl_cfg = { };
178
179 if (ipv6) {
180 udp_cfg.family = AF_INET6;
181 udp_cfg.ipv6_v6only = 1;
182 } else {
183 udp_cfg.family = AF_INET;
184 }
185
186 udp_cfg.local_udp_port = port;
187
188 /* Create UDP socket */
189 err = udp_sock_create(net, &udp_cfg, &sock);
190 if (err < 0)
191 return ERR_PTR(err);
192
193 tnl_cfg.encap_type = 1;
194 tnl_cfg.encap_rcv = rxe_udp_encap_recv;
195
196 /* Setup UDP tunnel */
197 setup_udp_tunnel_sock(net, sock, &tnl_cfg);
198
199 return sock;
200}
201
202static void rxe_release_udp_tunnel(struct socket *sk)
203{
204 if (sk)
205 udp_tunnel_sock_release(sk);
206}
207
208static void prepare_udp_hdr(struct sk_buff *skb, __be16 src_port,
209 __be16 dst_port)
210{
211 struct udphdr *udph;
212
213 __skb_push(skb, sizeof(*udph));
214 skb_reset_transport_header(skb);
215 udph = udp_hdr(skb);
216
217 udph->dest = dst_port;
218 udph->source = src_port;
219 udph->len = htons(skb->len);
220 udph->check = 0;
221}
222
223static void prepare_ipv4_hdr(struct dst_entry *dst, struct sk_buff *skb,
224 __be32 saddr, __be32 daddr, __u8 proto,
225 __u8 tos, __u8 ttl, __be16 df, bool xnet)
226{
227 struct iphdr *iph;
228
229 skb_scrub_packet(skb, xnet);
230
231 skb_clear_hash(skb);
232 skb_dst_set(skb, dst_clone(dst));
233 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
234
235 skb_push(skb, sizeof(struct iphdr));
236 skb_reset_network_header(skb);
237
238 iph = ip_hdr(skb);
239
240 iph->version = IPVERSION;
241 iph->ihl = sizeof(struct iphdr) >> 2;
242 iph->tot_len = htons(skb->len);
243 iph->frag_off = df;
244 iph->protocol = proto;
245 iph->tos = tos;
246 iph->daddr = daddr;
247 iph->saddr = saddr;
248 iph->ttl = ttl;
249 __ip_select_ident(dev_net(dst->dev), iph,
250 skb_shinfo(skb)->gso_segs ?: 1);
251}
252
253static void prepare_ipv6_hdr(struct dst_entry *dst, struct sk_buff *skb,
254 struct in6_addr *saddr, struct in6_addr *daddr,
255 __u8 proto, __u8 prio, __u8 ttl)
256{
257 struct ipv6hdr *ip6h;
258
259 memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
260 IPCB(skb)->flags &= ~(IPSKB_XFRM_TUNNEL_SIZE | IPSKB_XFRM_TRANSFORMED
261 | IPSKB_REROUTED);
262 skb_dst_set(skb, dst_clone(dst));
263
264 __skb_push(skb, sizeof(*ip6h));
265 skb_reset_network_header(skb);
266 ip6h = ipv6_hdr(skb);
267 ip6_flow_hdr(ip6h, prio, htonl(0));
268 ip6h->payload_len = htons(skb->len);
269 ip6h->nexthdr = proto;
270 ip6h->hop_limit = ttl;
271 ip6h->daddr = *daddr;
272 ip6h->saddr = *saddr;
273 ip6h->payload_len = htons(skb->len - sizeof(*ip6h));
274}
275
276static int prepare4(struct rxe_av *av, struct rxe_pkt_info *pkt,
277 struct sk_buff *skb)
278{
279 struct rxe_qp *qp = pkt->qp;
280 struct dst_entry *dst;
281 bool xnet = false;
282 __be16 df = htons(IP_DF);
283 struct in_addr *saddr = &av->sgid_addr._sockaddr_in.sin_addr;
284 struct in_addr *daddr = &av->dgid_addr._sockaddr_in.sin_addr;
285
286 dst = rxe_find_route(skb->dev, qp, av);
287 if (!dst) {
288 rxe_dbg_qp(qp, "Host not reachable\n");
289 return -EHOSTUNREACH;
290 }
291
292 prepare_udp_hdr(skb, cpu_to_be16(qp->src_port),
293 cpu_to_be16(ROCE_V2_UDP_DPORT));
294
295 prepare_ipv4_hdr(dst, skb, saddr->s_addr, daddr->s_addr, IPPROTO_UDP,
