net/smc/smc_ib.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  IB infrastructure:
   6  *  Establish SMC-R as an Infiniband Client to be notified about added and
   7  *  removed IB devices of type RDMA.
   8  *  Determine device and port characteristics for these IB devices.
   9  *
  10  *  Copyright IBM Corp. 2016
  11  *
  12  *  Author(s):  Ursula Braun <[email protected]>
  13  */
  14
  15 #include <linux/random.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/scatterlist.h>
  18 #include <rdma/ib_verbs.h>
  19 #include <rdma/ib_cache.h>
  20
  21 #include "smc_pnet.h"
  22 #include "smc_ib.h"
  23 #include "smc_core.h"
  24 #include "smc_wr.h"
  25 #include "smc.h"
  26
  27 #define SMC_MAX_CQE 32766       /* max. # of completion queue elements */
  28
  29 #define SMC_QP_MIN_RNR_TIMER            5
  30 #define SMC_QP_TIMEOUT                  15 /* 4096 * 2 ** timeout usec */
  31 #define SMC_QP_RETRY_CNT                        7 /* 7: infinite */
  32 #define SMC_QP_RNR_RETRY                        7 /* 7: infinite */
  33
  34 struct smc_ib_devices smc_ib_devices = {        /* smc-registered ib devices */
  35         .lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock),
  36         .list = LIST_HEAD_INIT(smc_ib_devices.list),
  37 };
  38
  39 #define SMC_LOCAL_SYSTEMID_RESET        "%%%%%%%"
  40
  41 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET; /* unique system
  42                                                                  * identifier
  43                                                                  */
  44
  45 static int smc_ib_modify_qp_init(struct smc_link *lnk)
  46 {
  47         struct ib_qp_attr qp_attr;
  48
  49         memset(&qp_attr, 0, sizeof(qp_attr));
  50         qp_attr.qp_state = IB_QPS_INIT;
  51         qp_attr.pkey_index = 0;
  52         qp_attr.port_num = lnk->ibport;
  53         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
  54                                 | IB_ACCESS_REMOTE_WRITE;
  55         return ib_modify_qp(lnk->roce_qp, &qp_attr,
  56                             IB_QP_STATE | IB_QP_PKEY_INDEX |
  57                             IB_QP_ACCESS_FLAGS | IB_QP_PORT);
  58 }
  59
  60 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
  61 {
  62         enum ib_qp_attr_mask qp_attr_mask =
  63                 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
  64                 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
  65         struct ib_qp_attr qp_attr;
  66
  67         memset(&qp_attr, 0, sizeof(qp_attr));
  68         qp_attr.qp_state = IB_QPS_RTR;
  69         qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
  70         qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
  71         rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
  72         rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, 1, 0);
  73         rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
  74         memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
  75                sizeof(lnk->peer_mac));
  76         qp_attr.dest_qp_num = lnk->peer_qpn;
  77         qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
  78         qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
  79                                          * requests
  80                                          */
  81         qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
  82
  83         return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
  84 }
  85
  86 int smc_ib_modify_qp_rts(struct smc_link *lnk)
  87 {
  88         struct ib_qp_attr qp_attr;
  89
  90         memset(&qp_attr, 0, sizeof(qp_attr));
  91         qp_attr.qp_state = IB_QPS_RTS;
  92         qp_attr.timeout = SMC_QP_TIMEOUT;       /* local ack timeout */
  93         qp_attr.retry_cnt = SMC_QP_RETRY_CNT;   /* retry count */
  94         qp_attr.rnr_retry = SMC_QP_RNR_RETRY;   /* RNR retries, 7=infinite */
  95         qp_attr.sq_psn = lnk->psn_initial;      /* starting send packet seq # */
  96         qp_attr.max_rd_atomic = 1;      /* # of outstanding RDMA reads and
  97                                          * atomic ops allowed
  98                                          */
  99         return ib_modify_qp(lnk->roce_qp, &qp_attr,
 100                             IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
 101                             IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
 102                             IB_QP_MAX_QP_RD_ATOMIC);
 103 }
 104
 105 int smc_ib_modify_qp_reset(struct smc_link *lnk)
 106 {
 107         struct ib_qp_attr qp_attr;
 108
 109         memset(&qp_attr, 0, sizeof(qp_attr));
 110         qp_attr.qp_state = IB_QPS_RESET;
 111         return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
 112 }
 113
