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 <linux/wait.h>
  19 #include <rdma/ib_verbs.h>
  20 #include <rdma/ib_cache.h>
  21
  22 #include "smc_pnet.h"
  23 #include "smc_ib.h"
  24 #include "smc_core.h"
  25 #include "smc_wr.h"
  26 #include "smc.h"
  27
  28 #define SMC_MAX_CQE 32766       /* max. # of completion queue elements */
  29
  30 #define SMC_QP_MIN_RNR_TIMER            5
  31 #define SMC_QP_TIMEOUT                  15 /* 4096 * 2 ** timeout usec */
  32 #define SMC_QP_RETRY_CNT                        7 /* 7: infinite */
  33 #define SMC_QP_RNR_RETRY                        7 /* 7: infinite */
  34
  35 struct smc_ib_devices smc_ib_devices = {        /* smc-registered ib devices */
  36         .lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock),
  37         .list = LIST_HEAD_INIT(smc_ib_devices.list),
  38 };
  39
  40 #define SMC_LOCAL_SYSTEMID_RESET        "%%%%%%%"
  41
  42 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET; /* unique system
  43                                                                  * identifier
  44                                                                  */
  45
  46 static int smc_ib_modify_qp_init(struct smc_link *lnk)
  47 {
  48         struct ib_qp_attr qp_attr;
  49
  50         memset(&qp_attr, 0, sizeof(qp_attr));
  51         qp_attr.qp_state = IB_QPS_INIT;
  52         qp_attr.pkey_index = 0;
  53         qp_attr.port_num = lnk->ibport;
  54         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
  55                                 | IB_ACCESS_REMOTE_WRITE;
  56         return ib_modify_qp(lnk->roce_qp, &qp_attr,
  57                             IB_QP_STATE | IB_QP_PKEY_INDEX |
  58                             IB_QP_ACCESS_FLAGS | IB_QP_PORT);
  59 }
  60
  61 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
  62 {
  63         enum ib_qp_attr_mask qp_attr_mask =
  64                 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
  65                 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
  66         struct ib_qp_attr qp_attr;
  67
  68         memset(&qp_attr, 0, sizeof(qp_attr));
  69         qp_attr.qp_state = IB_QPS_RTR;
  70         qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
  71         qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
  72         rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
  73         rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, 1, 0);
  74         rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
  75         memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
  76                sizeof(lnk->peer_mac));
  77         qp_attr.dest_qp_num = lnk->peer_qpn;
  78         qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
  79         qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
  80                                          * requests
  81                                          */
  82         qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
  83
  84         return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
  85 }
  86
  87 int smc_ib_modify_qp_rts(struct smc_link *lnk)
  88 {
  89         struct ib_qp_attr qp_attr;
  90
  91         memset(&qp_attr, 0, sizeof(qp_attr));
  92         qp_attr.qp_state = IB_QPS_RTS;
  93         qp_attr.timeout = SMC_QP_TIMEOUT;       /* local ack timeout */
  94         qp_attr.retry_cnt = SMC_QP_RETRY_CNT;   /* retry count */
  95         qp_attr.rnr_retry = SMC_QP_RNR_RETRY;   /* RNR retries, 7=infinite */
  96         qp_attr.sq_psn = lnk->psn_initial;      /* starting send packet seq # */
  97         qp_attr.max_rd_atomic = 1;      /* # of outstanding RDMA reads and
  98                                          * atomic ops allowed
  99                                          */
 100         return ib_modify_qp(lnk->roce_qp, &qp_attr,
 101                             IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
 102                             IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
 103                             IB_QP_MAX_QP_RD_ATOMIC);
 104 }
 105
 106 int smc_ib_modify_qp_reset(struct smc_link *lnk)
 107 {
 108         struct ib_qp_attr qp_attr;
 109
 110         memset(&qp_attr, 0, sizeof(qp_attr));
 111         qp_attr.qp_state = IB_QPS_RESET;
 112         return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
 113 }
 114
 115 int smc_ib_ready_link(struct smc_link *lnk)
 116 {
 117         struct smc_link_group *lgr = smc_get_lgr(lnk);
 118         int rc = 0;
 119
 120         rc = smc_ib_modify_qp_init(lnk);
 121         if (rc)
 122                 goto out;
 123
 124         rc = smc_ib_modify_qp_rtr(lnk);
 125         if (rc)
 126                 goto out;
 127         smc_wr_remember_qp_attr(lnk);
 128         rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
 129                               IB_CQ_SOLICITED_MASK);
 130         if (rc)
 131                 goto out;
 132         rc = smc_wr_rx_post_init(lnk);
 133         if (rc)
 134                 goto out;
 135         smc_wr_remember_qp_attr(lnk);
 136
 137         if (lgr->role == SMC_SERV) {
 138                 rc = smc_ib_modify_qp_rts(lnk);
 139                 if (rc)
 140                         goto out;
 141                 smc_wr_remember_qp_attr(lnk);
 142         }
 143 out:
 144         return rc;
 145 }
 146
 147 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
 148 {
 149         const struct ib_gid_attr *attr;
 150         int rc;
 151
 152         attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
 153         if (IS_ERR(attr))
 154                 return -ENODEV;
 155
 156         rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]);
 157         rdma_put_gid_attr(attr);
 158         return rc;
 159 }
 160
 161 /* Create an identifier unique for this instance of SMC-R.
