1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/slab.h>
35 
36 #include "ipoib.h"
37 
38 int ipoib_mcast_attach(struct net_device *dev, u16 mlid, union ib_gid *mgid, int set_qkey)
39 {
40 	struct ipoib_dev_priv *priv = netdev_priv(dev);
41 	struct ib_qp_attr *qp_attr = NULL;
42 	int ret;
43 	u16 pkey_index;
44 
45 	if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index)) {
46 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
47 		ret = -ENXIO;
48 		goto out;
49 	}
50 	set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
51 
52 	if (set_qkey) {
53 		ret = -ENOMEM;
54 		qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
55 		if (!qp_attr)
56 			goto out;
57 
58 		/* set correct QKey for QP */
59 		qp_attr->qkey = priv->qkey;
60 		ret = ib_modify_qp(priv->qp, qp_attr, IB_QP_QKEY);
61 		if (ret) {
62 			ipoib_warn(priv, "failed to modify QP, ret = %d\n", ret);
63 			goto out;
64 		}
65 	}
66 
67 	/* attach QP to multicast group */
68 	ret = ib_attach_mcast(priv->qp, mgid, mlid);
69 	if (ret)
70 		ipoib_warn(priv, "failed to attach to multicast group, ret = %d\n", ret);
71 
72 out:
73 	kfree(qp_attr);
74 	return ret;
75 }
76 
77 int ipoib_init_qp(struct net_device *dev)
78 {
79 	struct ipoib_dev_priv *priv = netdev_priv(dev);
80 	int ret;
81 	struct ib_qp_attr qp_attr;
82 	int attr_mask;
83 
84 	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
85 		return -1;
86 
87 	qp_attr.qp_state = IB_QPS_INIT;
88 	qp_attr.qkey = 0;
89 	qp_attr.port_num = priv->port;
90 	qp_attr.pkey_index = priv->pkey_index;
91 	attr_mask =
92 	    IB_QP_QKEY |
93 	    IB_QP_PORT |
94 	    IB_QP_PKEY_INDEX |
95 	    IB_QP_STATE;
96 	ret = ib_modify_qp(priv->qp, &qp_attr, attr_mask);
97 	if (ret) {
98 		ipoib_warn(priv, "failed to modify QP to init, ret = %d\n", ret);
99 		goto out_fail;
100 	}
101 
102 	qp_attr.qp_state = IB_QPS_RTR;
103 	/* Can't set this in a INIT->RTR transition */
104 	attr_mask &= ~IB_QP_PORT;
105 	ret = ib_modify_qp(priv->qp, &qp_attr, attr_mask);
106 	if (ret) {
107 		ipoib_warn(priv, "failed to modify QP to RTR, ret = %d\n", ret);
108 		goto out_fail;
109 	}
110 
111 	qp_attr.qp_state = IB_QPS_RTS;
112 	qp_attr.sq_psn = 0;
113 	attr_mask |= IB_QP_SQ_PSN;
114 	attr_mask &= ~IB_QP_PKEY_INDEX;
115 	ret = ib_modify_qp(priv->qp, &qp_attr, attr_mask);
116 	if (ret) {
117 		ipoib_warn(priv, "failed to modify QP to RTS, ret = %d\n", ret);
118 		goto out_fail;
119 	}
120 
121 	return 0;
122 
123 out_fail:
124 	qp_attr.qp_state = IB_QPS_RESET;
125 	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))
126 		ipoib_warn(priv, "Failed to modify QP to RESET state\n");
127 
128 	return ret;
129 }
130 
131 int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
132 {
133 	struct ipoib_dev_priv *priv = netdev_priv(dev);
134 	struct ib_qp_init_attr init_attr = {
135 		.cap = {
136 			.max_send_wr  = ipoib_sendq_size,
137 			.max_recv_wr  = ipoib_recvq_size,
138 			.max_send_sge = 1,
139 			.max_recv_sge = IPOIB_UD_RX_SG
140 		},
141 		.sq_sig_type = IB_SIGNAL_ALL_WR,
142 		.qp_type     = IB_QPT_UD
143 	};
144 	struct ib_cq_init_attr cq_attr = {};
145 
146 	int ret, size;
147 	int i;
148 
149 	priv->pd = ib_alloc_pd(priv->ca);
150 	if (IS_ERR(priv->pd)) {
151 		printk(KERN_WARNING "%s: failed to allocate PD\n", ca->name);
152 		return -ENODEV;
153 	}
154 
155 	/*
156 	 * the various IPoIB tasks assume they will never race against
157 	 * themselves, so always use a single thread workqueue
158 	 */
159 	priv->wq = create_singlethread_workqueue("ipoib_wq");
160 	if (!priv->wq) {
161 		printk(KERN_WARNING "ipoib: failed to allocate device WQ\n");
162 		goto out_free_pd;
163 	}
164 
165 	size = ipoib_recvq_size + 1;
166 	ret = ipoib_cm_dev_init(dev);
167 	if (!ret) {
168 		size += ipoib_sendq_size;
169 		if (ipoib_cm_has_srq(dev))
170 			size += ipoib_recvq_size + 1; /* 1 extra for rx_drain_qp */
171 		else
172 			size += ipoib_recvq_size * ipoib_max_conn_qp;
173 	} else
174 		if (ret != -ENOSYS)
175 			goto out_free_wq;
176 
177 	cq_attr.cqe = size;
178 	priv->recv_cq = ib_create_cq(priv->ca, ipoib_ib_completion, NULL,
179 				     dev, &cq_attr);
180 	if (IS_ERR(priv->recv_cq)) {