296 av->grh.traffic_class, av->grh.hop_limit, df, xnet);
297
298 dst_release(dst);
299 return 0;
300}
301
302static int prepare6(struct rxe_av *av, struct rxe_pkt_info *pkt,
303 struct sk_buff *skb)
304{
305 struct rxe_qp *qp = pkt->qp;
306 struct dst_entry *dst;
307 struct in6_addr *saddr = &av->sgid_addr._sockaddr_in6.sin6_addr;
308 struct in6_addr *daddr = &av->dgid_addr._sockaddr_in6.sin6_addr;
309
310 dst = rxe_find_route(skb->dev, qp, av);
311 if (!dst) {
312 rxe_dbg_qp(qp, "Host not reachable\n");
313 return -EHOSTUNREACH;
314 }
315
316 prepare_udp_hdr(skb, cpu_to_be16(qp->src_port),
317 cpu_to_be16(ROCE_V2_UDP_DPORT));
318
319 prepare_ipv6_hdr(dst, skb, saddr, daddr, IPPROTO_UDP,
320 av->grh.traffic_class,
321 av->grh.hop_limit);
322
323 dst_release(dst);
324 return 0;
325}
326
327int rxe_prepare(struct rxe_av *av, struct rxe_pkt_info *pkt,
328 struct sk_buff *skb)
329{
330 int err = 0;
331
332 if (skb->protocol == htons(ETH_P_IP))
333 err = prepare4(av, pkt, skb);
334 else if (skb->protocol == htons(ETH_P_IPV6))
335 err = prepare6(av, pkt, skb);
336
337 if (ether_addr_equal(skb->dev->dev_addr, av->dmac))
338 pkt->mask |= RXE_LOOPBACK_MASK;
339
340 return err;
341}
342
343static void rxe_skb_tx_dtor(struct sk_buff *skb)
344{
345 struct sock *sk = skb->sk;
346 struct rxe_qp *qp = sk->sk_user_data;
347 int skb_out = atomic_dec_return(&qp->skb_out);
348
349 if (unlikely(qp->need_req_skb &&
350 skb_out < RXE_INFLIGHT_SKBS_PER_QP_LOW))
351 rxe_sched_task(&qp->req.task);
352
353 rxe_put(qp);
354}
355
356static int rxe_send(struct sk_buff *skb, struct rxe_pkt_info *pkt)
357{
358 int err;
359
360 skb->destructor = rxe_skb_tx_dtor;
361 skb->sk = pkt->qp->sk->sk;
362
363 rxe_get(pkt->qp);
364 atomic_inc(&pkt->qp->skb_out);
365
366 if (skb->protocol == htons(ETH_P_IP)) {
367 err = ip_local_out(dev_net(skb_dst(skb)->dev), skb->sk, skb);
368 } else if (skb->protocol == htons(ETH_P_IPV6)) {
369 err = ip6_local_out(dev_net(skb_dst(skb)->dev), skb->sk, skb);
370 } else {
371 rxe_dbg_qp(pkt->qp, "Unknown layer 3 protocol: %d\n",
372 skb->protocol);
373 atomic_dec(&pkt->qp->skb_out);
374 rxe_put(pkt->qp);
375 kfree_skb(skb);
376 return -EINVAL;
377 }
378
379 if (unlikely(net_xmit_eval(err))) {
380 rxe_dbg_qp(pkt->qp, "error sending packet: %d\n", err);
381 return -EAGAIN;
382 }
383
384 return 0;
385}
386
387/* fix up a send packet to match the packets
388 * received from UDP before looping them back
389 */
390static int rxe_loopback(struct sk_buff *skb, struct rxe_pkt_info *pkt)
391{
392 memcpy(SKB_TO_PKT(skb), pkt, sizeof(*pkt));
393
394 if (skb->protocol == htons(ETH_P_IP))
395 skb_pull(skb, sizeof(struct iphdr));
396 else
397 skb_pull(skb, sizeof(struct ipv6hdr));
398
399 if (WARN_ON(!ib_device_try_get(&pkt->rxe->ib_dev))) {
400 kfree_skb(skb);
401 return -EIO;
402 }
403
404 rxe_rcv(skb);
405
406 return 0;
407}
408
409int rxe_xmit_packet(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
410 struct sk_buff *skb)
411{
412 int err;
413 int is_request = pkt->mask & RXE_REQ_MASK;
414 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
415
416 if ((is_request && (qp->req.state != QP_STATE_READY)) ||
