tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
RDMA/core: Add LAG functionality
Add support to get the RoCE LAG xmit slave by building skb of the RoCE
packet and call to master_get_xmit_slave. If driver wants to get the
slave assume all slaves are available, then need to set
RDMA_LAG_FLAGS_HASH_ALL_SLAVES in flags.
Link: https://lore.kernel.org/r/20200430192146.12863-14-maorg@mellanox.com
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
authored by
Maor Gottlieb
and committed by
Jason Gunthorpe
bd3920ea
fa5d010c
+161
-1
+1
-1
drivers/infiniband/core/Makefile
+1
-1
drivers/infiniband/core/Makefile
···
12
12
roce_gid_mgmt.o mr_pool.o addr.o sa_query.o \
13
13
multicast.o mad.o smi.o agent.o mad_rmpp.o \
14
14
nldev.o restrack.o counters.o ib_core_uverbs.o \
15
-
trace.o
15
+
trace.o lag.o
16
16
17
17
ib_core-$(CONFIG_SECURITY_INFINIBAND) += security.o
18
18
ib_core-$(CONFIG_CGROUP_RDMA) += cgroup.o
+136
drivers/infiniband/core/lag.c
+136
drivers/infiniband/core/lag.c
···
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 = htons(0xC000);
38
+
uh->dest = htons(ROCE_V2_UDP_DPORT);
39
+
uh->len = htons(sizeof(struct udphdr));
40
+
41
+
if (is_ipv4) {
42
+
skb_push(skb, sizeof(struct iphdr));
43
+
skb_reset_network_header(skb);
44
+
iph = ip_hdr(skb);
45
+
iph->frag_off = 0;
46
+
iph->version = 4;
47
+
iph->protocol = IPPROTO_UDP;
48
+
iph->ihl = 0x5;
49
+
iph->tot_len = htons(sizeof(struct udphdr) + sizeof(struct
50
+
iphdr));
51
+
memcpy(&iph->saddr, ah_attr->grh.sgid_attr->gid.raw + 12,
52
+
sizeof(struct in_addr));
53
+
memcpy(&iph->daddr, ah_attr->grh.dgid.raw + 12,
54
+
sizeof(struct in_addr));
55
+
} else {
56
+
skb_push(skb, sizeof(struct ipv6hdr));
57
+
skb_reset_network_header(skb);
58
+
ip6h = ipv6_hdr(skb);
59
+
ip6h->version = 6;
60
+
ip6h->nexthdr = IPPROTO_UDP;
61
+
memcpy(&ip6h->flow_lbl, &ah_attr->grh.flow_label,
62
+
sizeof(*ip6h->flow_lbl));
63
+
memcpy(&ip6h->saddr, ah_attr->grh.sgid_attr->gid.raw,
64
+
sizeof(struct in6_addr));
65
+
memcpy(&ip6h->daddr, ah_attr->grh.dgid.raw,
66
+
sizeof(struct in6_addr));
67
+
}
68
+
69
+
skb_push(skb, sizeof(struct ethhdr));
70
+
skb_reset_mac_header(skb);
71
+
eth = eth_hdr(skb);
72
+
skb->protocol = eth->h_proto = htons(is_ipv4 ? ETH_P_IP : ETH_P_IPV6);
73
+
rdma_read_gid_l2_fields(ah_attr->grh.sgid_attr, NULL, smac);
74
+
memcpy(eth->h_source, smac, ETH_ALEN);
75
+
memcpy(eth->h_dest, ah_attr->roce.dmac, ETH_ALEN);
76
+
77
+
return skb;
78
+
}
79
+
80
+
static struct net_device *rdma_get_xmit_slave_udp(struct ib_device *device,
81
+
struct net_device *master,
82
+
struct rdma_ah_attr *ah_attr,
83
+
gfp_t flags)
84
+
{
85
+
struct net_device *slave;
86
+
struct sk_buff *skb;
87
+
88
+
skb = rdma_build_skb(device, master, ah_attr, flags);
89
+
if (!skb)
90
+
return ERR_PTR(-ENOMEM);
91
+
92
+
rcu_read_lock();
93
+
slave = netdev_get_xmit_slave(master, skb,
94
+
!!(device->lag_flags &
95
+
RDMA_LAG_FLAGS_HASH_ALL_SLAVES));
96
+
if (slave)
97
+
dev_hold(slave);
98
+
rcu_read_unlock();
99
+
kfree_skb(skb);
100
+
return slave;
101
+
}
102
+
103
+
void rdma_lag_put_ah_roce_slave(struct net_device *xmit_slave)
104
+
{
105
+
if (xmit_slave)
106
+
dev_put(xmit_slave);
107
+
}
108
+
109
+
struct net_device *rdma_lag_get_ah_roce_slave(struct ib_device *device,
110
+
struct rdma_ah_attr *ah_attr,
111
+
gfp_t flags)
112
+
{
113
+
struct net_device *slave = NULL;
114
+
struct net_device *master;
115
+
116
+
if (!(ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE &&
117
+
ah_attr->grh.sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP))
118
+
return NULL;
119
+
120
+
rcu_read_lock();
121
+
master = rdma_read_gid_attr_ndev_rcu(ah_attr->grh.sgid_attr);
122
+
if (IS_ERR(master)) {
123
+
rcu_read_unlock();
124
+
return master;
125
+
}
126
+
dev_hold(master);
127
+
rcu_read_unlock();
128
+
129
+
if (!netif_is_bond_master(master))
130
+
goto put;
131
+
132
+
slave = rdma_get_xmit_slave_udp(device, master, ah_attr, flags);
133
+
put:
134
+
dev_put(master);
135
+
return slave;
136
+
}
+1
include/rdma/ib_verbs.h
+1
include/rdma/ib_verbs.h
+23
include/rdma/lag.h
+23
include/rdma/lag.h
···
1
+
/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
2
+
/*
3
+
* Copyright (c) 2020 Mellanox Technologies. All rights reserved.
4
+
*/
5
+
6
+
#ifndef _RDMA_LAG_H_
7
+
#define _RDMA_LAG_H_
8
+
9
+
#include <net/lag.h>
10
+
11
+
struct ib_device;
12
+
struct rdma_ah_attr;
13
+
14
+
enum rdma_lag_flags {
15
+
RDMA_LAG_FLAGS_HASH_ALL_SLAVES = 1 << 0
16
+
};
17
+
18
+
void rdma_lag_put_ah_roce_slave(struct net_device *xmit_slave);
19
+
struct net_device *rdma_lag_get_ah_roce_slave(struct ib_device *device,
20
+
struct rdma_ah_attr *ah_attr,
21
+
gfp_t flags);
22
+
23
+
#endif /* _RDMA_LAG_H_ */