1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34#include <linux/slab.h>
35
36#include "ipoib.h"
37
38int ipoib_mcast_attach(struct net_device *dev, struct ib_device *hca,
39 union ib_gid *mgid, u16 mlid, int set_qkey, u32 qkey)
40{
41 struct ipoib_dev_priv *priv = ipoib_priv(dev);
42 struct ib_qp_attr *qp_attr = NULL;
43 int ret;
44 u16 pkey_index;
45
46 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index)) {
47 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
48 ret = -ENXIO;
49 goto out;
50 }
51 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
52
53 if (set_qkey) {
54 ret = -ENOMEM;
55 qp_attr = kmalloc(sizeof(*qp_attr), GFP_KERNEL);
56 if (!qp_attr)
57 goto out;
58
59
60 qp_attr->qkey = qkey;
61 ret = ib_modify_qp(priv->qp, qp_attr, IB_QP_QKEY);
62 if (ret) {
63 ipoib_warn(priv, "failed to modify QP, ret = %d\n", ret);
64 goto out;
65 }
66 }
67
68
69 ret = ib_attach_mcast(priv->qp, mgid, mlid);
70 if (ret)
71 ipoib_warn(priv, "failed to attach to multicast group, ret = %d\n", ret);
72
73out:
74 kfree(qp_attr);
75 return ret;
76}
77
78int ipoib_mcast_detach(struct net_device *dev, struct ib_device *hca,
79 union ib_gid *mgid, u16 mlid)
80{
81 struct ipoib_dev_priv *priv = ipoib_priv(dev);
82 int ret;
83
84 ret = ib_detach_mcast(priv->qp, mgid, mlid);
85
86 return ret;
87}
88
89int ipoib_init_qp(struct net_device *dev)
90{
91 struct ipoib_dev_priv *priv = ipoib_priv(dev);
92 int ret;
93 struct ib_qp_attr qp_attr;
94 int attr_mask;
95
96 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
97 return -1;
98
99 qp_attr.qp_state = IB_QPS_INIT;
100 qp_attr.qkey = 0;
101 qp_attr.port_num = priv->port;
102 qp_attr.pkey_index = priv->pkey_index;
103 attr_mask =
104 IB_QP_QKEY |
105 IB_QP_PORT |
106 IB_QP_PKEY_INDEX |
107 IB_QP_STATE;
108 ret = ib_modify_qp(priv->qp, &qp_attr, attr_mask);
109 if (ret) {
110 ipoib_warn(priv, "failed to modify QP to init, ret = %d\n", ret);
111 goto out_fail;
112 }
113
114 qp_attr.qp_state = IB_QPS_RTR;
115
116 attr_mask &= ~IB_QP_PORT;
117 ret = ib_modify_qp(priv->qp, &qp_attr, attr_mask);
118 if (ret) {
119 ipoib_warn(priv, "failed to modify QP to RTR, ret = %d\n", ret);
120 goto out_fail;
121 }
122
123 qp_attr.qp_state = IB_QPS_RTS;
124 qp_attr.sq_psn = 0;
125 attr_mask |= IB_QP_SQ_PSN;
126 attr_mask &= ~IB_QP_PKEY_INDEX;
127 ret = ib_modify_qp(priv->qp, &qp_attr, attr_mask);
128 if (ret) {
129 ipoib_warn(priv, "failed to modify QP to RTS, ret = %d\n", ret);
130 goto out_fail;
131 }
132
133 return 0;
134
135out_fail:
136 qp_attr.qp_state = IB_QPS_RESET;
137 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
138 ipoib_warn(priv, "Failed to modify QP to RESET state\n");
139
140 return ret;
141}
142
143int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
144{
145 struct ipoib_dev_priv *priv = ipoib_priv(dev);
146 struct ib_qp_init_attr init_attr = {
147 .cap = {
148 .max_send_wr = ipoib_sendq_size,
149 .max_recv_wr = ipoib_recvq_size,
150 .max_send_sge = min_t(u32, priv->ca->attrs.max_send_sge,
151 MAX_SKB_FRAGS + 1),
152 .max_recv_sge = IPOIB_UD_RX_SG
153 },
154 .sq_sig_type = IB_SIGNAL_ALL_WR,
155 .qp_type = IB_QPT_UD
156 };
157 struct ib_cq_init_attr cq_attr = {};
158
159 int ret, size, req_vec;
160 int i;
161 static atomic_t counter;
162
163 size = ipoib_recvq_size + 1;
164 ret = ipoib_cm_dev_init(dev);
165 if (!ret) {
166 size += ipoib_sendq_size;
167 if (ipoib_cm_has_srq(dev))
168 size += ipoib_recvq_size + 1;
169 else
170 size += ipoib_recvq_size * ipoib_max_conn_qp;
171 } else
172 if (ret != -EOPNOTSUPP)
173 return ret;
174
175 req_vec = atomic_inc_return(&counter) * 2;
176 cq_attr.cqe = size;
177 cq_attr.comp_vector = req_vec % priv->ca->num_comp_vectors;
