1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3 * Copyright (c) 2018 Mellanox Technologies. All rights reserved.
4 */
5
6 #include <linux/mlx5/vport.h>
7 #include "ib_rep.h"
8 #include "srq.h"
9
10 static int
mlx5_ib_set_vport_rep(struct mlx5_core_dev * dev,struct mlx5_eswitch_rep * rep,int vport_index)11 mlx5_ib_set_vport_rep(struct mlx5_core_dev *dev,
12 struct mlx5_eswitch_rep *rep,
13 int vport_index)
14 {
15 struct mlx5_ib_dev *ibdev;
16
17 ibdev = mlx5_eswitch_uplink_get_proto_dev(dev->priv.eswitch, REP_IB);
18 if (!ibdev)
19 return -EINVAL;
20
21 ibdev->port[vport_index].rep = rep;
22 rep->rep_data[REP_IB].priv = ibdev;
23 write_lock(&ibdev->port[vport_index].roce.netdev_lock);
24 ibdev->port[vport_index].roce.netdev =
25 mlx5_ib_get_rep_netdev(rep->esw, rep->vport);
26 write_unlock(&ibdev->port[vport_index].roce.netdev_lock);
27
28 return 0;
29 }
30
31 static void mlx5_ib_register_peer_vport_reps(struct mlx5_core_dev *mdev);
32
mlx5_ib_num_ports_update(struct mlx5_core_dev * dev,u32 * num_ports)33 static void mlx5_ib_num_ports_update(struct mlx5_core_dev *dev, u32 *num_ports)
34 {
35 struct mlx5_core_dev *peer_dev;
36 int i;
37
38 mlx5_lag_for_each_peer_mdev(dev, peer_dev, i) {
39 u32 peer_num_ports = mlx5_eswitch_get_total_vports(peer_dev);
40
41 if (mlx5_lag_is_mpesw(peer_dev))
42 *num_ports += peer_num_ports;
43 else
44 /* Only 1 ib port is the representor for all uplinks */
45 *num_ports += peer_num_ports - 1;
46 }
47 }
48
49 static int
mlx5_ib_vport_rep_load(struct mlx5_core_dev * dev,struct mlx5_eswitch_rep * rep)50 mlx5_ib_vport_rep_load(struct mlx5_core_dev *dev, struct mlx5_eswitch_rep *rep)
51 {
52 u32 num_ports = mlx5_eswitch_get_total_vports(dev);
53 struct mlx5_core_dev *lag_master = dev;
54 const struct mlx5_ib_profile *profile;
55 struct mlx5_core_dev *peer_dev;
56 struct mlx5_ib_dev *ibdev;
57 int new_uplink = false;
58 int vport_index;
59 int ret;
60 int i;
61
62 vport_index = rep->vport_index;
63
64 if (mlx5_lag_is_shared_fdb(dev)) {
65 if (mlx5_lag_is_master(dev)) {
66 mlx5_ib_num_ports_update(dev, &num_ports);
67 } else {
68 if (rep->vport == MLX5_VPORT_UPLINK) {
69 if (!mlx5_lag_is_mpesw(dev))
70 return 0;
71 new_uplink = true;
72 }
73 mlx5_lag_for_each_peer_mdev(dev, peer_dev, i) {
74 u32 peer_n_ports = mlx5_eswitch_get_total_vports(peer_dev);
75
76 if (mlx5_lag_is_master(peer_dev))
77 lag_master = peer_dev;
78 else if (!mlx5_lag_is_mpesw(dev))
79 /* Only 1 ib port is the representor for all uplinks */
80 peer_n_ports--;
81
82 if (mlx5_get_dev_index(peer_dev) < mlx5_get_dev_index(dev))
83 vport_index += peer_n_ports;
84 }
85 }
86 }
87
88 if (rep->vport == MLX5_VPORT_UPLINK && !new_uplink)
89 profile = &raw_eth_profile;
90 else
91 return mlx5_ib_set_vport_rep(lag_master, rep, vport_index);
92
93 ibdev = ib_alloc_device(mlx5_ib_dev, ib_dev);
94 if (!ibdev)
95 return -ENOMEM;
96
97 ibdev->port = kcalloc(num_ports, sizeof(*ibdev->port),
98 GFP_KERNEL);
99 if (!ibdev->port) {
100 ret = -ENOMEM;
101 goto fail_port;
102 }
103
104 ibdev->is_rep = true;
105 vport_index = rep->vport_index;
106 ibdev->port[vport_index].rep = rep;
107 ibdev->port[vport_index].roce.netdev =
108 mlx5_ib_get_rep_netdev(lag_master->priv.eswitch, rep->vport);
109 ibdev->mdev = lag_master;
110 ibdev->num_ports = num_ports;
111
112 ret = __mlx5_ib_add(ibdev, profile);
113 if (ret)
114 goto fail_add;
115
116 rep->rep_data[REP_IB].priv = ibdev;
117 if (mlx5_lag_is_shared_fdb(lag_master))
118 mlx5_ib_register_peer_vport_reps(lag_master);
119
120 return 0;
121
122 fail_add:
123 kfree(ibdev->port);
124 fail_port:
125 ib_dealloc_device(&ibdev->ib_dev);
126 return ret;
127 }
128
mlx5_ib_rep_to_dev(struct mlx5_eswitch_rep * rep)129 static void *mlx5_ib_rep_to_dev(struct mlx5_eswitch_rep *rep)
130 {
131 return rep->rep_data[REP_IB].priv;
132 }
133
134 static void
mlx5_ib_vport_rep_unload(struct mlx5_eswitch_rep * rep)135 mlx5_ib_vport_rep_unload(struct mlx5_eswitch_rep *rep)
136 {
137 struct mlx5_core_dev *mdev = mlx5_eswitch_get_core_dev(rep->esw);
