xref: /openbmc/linux/drivers/infiniband/core/lag.c (revision 9123e3a7)
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2020 Mellanox Technologies. All rights reserved.
4  */
5 
6 #include <rdma/ib_verbs.h>
7 #include <rdma/ib_cache.h>
8 #include <rdma/lag.h>
9 
10 static struct sk_buff *rdma_build_skb(struct ib_device *device,
11 				      struct net_device *netdev,
12 				      struct rdma_ah_attr *ah_attr,
13 				      gfp_t flags)
14 {
15 	struct ipv6hdr *ip6h;
16 	struct sk_buff *skb;
17 	struct ethhdr *eth;
18 	struct iphdr *iph;
19 	struct udphdr *uh;
20 	u8 smac[ETH_ALEN];
21 	bool is_ipv4;
22 	int hdr_len;
23 
24 	is_ipv4 = ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw);
25 	hdr_len = ETH_HLEN + sizeof(struct udphdr) + LL_RESERVED_SPACE(netdev);
26 	hdr_len += is_ipv4 ? sizeof(struct iphdr) : sizeof(struct ipv6hdr);
27 
28 	skb = alloc_skb(hdr_len, flags);
29 	if (!skb)
30 		return NULL;
31 
32 	skb->dev = netdev;
33 	skb_reserve(skb, hdr_len);
34 	skb_push(skb, sizeof(struct udphdr));
35 	skb_reset_transport_header(skb);
36 	uh = udp_hdr(skb);
37 	uh->source =
38 		htons(rdma_flow_label_to_udp_sport(ah_attr->grh.flow_label));
39 	uh->dest = htons(ROCE_V2_UDP_DPORT);
40 	uh->len = htons(sizeof(struct udphdr));
41 
42 	if (is_ipv4) {
43 		skb_push(skb, sizeof(struct iphdr));
44 		skb_reset_network_header(skb);
45 		iph = ip_hdr(skb);
46 		iph->frag_off = 0;
47 		iph->version = 4;
48 		iph->protocol = IPPROTO_UDP;
49 		iph->ihl = 0x5;
50 		iph->tot_len = htons(sizeof(struct udphdr) + sizeof(struct
51 								    iphdr));
52 		memcpy(&iph->saddr, ah_attr->grh.sgid_attr->gid.raw + 12,
53 		       sizeof(struct in_addr));
54 		memcpy(&iph->daddr, ah_attr->grh.dgid.raw + 12,
55 		       sizeof(struct in_addr));
56 	} else {
57 		skb_push(skb, sizeof(struct ipv6hdr));
58 		skb_reset_network_header(skb);
59 		ip6h = ipv6_hdr(skb);
60 		ip6h->version = 6;
61 		ip6h->nexthdr = IPPROTO_UDP;
62 		memcpy(&ip6h->flow_lbl, &ah_attr->grh.flow_label,
63 		       sizeof(*ip6h->flow_lbl));
64 		memcpy(&ip6h->saddr, ah_attr->grh.sgid_attr->gid.raw,
65 		       sizeof(struct in6_addr));
66 		memcpy(&ip6h->daddr, ah_attr->grh.dgid.raw,
67 		       sizeof(struct in6_addr));
68 	}
69 
70 	skb_push(skb, sizeof(struct ethhdr));
71 	skb_reset_mac_header(skb);
72 	eth = eth_hdr(skb);
73 	skb->protocol = eth->h_proto = htons(is_ipv4 ? ETH_P_IP : ETH_P_IPV6);
74 	rdma_read_gid_l2_fields(ah_attr->grh.sgid_attr, NULL, smac);
75 	memcpy(eth->h_source, smac, ETH_ALEN);
76 	memcpy(eth->h_dest, ah_attr->roce.dmac, ETH_ALEN);
77 
78 	return skb;
79 }
80 
81 static struct net_device *rdma_get_xmit_slave_udp(struct ib_device *device,
82 						  struct net_device *master,
83 						  struct rdma_ah_attr *ah_attr,
84 						  gfp_t flags)
85 {
86 	struct net_device *slave;
87 	struct sk_buff *skb;
88 
89 	skb = rdma_build_skb(device, master, ah_attr, flags);
90 	if (!skb)
91 		return ERR_PTR(-ENOMEM);
92 
93 	rcu_read_lock();
94 	slave = netdev_get_xmit_slave(master, skb,
95 				      !!(device->lag_flags &
96 					 RDMA_LAG_FLAGS_HASH_ALL_SLAVES));
97 	if (slave)
98 		dev_hold(slave);
99 	rcu_read_unlock();
100 	kfree_skb(skb);
101 	return slave;
102 }
103 
104 void rdma_lag_put_ah_roce_slave(struct net_device *xmit_slave)
105 {
106 	if (xmit_slave)
107 		dev_put(xmit_slave);
108 }
109 
110 struct net_device *rdma_lag_get_ah_roce_slave(struct ib_device *device,
111 					      struct rdma_ah_attr *ah_attr,
112 					      gfp_t flags)
113 {
114 	struct net_device *slave = NULL;
115 	struct net_device *master;
116 
117 	if (!(ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE &&
118 	      ah_attr->grh.sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP &&
119 	      ah_attr->grh.flow_label))
120 		return NULL;
121 
122 	rcu_read_lock();
123 	master = rdma_read_gid_attr_ndev_rcu(ah_attr->grh.sgid_attr);
124 	if (IS_ERR(master)) {
125 		rcu_read_unlock();
126 		return master;
127 	}
128 	dev_hold(master);
129 	rcu_read_unlock();
130 
131 	if (!netif_is_bond_master(master))
132 		goto put;
133 
134 	slave = rdma_get_xmit_slave_udp(device, master, ah_attr, flags);
135 put:
136 	dev_put(master);
137 	return slave;
138 }
139