 114 int smc_ib_ready_link(struct smc_link *lnk)
 115 {
 116         struct smc_link_group *lgr = smc_get_lgr(lnk);
 117         int rc = 0;
 118
 119         rc = smc_ib_modify_qp_init(lnk);
 120         if (rc)
 121                 goto out;
 122
 123         rc = smc_ib_modify_qp_rtr(lnk);
 124         if (rc)
 125                 goto out;
 126         smc_wr_remember_qp_attr(lnk);
 127         rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
 128                               IB_CQ_SOLICITED_MASK);
 129         if (rc)
 130                 goto out;
 131         rc = smc_wr_rx_post_init(lnk);
 132         if (rc)
 133                 goto out;
 134         smc_wr_remember_qp_attr(lnk);
 135
 136         if (lgr->role == SMC_SERV) {
 137                 rc = smc_ib_modify_qp_rts(lnk);
 138                 if (rc)
 139                         goto out;
 140                 smc_wr_remember_qp_attr(lnk);
 141         }
 142 out:
 143         return rc;
 144 }
 145
 146 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
 147 {
 148         const struct ib_gid_attr *attr;
 149         int rc;
 150
 151         attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
 152         if (IS_ERR(attr))
 153                 return -ENODEV;
 154
 155         rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]);
 156         rdma_put_gid_attr(attr);
 157         return rc;
 158 }
 159
 160 /* Create an identifier unique for this instance of SMC-R.
 161  * The MAC-address of the first active registered IB device
 162  * plus a random 2-byte number is used to create this identifier.
 163  * This name is delivered to the peer during connection initialization.
 164  */
 165 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
 166                                                 u8 ibport)
 167 {
 168         memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
 169                sizeof(smcibdev->mac[ibport - 1]));
 170         get_random_bytes(&local_systemid[0], 2);
 171 }
 172
 173 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
 174 {
 175         return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
 176 }
 177
 178 /* determine the gid for an ib-device port and vlan id */
 179 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
 180                          unsigned short vlan_id, u8 gid[], u8 *sgid_index)
 181 {
 182         const struct ib_gid_attr *attr;
 183         const struct net_device *ndev;
 184         int i;
 185
 186         for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
 187                 attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
 188                 if (IS_ERR(attr))
 189                         continue;
 190
 191                 rcu_read_lock();
 192                 ndev = rdma_read_gid_attr_ndev_rcu(attr);
 193                 if (!IS_ERR(ndev) &&
 194                     ((!vlan_id && !is_vlan_dev(attr->ndev)) ||
 195                      (vlan_id && is_vlan_dev(attr->ndev) &&
 196                       vlan_dev_vlan_id(attr->ndev) == vlan_id)) &&
 197                     attr->gid_type == IB_GID_TYPE_ROCE) {
 198                         rcu_read_unlock();
 199                         if (gid)
 200                                 memcpy(gid, &attr->gid, SMC_GID_SIZE);
 201                         if (sgid_index)
 202                                 *sgid_index = attr->index;
 203                         rdma_put_gid_attr(attr);
 204                         return 0;
 205                 }
 206                 rcu_read_unlock();
 207                 rdma_put_gid_attr(attr);
 208         }
 209         return -ENODEV;
 210 }
 211
 212 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
 213 {
 214         int rc;
 215
 216         memset(&smcibdev->pattr[ibport - 1], 0,
 217                sizeof(smcibdev->pattr[ibport - 1]));
 218         rc = ib_query_port(smcibdev->ibdev, ibport,
 219                            &smcibdev->pattr[ibport - 1]);
 220         if (rc)
 221                 goto out;
 222         /* the SMC protocol requires specification of the RoCE MAC address */
 223         rc = smc_ib_fill_mac(smcibdev, ibport);
 224         if (rc)
 225                 goto out;
 226         if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET,
 227                      sizeof(local_systemid)) &&
 228             smc_ib_port_active(smcibdev, ibport))
 229                 /* create unique system identifier */
 230                 smc_ib_define_local_systemid(smcibdev, ibport);
 231 out:
 232         return rc;
 233 }
 234
 235 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
 236 static void smc_ib_port_event_work(struct work_struct *work)
 237 {
 238         struct smc_ib_device *smcibdev = container_of(
 239                 work, struct smc_ib_device, port_event_work);