138 struct mlx5_ib_dev *dev = mlx5_ib_rep_to_dev(rep);
139 int vport_index = rep->vport_index;
140 struct mlx5_ib_port *port;
141 int i;
142
143 if (WARN_ON(!mdev))
144 return;
145
146 if (!dev)
147 return;
148
149 if (mlx5_lag_is_shared_fdb(mdev) &&
150 !mlx5_lag_is_master(mdev)) {
151 if (rep->vport == MLX5_VPORT_UPLINK && !mlx5_lag_is_mpesw(mdev))
152 return;
153 for (i = 0; i < dev->num_ports; i++) {
154 if (dev->port[i].rep == rep)
155 break;
156 }
157 if (WARN_ON(i == dev->num_ports))
158 return;
159 vport_index = i;
160 }
161
162 port = &dev->port[vport_index];
163 write_lock(&port->roce.netdev_lock);
164 port->roce.netdev = NULL;
165 write_unlock(&port->roce.netdev_lock);
166 rep->rep_data[REP_IB].priv = NULL;
167 port->rep = NULL;
168
169 if (rep->vport == MLX5_VPORT_UPLINK) {
170
171 if (mlx5_lag_is_shared_fdb(mdev) && !mlx5_lag_is_master(mdev))
172 return;
173
174 if (mlx5_lag_is_shared_fdb(mdev)) {
175 struct mlx5_core_dev *peer_mdev;
176 struct mlx5_eswitch *esw;
177
178 mlx5_lag_for_each_peer_mdev(mdev, peer_mdev, i) {
179 esw = peer_mdev->priv.eswitch;
180 mlx5_eswitch_unregister_vport_reps(esw, REP_IB);
181 }
182 }
183 __mlx5_ib_remove(dev, dev->profile, MLX5_IB_STAGE_MAX);
184 }
185 }
186
187 static const struct mlx5_eswitch_rep_ops rep_ops = {
188 .load = mlx5_ib_vport_rep_load,
189 .unload = mlx5_ib_vport_rep_unload,
190 .get_proto_dev = mlx5_ib_rep_to_dev,
191 };
192
mlx5_ib_register_peer_vport_reps(struct mlx5_core_dev * mdev)193 static void mlx5_ib_register_peer_vport_reps(struct mlx5_core_dev *mdev)
194 {
195 struct mlx5_core_dev *peer_mdev;
196 struct mlx5_eswitch *esw;
197 int i;
198
199 mlx5_lag_for_each_peer_mdev(mdev, peer_mdev, i) {
200 esw = peer_mdev->priv.eswitch;
201 mlx5_eswitch_register_vport_reps(esw, &rep_ops, REP_IB);
202 }
203 }
204
mlx5_ib_get_rep_netdev(struct mlx5_eswitch * esw,u16 vport_num)205 struct net_device *mlx5_ib_get_rep_netdev(struct mlx5_eswitch *esw,
206 u16 vport_num)
207 {
208 return mlx5_eswitch_get_proto_dev(esw, vport_num, REP_ETH);
209 }
210
create_flow_rule_vport_sq(struct mlx5_ib_dev * dev,struct mlx5_ib_sq * sq,u32 port)211 struct mlx5_flow_handle *create_flow_rule_vport_sq(struct mlx5_ib_dev *dev,
212 struct mlx5_ib_sq *sq,
213 u32 port)
214 {
215 struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
216 struct mlx5_eswitch_rep *rep;
217
218 if (!dev->is_rep || !port)
219 return NULL;
220
221 if (!dev->port[port - 1].rep)
222 return ERR_PTR(-EINVAL);
223
224 rep = dev->port[port - 1].rep;
225
226 return mlx5_eswitch_add_send_to_vport_rule(esw, esw, rep, sq->base.mqp.qpn);
227 }
228
mlx5r_rep_probe(struct auxiliary_device * adev,const struct auxiliary_device_id * id)229 static int mlx5r_rep_probe(struct auxiliary_device *adev,
230 const struct auxiliary_device_id *id)
231 {
232 struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
233 struct mlx5_core_dev *mdev = idev->mdev;
234 struct mlx5_eswitch *esw;
235
236 esw = mdev->priv.eswitch;
237 mlx5_eswitch_register_vport_reps(esw, &rep_ops, REP_IB);
238 return 0;
239 }
240
mlx5r_rep_remove(struct auxiliary_device * adev)241 static void mlx5r_rep_remove(struct auxiliary_device *adev)
242 {
243 struct mlx5_adev *idev = container_of(adev, struct mlx5_adev, adev);
244 struct mlx5_core_dev *mdev = idev->mdev;
245 struct mlx5_eswitch *esw;
246
247 esw = mdev->priv.eswitch;
248 mlx5_eswitch_unregister_vport_reps(esw, REP_IB);
249 }
250
251 static const struct auxiliary_device_id mlx5r_rep_id_table[] = {
252 { .name = MLX5_ADEV_NAME ".rdma-rep", },
253 {},
254 };
255
256 MODULE_DEVICE_TABLE(auxiliary, mlx5r_rep_id_table);
257
258 static struct auxiliary_driver mlx5r_rep_driver = {
259 .name = "rep",
260 .probe = mlx5r_rep_probe,
261 .remove = mlx5r_rep_remove,
262 .id_table = mlx5r_rep_id_table,
263 };
264
mlx5r_rep_init(void)265 int mlx5r_rep_init(void)
266 {
267 return auxiliary_driver_register(&mlx5r_rep_driver);
268 }
269
mlx5r_rep_cleanup(void)270 void mlx5r_rep_cleanup(void)
271 {
272 auxiliary_driver_unregister(&mlx5r_rep_driver);
273 }
274