]> err.no Git - linux-2.6/blob - drivers/infiniband/ulp/ipoib/ipoib_verbs.c
IPoIB: Refresh paths instead of flushing them on SM change events
[linux-2.6] / drivers / infiniband / ulp / ipoib / ipoib_verbs.c
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 "ipoib.h"
35
36 int ipoib_mcast_attach(struct net_device *dev, u16 mlid, union ib_gid *mgid)
37 {
38         struct ipoib_dev_priv *priv = netdev_priv(dev);
39         struct ib_qp_attr *qp_attr;
40         int ret;
41         u16 pkey_index;
42
43         ret = -ENOMEM;
44         qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
45         if (!qp_attr)
46                 goto out;
47
48         if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index)) {
49                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
50                 ret = -ENXIO;
51                 goto out;
52         }
53         set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
54
55         /* set correct QKey for QP */
56         qp_attr->qkey = priv->qkey;
57         ret = ib_modify_qp(priv->qp, qp_attr, IB_QP_QKEY);
58         if (ret) {
59                 ipoib_warn(priv, "failed to modify QP, ret = %d\n", ret);
60                 goto out;
61         }
62
63         /* attach QP to multicast group */
64         mutex_lock(&priv->mcast_mutex);
65         ret = ib_attach_mcast(priv->qp, mgid, mlid);
66         mutex_unlock(&priv->mcast_mutex);
67         if (ret)
68                 ipoib_warn(priv, "failed to attach to multicast group, ret = %d\n", ret);
69
70 out:
71         kfree(qp_attr);
72         return ret;
73 }
74
75 int ipoib_mcast_detach(struct net_device *dev, u16 mlid, union ib_gid *mgid)
76 {
77         struct ipoib_dev_priv *priv = netdev_priv(dev);
78         int ret;
79
80         mutex_lock(&priv->mcast_mutex);
81         ret = ib_detach_mcast(priv->qp, mgid, mlid);
82         mutex_unlock(&priv->mcast_mutex);
83         if (ret)
84                 ipoib_warn(priv, "ib_detach_mcast failed (result = %d)\n", ret);
85
86         return ret;
87 }
88
89 int ipoib_init_qp(struct net_device *dev)
90 {
91         struct ipoib_dev_priv *priv = netdev_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         /* Can't set this in a INIT->RTR transition */
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
135 out_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
143 int ipoib_transport_dev_init(struct net_device *dev, struct ib_device *ca)
144 {
145         struct ipoib_dev_priv *priv = netdev_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 = 1,
151                         .max_recv_sge = IPOIB_UD_RX_SG
152                 },
153                 .sq_sig_type = IB_SIGNAL_ALL_WR,
154                 .qp_type     = IB_QPT_UD
155         };
156
157         int ret, size;
158         int i;
159
160         priv->pd = ib_alloc_pd(priv->ca);
161         if (IS_ERR(priv->pd)) {
162                 printk(KERN_WARNING "%s: failed to allocate PD\n", ca->name);
163                 return -ENODEV;
164         }
165
166         priv->mr = ib_get_dma_mr(priv->pd, IB_ACCESS_LOCAL_WRITE);
167         if (IS_ERR(priv->mr)) {
168                 printk(KERN_WARNING "%s: ib_get_dma_mr failed\n", ca->name);
169                 goto out_free_pd;
170         }
171
172         size = ipoib_recvq_size + 1;
173         ret = ipoib_cm_dev_init(dev);
174         if (!ret) {
175                 size += ipoib_sendq_size;
176                 if (ipoib_cm_has_srq(dev))
177                         size += ipoib_recvq_size + 1; /* 1 extra for rx_drain_qp */
178                 else
179                         size += ipoib_recvq_size * ipoib_max_conn_qp;
180         }
181
182         priv->recv_cq = ib_create_cq(priv->ca, ipoib_ib_completion, NULL, dev, size, 0);
183         if (IS_ERR(priv->recv_cq)) {
184                 printk(KERN_WARNING "%s: failed to create receive CQ\n", ca->name);
185                 goto out_free_mr;
186         }
187
188         priv->send_cq = ib_create_cq(priv->ca, ipoib_send_comp_handler, NULL,
189                                      dev, ipoib_sendq_size, 0);
190         if (IS_ERR(priv->send_cq)) {
191                 printk(KERN_WARNING "%s: failed to create send CQ\n", ca->name);
192                 goto out_free_recv_cq;
193         }
194