181 		printk(KERN_WARNING "%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 	priv->send_cq = ib_create_cq(priv->ca, ipoib_send_comp_handler, NULL,
187 				     dev, &cq_attr);
188 	if (IS_ERR(priv->send_cq)) {
189 		printk(KERN_WARNING "%s: failed to create send CQ\n", ca->name);
190 		goto out_free_recv_cq;
191 	}
192 
193 	if (ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP))
194 		goto out_free_send_cq;
195 
196 	init_attr.send_cq = priv->send_cq;
197 	init_attr.recv_cq = priv->recv_cq;
198 
199 	if (priv->hca_caps & IB_DEVICE_UD_TSO)
200 		init_attr.create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
201 
202 	if (priv->hca_caps & IB_DEVICE_BLOCK_MULTICAST_LOOPBACK)
203 		init_attr.create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
204 
205 	if (priv->hca_caps & IB_DEVICE_MANAGED_FLOW_STEERING)
206 		init_attr.create_flags |= IB_QP_CREATE_NETIF_QP;
207 
208 	if (dev->features & NETIF_F_SG)
209 		init_attr.cap.max_send_sge =
210 			min_t(u32, priv->ca->attrs.max_sge, MAX_SKB_FRAGS + 1);
211 
212 	priv->qp = ib_create_qp(priv->pd, &init_attr);
213 	if (IS_ERR(priv->qp)) {
214 		printk(KERN_WARNING "%s: failed to create QP\n", ca->name);
215 		goto out_free_send_cq;
216 	}
217 
218 	priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
219 	priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
220 	priv->dev->dev_addr[3] = (priv->qp->qp_num      ) & 0xff;
221 
222 	for (i = 0; i < MAX_SKB_FRAGS + 1; ++i)
223 		priv->tx_sge[i].lkey = priv->pd->local_dma_lkey;
224 
225 	priv->tx_wr.wr.opcode		= IB_WR_SEND;
226 	priv->tx_wr.wr.sg_list		= priv->tx_sge;
227 	priv->tx_wr.wr.send_flags	= IB_SEND_SIGNALED;
228 
229 	priv->rx_sge[0].lkey = priv->pd->local_dma_lkey;
230 
231 	priv->rx_sge[0].length = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
232 	priv->rx_wr.num_sge = 1;
233 
234 	priv->rx_wr.next = NULL;
235 	priv->rx_wr.sg_list = priv->rx_sge;
236 
237 	priv->max_send_sge = init_attr.cap.max_send_sge;
238 
239 	return 0;
240 
241 out_free_send_cq:
242 	ib_destroy_cq(priv->send_cq);
243 
244 out_free_recv_cq:
245 	ib_destroy_cq(priv->recv_cq);
246 
247 out_cm_dev_cleanup:
248 	ipoib_cm_dev_cleanup(dev);
249 
250 out_free_wq:
251 	destroy_workqueue(priv->wq);
252 	priv->wq = NULL;
253 
254 out_free_pd:
255 	ib_dealloc_pd(priv->pd);
256 
257 	return -ENODEV;
258 }
259 
260 void ipoib_transport_dev_cleanup(struct net_device *dev)
261 {
262 	struct ipoib_dev_priv *priv = netdev_priv(dev);
263 
264 	if (priv->qp) {
265 		if (ib_destroy_qp(priv->qp))
266 			ipoib_warn(priv, "ib_qp_destroy failed\n");
267 
268 		priv->qp = NULL;
269 		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
270 	}
271 
272 	if (ib_destroy_cq(priv->send_cq))
273 		ipoib_warn(priv, "ib_cq_destroy (send) failed\n");
274 
275 	if (ib_destroy_cq(priv->recv_cq))
276 		ipoib_warn(priv, "ib_cq_destroy (recv) failed\n");
277 
278 	ipoib_cm_dev_cleanup(dev);
279 
280 	if (priv->wq) {
281 		flush_workqueue(priv->wq);
282 		destroy_workqueue(priv->wq);
283 		priv->wq = NULL;
284 	}
285 
286 	ib_dealloc_pd(priv->pd);
287 }
288 
289 void ipoib_event(struct ib_event_handler *handler,
290 		 struct ib_event *record)
291 {
292 	struct ipoib_dev_priv *priv =
293 		container_of(handler, struct ipoib_dev_priv, event_handler);
294 
295 	if (record->element.port_num != priv->port)
296 		return;
297 
298 	ipoib_dbg(priv, "Event %d on device %s port %d\n", record->event,
299 		  record->device->name, record->element.port_num);
300 
301 	if (record->event == IB_EVENT_SM_CHANGE ||
302 	    record->event == IB_EVENT_CLIENT_REREGISTER) {
303 		queue_work(ipoib_workqueue, &priv->flush_light);
304 	} else if (record->event == IB_EVENT_PORT_ERR ||
305 		   record->event == IB_EVENT_PORT_ACTIVE ||
306 		   record->event == IB_EVENT_LID_CHANGE) {
307 		queue_work(ipoib_workqueue, &priv->flush_normal);
308 	} else if (record->event == IB_EVENT_PKEY_CHANGE) {
309 		queue_work(ipoib_workqueue, &priv->flush_heavy);
310 	} else if (record->event == IB_EVENT_GID_CHANGE &&
311 		   !test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {
312 		queue_work(ipoib_workqueue, &priv->flush_light);
313 	}
314 }
315