417 (!is_request && (qp->resp.state != QP_STATE_READY))) {
418 rxe_dbg_qp(qp, "Packet dropped. QP is not in ready state\n");
419 goto drop;
420 }
421
422 rxe_icrc_generate(skb, pkt);
423
424 if (pkt->mask & RXE_LOOPBACK_MASK)
425 err = rxe_loopback(skb, pkt);
426 else
427 err = rxe_send(skb, pkt);
428 if (err) {
429 rxe_counter_inc(rxe, RXE_CNT_SEND_ERR);
430 return err;
431 }
432
433 if ((qp_type(qp) != IB_QPT_RC) &&
434 (pkt->mask & RXE_END_MASK)) {
435 pkt->wqe->state = wqe_state_done;
436 rxe_sched_task(&qp->comp.task);
437 }
438
439 rxe_counter_inc(rxe, RXE_CNT_SENT_PKTS);
440 goto done;
441
442drop:
443 kfree_skb(skb);
444 err = 0;
445done:
446 return err;
447}
448
449struct sk_buff *rxe_init_packet(struct rxe_dev *rxe, struct rxe_av *av,
450 int paylen, struct rxe_pkt_info *pkt)
451{
452 unsigned int hdr_len;
453 struct sk_buff *skb = NULL;
454 struct net_device *ndev;
455 const struct ib_gid_attr *attr;
456 const int port_num = 1;
457
458 attr = rdma_get_gid_attr(&rxe->ib_dev, port_num, av->grh.sgid_index);
459 if (IS_ERR(attr))
460 return NULL;
461
462 if (av->network_type == RXE_NETWORK_TYPE_IPV4)
463 hdr_len = ETH_HLEN + sizeof(struct udphdr) +
464 sizeof(struct iphdr);
465 else
466 hdr_len = ETH_HLEN + sizeof(struct udphdr) +
467 sizeof(struct ipv6hdr);
468
469 rcu_read_lock();
470 ndev = rdma_read_gid_attr_ndev_rcu(attr);
471 if (IS_ERR(ndev)) {
472 rcu_read_unlock();
473 goto out;
474 }
475 skb = alloc_skb(paylen + hdr_len + LL_RESERVED_SPACE(ndev),
476 GFP_ATOMIC);
477
478 if (unlikely(!skb)) {
479 rcu_read_unlock();
480 goto out;
481 }
482
483 skb_reserve(skb, hdr_len + LL_RESERVED_SPACE(ndev));
484
485 /* FIXME: hold reference to this netdev until life of this skb. */
486 skb->dev = ndev;
487 rcu_read_unlock();
488
489 if (av->network_type == RXE_NETWORK_TYPE_IPV4)
490 skb->protocol = htons(ETH_P_IP);
491 else
492 skb->protocol = htons(ETH_P_IPV6);
493
494 pkt->rxe = rxe;
495 pkt->port_num = port_num;
496 pkt->hdr = skb_put(skb, paylen);
497 pkt->mask |= RXE_GRH_MASK;
498
499out:
500 rdma_put_gid_attr(attr);
501 return skb;
502}
503
504/*
505 * this is required by rxe_cfg to match rxe devices in
506 * /sys/class/infiniband up with their underlying ethernet devices
507 */
508const char *rxe_parent_name(struct rxe_dev *rxe, unsigned int port_num)
509{
510 return rxe->ndev->name;
511}
512
513int rxe_net_add(const char *ibdev_name, struct net_device *ndev)
514{
515 int err;
516 struct rxe_dev *rxe = NULL;
517
518 rxe = ib_alloc_device(rxe_dev, ib_dev);
519 if (!rxe)
520 return -ENOMEM;
521
522 rxe->ndev = ndev;
523
524 err = rxe_add(rxe, ndev->mtu, ibdev_name);
525 if (err) {
526 ib_dealloc_device(&rxe->ib_dev);
527 return err;
528 }
529
530 return 0;
531}
532
533static void rxe_port_event(struct rxe_dev *rxe,
534 enum ib_event_type event)
535{
536 struct ib_event ev;
537
538 ev.device = &rxe->ib_dev;
539 ev.element.port_num = 1;
540 ev.event = event;
541
542 ib_dispatch_event(&ev);
543}
544
545/* Caller must hold net_info_lock */
546void rxe_port_up(struct rxe_dev *rxe)
547{
548 struct rxe_port *port;
549
550 port = &rxe->port;
551 port->attr.state = IB_PORT_ACTIVE;
