kernel/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 <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
 145         int ret, size;
 146         int i;
 147
 148         priv->pd = ib_alloc_pd(priv->ca);
 149         if (IS_ERR(priv->pd)) {
 150                 printk(KERN_WARNING "%s: failed to allocate PD\n", ca->name);
 151                 return -ENODEV;
 152         }
 153
 154         priv->mr = ib_get_dma_mr(priv->pd, IB_ACCESS_LOCAL_WRITE);
 155         if (IS_ERR(priv->mr)) {
 156                 printk(KERN_WARNING "%s: ib_get_dma_mr failed\n", ca->name);
 157                 goto out_free_pd;
 158         }
 159
 160         /*
 161          * the various IPoIB tasks assume they will never race against
 162          * themselves, so always use a single thread workqueue
 163          */
 164         priv->wq = create_singlethread_workqueue("ipoib_wq");
 165         if (!priv->wq) {
 166                 printk(KERN_WARNING "ipoib: failed to allocate device WQ\n");
 167                 goto out_free_mr;
 168         }
 169
 170         size = ipoib_recvq_size + 1;
 171         ret = ipoib_cm_dev_init(dev);
 172         if (!ret) {
 173                 size += ipoib_sendq_size;
 174                 if (ipoib_cm_has_srq(dev))
 175                         size += ipoib_recvq_size + 1; /* 1 extra for rx_drain_qp */
 176                 else
 177                         size += ipoib_recvq_size * ipoib_max_conn_qp;
 178         } else
 179                 if (ret != -ENOSYS)
 180                         goto out_free_wq;
 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_cm_dev_cleanup;
 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 (priv->hca_caps & IB_DEVICE_MANAGED_FLOW_STEERING)
 208                 init_attr.create_flags |= IB_QP_CREATE_NETIF_QP;
 209
 210         if (dev->features & NETIF_F_SG)
 211                 init_attr.cap.max_send_sge = MAX_SKB_FRAGS + 1;
 212
 213         priv->qp = ib_create_qp(priv->pd, &init_attr);
 214         if (IS_ERR(priv->qp)) {
 215                 printk(KERN_WARNING "%s: failed to create QP\n", ca->name);
 216                 goto out_free_send_cq;
 217         }
 218
 219         priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
 220         priv->dev->dev_addr[2] = (priv->qp->qp_num >>  8) & 0xff;
 221         priv->dev->dev_addr[3] = (priv->qp->qp_num      ) & 0xff;
 222
 223         for (i = 0; i < MAX_SKB_FRAGS + 1; ++i)
 224                 priv->tx_sge[i].lkey = priv->mr->lkey;
 225
 226         priv->tx_wr.opcode      = IB_WR_SEND;
 227         priv->tx_wr.sg_list     = priv->tx_sge;
 228         priv->tx_wr.send_flags  = IB_SEND_SIGNALED;
 229
 230         priv->rx_sge[0].lkey = priv->mr->lkey;
 231
 232         priv->rx_sge[0].length = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);
 233         priv->rx_wr.num_sge = 1;
 234
 235         priv->rx_wr.next = NULL;
 236         priv->rx_wr.sg_list = priv->rx_sge;
 237
 238         return 0;
 239
 240 out_free_send_cq:
 241         ib_destroy_cq(priv->send_cq);
 242
 243 out_free_recv_cq:
 244         ib_destroy_cq(priv->recv_cq);
 245
 246 out_cm_dev_cleanup:
 247         ipoib_cm_dev_cleanup(dev);
 248
 249 out_free_wq:
 250         destroy_workqueue(priv->wq);
 251         priv->wq = NULL;
 252
 253 out_free_mr:
 254         ib_dereg_mr(priv->mr);
 255
 256 out_free_pd:
 257         ib_dealloc_pd(priv->pd);
 258
 259         return -ENODEV;
 260 }
 261
 262 void ipoib_transport_dev_cleanup(struct net_device *dev)
 263 {
 264         struct ipoib_dev_priv *priv = netdev_priv(dev);
 265
 266         if (priv->qp) {
 267                 if (ib_destroy_qp(priv->qp))
 268                         ipoib_warn(priv, "ib_qp_destroy failed\n");
 269
 270                 priv->qp = NULL;
 271                 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);
 272         }
 273
 274         if (ib_destroy_cq(priv->send_cq))
 275                 ipoib_warn(priv, "ib_cq_destroy (send) failed\n");
 276
 277         if (ib_destroy_cq(priv->recv_cq))
 278                 ipoib_warn(priv, "ib_cq_destroy (recv) failed\n");
 279
 280         ipoib_cm_dev_cleanup(dev);
 281
 282         if (priv->wq) {
 283                 flush_workqueue(priv->wq);
 284                 destroy_workqueue(priv->wq);
 285                 priv->wq = NULL;
 286         }
 287
 288         if (ib_dereg_mr(priv->mr))
 289                 ipoib_warn(priv, "ib_dereg_mr failed\n");
 290
 291         if (ib_dealloc_pd(priv->pd))
 292                 ipoib_warn(priv, "ib_dealloc_pd failed\n");
 293
 294 }
 295
 296 void ipoib_event(struct ib_event_handler *handler,
 297                  struct ib_event *record)
 298 {
 299         struct ipoib_dev_priv *priv =
 300                 container_of(handler, struct ipoib_dev_priv, event_handler);
 301
 302         if (record->element.port_num != priv->port)
 303                 return;
 304
 305         ipoib_dbg(priv, "Event %d on device %s port %d\n", record->event,
 306                   record->device->name, record->element.port_num);
 307
 308         if (record->event == IB_EVENT_SM_CHANGE ||
 309             record->event == IB_EVENT_CLIENT_REREGISTER) {
 310                 queue_work(ipoib_workqueue, &priv->flush_light);
 311         } else if (record->event == IB_EVENT_PORT_ERR ||
 312                    record->event == IB_EVENT_PORT_ACTIVE ||
 313                    record->event == IB_EVENT_LID_CHANGE) {
 314                 queue_work(ipoib_workqueue, &priv->flush_normal);
 315         } else if (record->event == IB_EVENT_PKEY_CHANGE) {
 316                 queue_work(ipoib_workqueue, &priv->flush_heavy);
 317         }
 318 }