195         if (ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP))
196                 goto out_free_send_cq;
197
198         init_attr.send_cq = priv->send_cq;
199         init_attr.recv_cq = priv->recv_cq;
200
201         if (priv->hca_caps & IB_DEVICE_UD_TSO)
202                 init_attr.create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
203
204         if (priv->hca_caps & IB_DEVICE_BLOCK_MULTICAST_LOOPBACK)
205                 init_attr.create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
206
207         if (dev->features & NETIF_F_SG)
208                 init_attr.cap.max_send_sge = MAX_SKB_FRAGS + 1;
209
210         priv->qp = ib_create_qp(priv->pd, &init_attr);
211         if (IS_ERR(priv->qp)) {
212                 printk(KERN_WARNING "%s: failed to create QP\n", ca->name);
213                 goto out_free_send_cq;
214         }
215
216         priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
217         priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
218         priv->dev->dev_addr[3] = (priv->qp->qp_num      ) & 0xff;
219
220         for (i = 0; i < MAX_SKB_FRAGS + 1; ++i)
221                 priv->tx_sge[i].lkey = priv->mr->lkey;
222
223         priv->tx_wr.opcode      = IB_WR_SEND;
224         priv->tx_wr.sg_list     = priv->tx_sge;
225         priv->tx_wr.send_flags  = IB_SEND_SIGNALED;
226
227         priv->rx_sge[0].lkey = priv->mr->lkey;
228         if (ipoib_ud_need_sg(priv->max_ib_mtu)) {
229                 priv->rx_sge[0].length = IPOIB_UD_HEAD_SIZE;
230                 priv->rx_sge[1].length = PAGE_SIZE;
231                 priv->rx_sge[1].lkey = priv->mr->lkey;
232                 priv->rx_wr.num_sge = IPOIB_UD_RX_SG;
233         } else {
234                 priv->rx_sge[0].length = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
235                 priv->rx_wr.num_sge = 1;
236         }
237         priv->rx_wr.next = NULL;
238         priv->rx_wr.sg_list = priv->rx_sge;
239
240         return 0;
241
242 out_free_send_cq:
243         ib_destroy_cq(priv->send_cq);
244
245 out_free_recv_cq:
246         ib_destroy_cq(priv->recv_cq);
247
248 out_free_mr:
249         ib_dereg_mr(priv->mr);
250         ipoib_cm_dev_cleanup(dev);
251
252 out_free_pd:
253         ib_dealloc_pd(priv->pd);
254         return -ENODEV;
255 }
256
257 void ipoib_transport_dev_cleanup(struct net_device *dev)
258 {
259         struct ipoib_dev_priv *priv = netdev_priv(dev);
260
261         if (priv->qp) {
262                 if (ib_destroy_qp(priv->qp))
263                         ipoib_warn(priv, "ib_qp_destroy failed\n");
264
265                 priv->qp = NULL;
266                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
267         }
268
269         if (ib_destroy_cq(priv->send_cq))
270                 ipoib_warn(priv, "ib_cq_destroy (send) failed\n");
271
272         if (ib_destroy_cq(priv->recv_cq))
273                 ipoib_warn(priv, "ib_cq_destroy (recv) failed\n");
274
275         ipoib_cm_dev_cleanup(dev);
276
277         if (ib_dereg_mr(priv->mr))
278                 ipoib_warn(priv, "ib_dereg_mr failed\n");
279
280         if (ib_dealloc_pd(priv->pd))
281                 ipoib_warn(priv, "ib_dealloc_pd failed\n");
282 }
283
284 void ipoib_event(struct ib_event_handler *handler,
285                  struct ib_event *record)
286 {
287         struct ipoib_dev_priv *priv =
288                 container_of(handler, struct ipoib_dev_priv, event_handler);
289
290         if (record->element.port_num != priv->port)
291                 return;
292
293         ipoib_dbg(priv, "Event %d on device %s port %d\n", record->event,
294                   record->device->name, record->element.port_num);
295
296         if (record->event == IB_EVENT_SM_CHANGE ||
297             record->event == IB_EVENT_CLIENT_REREGISTER) {
298                 queue_work(ipoib_workqueue, &priv->flush_light);
299         } else if (record->event == IB_EVENT_PORT_ERR ||
300                    record->event == IB_EVENT_PORT_ACTIVE ||
301                    record->event == IB_EVENT_LID_CHANGE) {
302                 queue_work(ipoib_workqueue, &priv->flush_normal);
303         } else if (record->event == IB_EVENT_PKEY_CHANGE) {
304                 queue_work(ipoib_workqueue, &priv->flush_heavy);
305         }
306 }