 240         u8 port_idx;
 241
 242         for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
 243                 smc_ib_remember_port_attr(smcibdev, port_idx + 1);
 244                 clear_bit(port_idx, &smcibdev->port_event_mask);
 245                 if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
 246                         set_bit(port_idx, smcibdev->ports_going_away);
 247                         smc_port_terminate(smcibdev, port_idx + 1);
 248                 } else {
 249                         clear_bit(port_idx, smcibdev->ports_going_away);
 250                 }
 251         }
 252 }
 253
 254 /* can be called in IRQ context */
 255 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
 256                                         struct ib_event *ibevent)
 257 {
 258         struct smc_ib_device *smcibdev;
 259         u8 port_idx;
 260
 261         smcibdev = container_of(handler, struct smc_ib_device, event_handler);
 262
 263         switch (ibevent->event) {
 264         case IB_EVENT_DEVICE_FATAL:
 265                 /* terminate all ports on device */
 266                 for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
 267                         set_bit(port_idx, &smcibdev->port_event_mask);
 268                         set_bit(port_idx, smcibdev->ports_going_away);
 269                 }
 270                 schedule_work(&smcibdev->port_event_work);
 271                 break;
 272         case IB_EVENT_PORT_ERR:
 273         case IB_EVENT_PORT_ACTIVE:
 274         case IB_EVENT_GID_CHANGE:
 275                 port_idx = ibevent->element.port_num - 1;
 276                 if (port_idx < SMC_MAX_PORTS) {
 277                         set_bit(port_idx, &smcibdev->port_event_mask);
 278                         if (ibevent->event == IB_EVENT_PORT_ERR)
 279                                 set_bit(port_idx, smcibdev->ports_going_away);
 280                         else if (ibevent->event == IB_EVENT_PORT_ACTIVE)
 281                                 clear_bit(port_idx, smcibdev->ports_going_away);
 282                         schedule_work(&smcibdev->port_event_work);
 283                 }
 284                 break;
 285         default:
 286                 break;
 287         }
 288 }
 289
 290 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
 291 {
 292         if (lnk->roce_pd)
 293                 ib_dealloc_pd(lnk->roce_pd);
 294         lnk->roce_pd = NULL;
 295 }
 296
 297 int smc_ib_create_protection_domain(struct smc_link *lnk)
 298 {
 299         int rc;
 300
 301         lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
 302         rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
 303         if (IS_ERR(lnk->roce_pd))
 304                 lnk->roce_pd = NULL;
 305         return rc;
 306 }
 307
 308 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
 309 {
 310         struct smc_link *lnk = (struct smc_link *)priv;
 311         struct smc_ib_device *smcibdev = lnk->smcibdev;
 312         u8 port_idx;
 313
 314         switch (ibevent->event) {
 315         case IB_EVENT_QP_FATAL:
 316         case IB_EVENT_QP_ACCESS_ERR:
 317                 port_idx = ibevent->element.qp->port - 1;
 318                 if (port_idx < SMC_MAX_PORTS) {
 319                         set_bit(port_idx, &smcibdev->port_event_mask);
 320                         set_bit(port_idx, smcibdev->ports_going_away);
 321                         schedule_work(&smcibdev->port_event_work);
 322                 }
 323                 break;
 324         default:
 325                 break;
 326         }
 327 }
 328
 329 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
 330 {
 331         if (lnk->roce_qp)
 332                 ib_destroy_qp(lnk->roce_qp);
 333         lnk->roce_qp = NULL;
 334 }
 335
 336 /* create a queue pair within the protection domain for a link */
 337 int smc_ib_create_queue_pair(struct smc_link *lnk)
 338 {
 339         struct ib_qp_init_attr qp_attr = {
 340                 .event_handler = smc_ib_qp_event_handler,
 341                 .qp_context = lnk,
 342                 .send_cq = lnk->smcibdev->roce_cq_send,
 343                 .recv_cq = lnk->smcibdev->roce_cq_recv,
 344                 .srq = NULL,
 345                 .cap = {
 346                                 /* include unsolicited rdma_writes as well,
 347                                  * there are max. 2 RDMA_WRITE per 1 WR_SEND
 348                                  */
 349                         .max_send_wr = SMC_WR_BUF_CNT * 3,
 350                         .max_recv_wr = SMC_WR_BUF_CNT * 3,
 351                         .max_send_sge = SMC_IB_MAX_SEND_SGE,
 352                         .max_recv_sge = 1,
 353                 },
 354                 .sq_sig_type = IB_SIGNAL_REQ_WR,
 355                 .qp_type = IB_QPT_RC,