178 priv->recv_cq = ib_create_cq(priv->ca, ipoib_ib_rx_completion, NULL,
179 priv, &cq_attr);
180 if (IS_ERR(priv->recv_cq)) {
181 pr_warn("%s: failed to create receive CQ\n", ca->name);
182 goto out_cm_dev_cleanup;
183 }
184
185 cq_attr.cqe = ipoib_sendq_size;
186 cq_attr.comp_vector = (req_vec + 1) % priv->ca->num_comp_vectors;
187 priv->send_cq = ib_create_cq(priv->ca, ipoib_ib_tx_completion, NULL,
188 priv, &cq_attr);
189 if (IS_ERR(priv->send_cq)) {
190 pr_warn("%s: failed to create send CQ\n", ca->name);
191 goto out_free_recv_cq;
192 }
193
194 if (ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP))
195 goto out_free_send_cq;
196
197 init_attr.send_cq = priv->send_cq;
198 init_attr.recv_cq = priv->recv_cq;
199
200 if (priv->hca_caps & IB_DEVICE_UD_TSO)
201 init_attr.create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
202
203 if (priv->hca_caps & IB_DEVICE_BLOCK_MULTICAST_LOOPBACK)
204 init_attr.create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
205
206 if (priv->hca_caps & IB_DEVICE_MANAGED_FLOW_STEERING)
207 init_attr.create_flags |= IB_QP_CREATE_NETIF_QP;
208
209 if (priv->hca_caps & IB_DEVICE_RDMA_NETDEV_OPA)
210 init_attr.create_flags |= IB_QP_CREATE_NETDEV_USE;
211
212 priv->qp = ib_create_qp(priv->pd, &init_attr);
213 if (IS_ERR(priv->qp)) {
214 pr_warn("%s: failed to create QP\n", ca->name);
215 goto out_free_send_cq;
216 }
217
218 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
219 goto out_free_send_cq;
220
221 for (i = 0; i < MAX_SKB_FRAGS + 1; ++i)
222 priv->tx_sge[i].lkey = priv->pd->local_dma_lkey;
223
224 priv->tx_wr.wr.opcode = IB_WR_SEND;
225 priv->tx_wr.wr.sg_list = priv->tx_sge;
226 priv->tx_wr.wr.send_flags = IB_SEND_SIGNALED;
227
228 priv->rx_sge[0].lkey = priv->pd->local_dma_lkey;
229
230 priv->rx_sge[0].length = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
231 priv->rx_wr.num_sge = 1;
232
233 priv->rx_wr.next = NULL;
234 priv->rx_wr.sg_list = priv->rx_sge;
235
236 if (init_attr.cap.max_send_sge > 1)
237 dev->features |= NETIF_F_SG;
238
239 priv->max_send_sge = init_attr.cap.max_send_sge;
240
241 return 0;
242
243out_free_send_cq:
244 ib_destroy_cq(priv->send_cq);
245
246out_free_recv_cq:
247 ib_destroy_cq(priv->recv_cq);
248
249out_cm_dev_cleanup:
250 ipoib_cm_dev_cleanup(dev);
251
252 return -ENODEV;
253}
254
255void ipoib_transport_dev_cleanup(struct net_device *dev)
256{
257 struct ipoib_dev_priv *priv = ipoib_priv(dev);
258
259 if (priv->qp) {
260 if (ib_destroy_qp(priv->qp))
261 ipoib_warn(priv, "ib_qp_destroy failed\n");
262
263 priv->qp = NULL;
264 }
265
266 ib_destroy_cq(priv->send_cq);
267 ib_destroy_cq(priv->recv_cq);
268}
269
270void ipoib_event(struct ib_event_handler *handler,
271 struct ib_event *record)
272{
273 struct ipoib_dev_priv *priv =
274 container_of(handler, struct ipoib_dev_priv, event_handler);
275
276 if (record->element.port_num != priv->port)
277 return;
278
279 ipoib_dbg(priv, "Event %d on device %s port %d\n", record->event,
280 dev_name(&record->device->dev), record->element.port_num);
281
282 if (record->event == IB_EVENT_CLIENT_REREGISTER) {
283 queue_work(ipoib_workqueue, &priv->flush_light);
284 } else if (record->event == IB_EVENT_PORT_ERR ||
285 record->event == IB_EVENT_PORT_ACTIVE ||
286 record->event == IB_EVENT_LID_CHANGE) {
287 queue_work(ipoib_workqueue, &priv->flush_normal);
288 } else if (record->event == IB_EVENT_PKEY_CHANGE) {
289 queue_work(ipoib_workqueue, &priv->flush_heavy);
290 } else if (record->event == IB_EVENT_GID_CHANGE &&
291 !test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
292 queue_work(ipoib_workqueue, &priv->flush_light);
293 }
294}
295