552
553 rxe_port_event(rxe, IB_EVENT_PORT_ACTIVE);
554 dev_info(&rxe->ib_dev.dev, "set active\n");
555}
556
557/* Caller must hold net_info_lock */
558void rxe_port_down(struct rxe_dev *rxe)
559{
560 struct rxe_port *port;
561
562 port = &rxe->port;
563 port->attr.state = IB_PORT_DOWN;
564
565 rxe_port_event(rxe, IB_EVENT_PORT_ERR);
566 rxe_counter_inc(rxe, RXE_CNT_LINK_DOWNED);
567 dev_info(&rxe->ib_dev.dev, "set down\n");
568}
569
570void rxe_set_port_state(struct rxe_dev *rxe)
571{
572 if (netif_running(rxe->ndev) && netif_carrier_ok(rxe->ndev))
573 rxe_port_up(rxe);
574 else
575 rxe_port_down(rxe);
576}
577
578static int rxe_notify(struct notifier_block *not_blk,
579 unsigned long event,
580 void *arg)
581{
582 struct net_device *ndev = netdev_notifier_info_to_dev(arg);
583 struct rxe_dev *rxe = rxe_get_dev_from_net(ndev);
584
585 if (!rxe)
586 return NOTIFY_OK;
587
588 switch (event) {
589 case NETDEV_UNREGISTER:
590 ib_unregister_device_queued(&rxe->ib_dev);
591 break;
592 case NETDEV_UP:
593 rxe_port_up(rxe);
594 break;
595 case NETDEV_DOWN:
596 rxe_port_down(rxe);
597 break;
598 case NETDEV_CHANGEMTU:
599 rxe_dbg(rxe, "%s changed mtu to %d\n", ndev->name, ndev->mtu);
600 rxe_set_mtu(rxe, ndev->mtu);
601 break;
602 case NETDEV_CHANGE:
603 rxe_set_port_state(rxe);
604 break;
605 case NETDEV_REBOOT:
606 case NETDEV_GOING_DOWN:
607 case NETDEV_CHANGEADDR:
608 case NETDEV_CHANGENAME:
609 case NETDEV_FEAT_CHANGE:
610 default:
611 rxe_dbg(rxe, "ignoring netdev event = %ld for %s\n",
612 event, ndev->name);
613 break;
614 }
615
616 ib_device_put(&rxe->ib_dev);
617 return NOTIFY_OK;
618}
619
620static struct notifier_block rxe_net_notifier = {
621 .notifier_call = rxe_notify,
622};
623
624static int rxe_net_ipv4_init(void)
625{
626 recv_sockets.sk4 = rxe_setup_udp_tunnel(&init_net,
627 htons(ROCE_V2_UDP_DPORT), false);
628 if (IS_ERR(recv_sockets.sk4)) {
629 recv_sockets.sk4 = NULL;
630 pr_err("Failed to create IPv4 UDP tunnel\n");
631 return -1;
632 }
633
634 return 0;
635}
636
637static int rxe_net_ipv6_init(void)
638{
639#if IS_ENABLED(CONFIG_IPV6)
640
641 recv_sockets.sk6 = rxe_setup_udp_tunnel(&init_net,
642 htons(ROCE_V2_UDP_DPORT), true);
643 if (PTR_ERR(recv_sockets.sk6) == -EAFNOSUPPORT) {
644 recv_sockets.sk6 = NULL;
645 pr_warn("IPv6 is not supported, can not create a UDPv6 socket\n");
646 return 0;
647 }
648
649 if (IS_ERR(recv_sockets.sk6)) {
650 recv_sockets.sk6 = NULL;
651 pr_err("Failed to create IPv6 UDP tunnel\n");
652 return -1;
653 }
654#endif
655 return 0;
656}
657
658void rxe_net_exit(void)
659{
660 rxe_release_udp_tunnel(recv_sockets.sk6);
661 rxe_release_udp_tunnel(recv_sockets.sk4);
662 unregister_netdevice_notifier(&rxe_net_notifier);
663}
664
665int rxe_net_init(void)
666{
667 int err;
668
669 recv_sockets.sk6 = NULL;
670
671 err = rxe_net_ipv4_init();
672 if (err)
673 return err;
674 err = rxe_net_ipv6_init();
675 if (err)
676 goto err_out;
677 err = register_netdevice_notifier(&rxe_net_notifier);
678 if (err) {
679 pr_err("Failed to register netdev notifier\n");
680 goto err_out;
681 }
682 return 0;
683err_out:
684 rxe_net_exit();
685 return err;
686}