 356         };
 357         int rc;
 358
 359         lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
 360         rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
 361         if (IS_ERR(lnk->roce_qp))
 362                 lnk->roce_qp = NULL;
 363         else
 364                 smc_wr_remember_qp_attr(lnk);
 365         return rc;
 366 }
 367
 368 void smc_ib_put_memory_region(struct ib_mr *mr)
 369 {
 370         ib_dereg_mr(mr);
 371 }
 372
 373 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot)
 374 {
 375         unsigned int offset = 0;
 376         int sg_num;
 377
 378         /* map the largest prefix of a dma mapped SG list */
 379         sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK],
 380                               buf_slot->sgt[SMC_SINGLE_LINK].sgl,
 381                               buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
 382                               &offset, PAGE_SIZE);
 383
 384         return sg_num;
 385 }
 386
 387 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
 388 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
 389                              struct smc_buf_desc *buf_slot)
 390 {
 391         if (buf_slot->mr_rx[SMC_SINGLE_LINK])
 392                 return 0; /* already done */
 393
 394         buf_slot->mr_rx[SMC_SINGLE_LINK] =
 395                 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
 396         if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) {
 397                 int rc;
 398
 399                 rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]);
 400                 buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL;
 401                 return rc;
 402         }
 403
 404         if (smc_ib_map_mr_sg(buf_slot) != 1)
 405                 return -EINVAL;
 406
 407         return 0;
 408 }
 409
 410 /* synchronize buffer usage for cpu access */
 411 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev,
 412                             struct smc_buf_desc *buf_slot,
 413                             enum dma_data_direction data_direction)
 414 {
 415         struct scatterlist *sg;
 416         unsigned int i;
 417
 418         /* for now there is just one DMA address */
 419         for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
 420                     buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
 421                 if (!sg_dma_len(sg))
 422                         break;
 423                 ib_dma_sync_single_for_cpu(smcibdev->ibdev,
 424                                            sg_dma_address(sg),
 425                                            sg_dma_len(sg),
 426                                            data_direction);
 427         }
 428 }
 429
 430 /* synchronize buffer usage for device access */
 431 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev,
 432                                struct smc_buf_desc *buf_slot,
 433                                enum dma_data_direction data_direction)
 434 {
 435         struct scatterlist *sg;
 436         unsigned int i;
 437
 438         /* for now there is just one DMA address */
 439         for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
 440                     buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
 441                 if (!sg_dma_len(sg))
 442                         break;
 443                 ib_dma_sync_single_for_device(smcibdev->ibdev,
 444                                               sg_dma_address(sg),
 445                                               sg_dma_len(sg),
 446                                               data_direction);
 447         }
 448 }
 449
 450 /* Map a new TX or RX buffer SG-table to DMA */
 451 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev,
 452                       struct smc_buf_desc *buf_slot,
 453                       enum dma_data_direction data_direction)
 454 {
 455         int mapped_nents;
 456
 457         mapped_nents = ib_dma_map_sg(smcibdev->ibdev,
 458                                      buf_slot->sgt[SMC_SINGLE_LINK].sgl,
 459                                      buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
 460                                      data_direction);
 461         if (!mapped_nents)
 462                 return -ENOMEM;
 463
 464         return mapped_nents;
 465 }
 466
 467 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev,
 468                          struct smc_buf_desc *buf_slot,
 469                          enum dma_data_direction data_direction)
 470 {
 471         if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address)
 472                 return; /* already unmapped */
 473
 474         ib_dma_unmap_sg(smcibdev->ibdev,
 475                         buf_slot->sgt[SMC_SINGLE_LINK].sgl,
 476                         buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
 477                         data_direction);
 478         buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0;
 479 }
 480
 481 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
 482 {
 483         struct ib_cq_init_attr cqattr = {
 484                 .cqe = SMC_MAX_CQE, .comp_vector = 0 };
 485         int cqe_size_order, smc_order;
 486         long rc;
 487
 488         /* the calculated number of cq entries fits to mlx5 cq allocation */
 489         cqe_size_order = cache_line_size() == 128 ? 7 : 6;
 490         smc_order = MAX_ORDER - cqe_size_order - 1;
 491         if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
 492                 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
 493         smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
 494                                               smc_wr_tx_cq_handler, NULL,
 495                                               smcibdev, &cqattr);
 496         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
 497         if (IS_ERR(smcibdev->roce_cq_send)) {
 498                 smcibdev->roce_cq_send = NULL;
 499                 return rc;
 500         }
 501         smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
 502                                               smc_wr_rx_cq_handler, NULL,
 503                                               smcibdev, &cqattr);
 504         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
 505         if (IS_ERR(smcibdev->roce_cq_recv)) {
 506                 smcibdev->roce_cq_recv = NULL;
 507                 goto err;
 508         }
 509         smc_wr_add_dev(smcibdev);
 510         smcibdev->initialized = 1;
 511         return rc;
 512
 513 err:
 514         ib_destroy_cq(smcibdev->roce_cq_send);
 515         return rc;
 516 }
 517
 518 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
 519 {
 520         if (!smcibdev->initialized)
 521                 return;
 522         smcibdev->initialized = 0;
 523         smc_wr_remove_dev(smcibdev);
 524         ib_destroy_cq(smcibdev->roce_cq_recv);
 525         ib_destroy_cq(smcibdev->roce_cq_send);
 526 }
 527
 528 static struct ib_client smc_ib_client;
 529
 530 /* callback function for ib_register_client() */
 531 static void smc_ib_add_dev(struct ib_device *ibdev)
 532 {
 533         struct smc_ib_device *smcibdev;
 534         u8 port_cnt;
 535         int i;
 536
 537         if (ibdev->node_type != RDMA_NODE_IB_CA)
 538                 return;
 539
 540         smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
 541         if (!smcibdev)
 542                 return;
 543
 544         smcibdev->ibdev = ibdev;
 545         INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
 546
 547         spin_lock(&smc_ib_devices.lock);
 548         list_add_tail(&smcibdev->list, &smc_ib_devices.list);
 549         spin_unlock(&smc_ib_devices.lock);
 550         ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
 551         INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
 552                               smc_ib_global_event_handler);
 553         ib_register_event_handler(&smcibdev->event_handler);
 554
 555         /* trigger reading of the port attributes */
 556         port_cnt = smcibdev->ibdev->phys_port_cnt;
 557         for (i = 0;
 558              i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
 559              i++) {
 560                 set_bit(i, &smcibdev->port_event_mask);
 561                 /* determine pnetids of the port */
 562                 smc_pnetid_by_dev_port(ibdev->dev.parent, i,
 563                                        smcibdev->pnetid[i]);
 564         }
 565         schedule_work(&smcibdev->port_event_work);
 566 }
 567
 568 /* callback function for ib_register_client() */
 569 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
 570 {
 571         struct smc_ib_device *smcibdev;
 572
 573         smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
 574         ib_set_client_data(ibdev, &smc_ib_client, NULL);
 575         spin_lock(&smc_ib_devices.lock);
 576         list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
 577         spin_unlock(&smc_ib_devices.lock);
 578         smc_ib_cleanup_per_ibdev(smcibdev);
 579         ib_unregister_event_handler(&smcibdev->event_handler);
 580         kfree(smcibdev);
 581 }
 582
 583 static struct ib_client smc_ib_client = {
 584         .name   = "smc_ib",
 585         .add    = smc_ib_add_dev,
 586         .remove = smc_ib_remove_dev,
 587 };
 588
 589 int __init smc_ib_register_client(void)
 590 {
 591         return ib_register_client(&smc_ib_client);
 592 }
 593
 594 void smc_ib_unregister_client(void)
 595 {
 596         ib_unregister_client(&smc_ib_client);
 597 }