 162  * The MAC-address of the first active registered IB device
 163  * plus a random 2-byte number is used to create this identifier.
 164  * This name is delivered to the peer during connection initialization.
 165  */
 166 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
 167                                                 u8 ibport)
 168 {
 169         memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
 170                sizeof(smcibdev->mac[ibport - 1]));
 171         get_random_bytes(&local_systemid[0], 2);
 172 }
 173
 174 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
 175 {
 176         return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
 177 }
 178
 179 /* determine the gid for an ib-device port and vlan id */
 180 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
 181                          unsigned short vlan_id, u8 gid[], u8 *sgid_index)
 182 {
 183         const struct ib_gid_attr *attr;
 184         const struct net_device *ndev;
 185         int i;
 186
 187         for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
 188                 attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
 189                 if (IS_ERR(attr))
 190                         continue;
 191
 192                 rcu_read_lock();
 193                 ndev = rdma_read_gid_attr_ndev_rcu(attr);
 194                 if (!IS_ERR(ndev) &&
 195                     ((!vlan_id && !is_vlan_dev(attr->ndev)) ||
 196                      (vlan_id && is_vlan_dev(attr->ndev) &&
 197                       vlan_dev_vlan_id(attr->ndev) == vlan_id)) &&
 198                     attr->gid_type == IB_GID_TYPE_ROCE) {
 199                         rcu_read_unlock();
 200                         if (gid)
 201                                 memcpy(gid, &attr->gid, SMC_GID_SIZE);
 202                         if (sgid_index)
 203                                 *sgid_index = attr->index;
 204                         rdma_put_gid_attr(attr);
 205                         return 0;
 206                 }
 207                 rcu_read_unlock();
 208                 rdma_put_gid_attr(attr);
 209         }
 210         return -ENODEV;
 211 }
 212
 213 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
 214 {
 215         int rc;
 216
 217         memset(&smcibdev->pattr[ibport - 1], 0,
 218                sizeof(smcibdev->pattr[ibport - 1]));
 219         rc = ib_query_port(smcibdev->ibdev, ibport,
 220                            &smcibdev->pattr[ibport - 1]);
 221         if (rc)
 222                 goto out;
 223         /* the SMC protocol requires specification of the RoCE MAC address */
 224         rc = smc_ib_fill_mac(smcibdev, ibport);
 225         if (rc)
 226                 goto out;
 227         if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET,
 228                      sizeof(local_systemid)) &&
 229             smc_ib_port_active(smcibdev, ibport))
 230                 /* create unique system identifier */
 231                 smc_ib_define_local_systemid(smcibdev, ibport);
 232 out:
 233         return rc;
 234 }
 235
 236 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
 237 static void smc_ib_port_event_work(struct work_struct *work)
 238 {
 239         struct smc_ib_device *smcibdev = container_of(
 240                 work, struct smc_ib_device, port_event_work);
 241         u8 port_idx;
 242
 243         for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
 244                 smc_ib_remember_port_attr(smcibdev, port_idx + 1);
 245                 clear_bit(port_idx, &smcibdev->port_event_mask);
 246                 if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
 247                         set_bit(port_idx, smcibdev->ports_going_away);
 248                         smc_port_terminate(smcibdev, port_idx + 1);
 249                 } else {
 250                         clear_bit(port_idx, smcibdev->ports_going_away);
 251                 }
 252         }
 253 }
 254
 255 /* can be called in IRQ context */
 256 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
 257                                         struct ib_event *ibevent)
 258 {
 259         struct smc_ib_device *smcibdev;
 260         u8 port_idx;
 261
 262         smcibdev = container_of(handler, struct smc_ib_device, event_handler);
 263
 264         switch (ibevent->event) {
 265         case IB_EVENT_DEVICE_FATAL:
 266                 /* terminate all ports on device */
 267                 for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
 268                         set_bit(port_idx, &smcibdev->port_event_mask);
 269                         set_bit(port_idx, smcibdev->ports_going_away);
 270                 }
 271                 schedule_work(&smcibdev->port_event_work);
 272                 break;
 273         case IB_EVENT_PORT_ERR:
 274         case IB_EVENT_PORT_ACTIVE:
 275         case IB_EVENT_GID_CHANGE:
 276                 port_idx = ibevent->element.port_num - 1;
 277                 if (port_idx < SMC_MAX_PORTS) {
 278                         set_bit(port_idx, &smcibdev->port_event_mask);
 279                         if (ibevent->event == IB_EVENT_PORT_ERR)
 280                                 set_bit(port_idx, smcibdev->ports_going_away);
 281                         else if (ibevent->event == IB_EVENT_PORT_ACTIVE)
 282                                 clear_bit(port_idx, smcibdev->ports_going_away);
 283                         schedule_work(&smcibdev->port_event_work);
 284                 }
 285                 break;
 286         default:
 287                 break;
 288         }
 289 }
 290
 291 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
 292 {
 293         if (lnk->roce_pd)
 294                 ib_dealloc_pd(lnk->roce_pd);
 295         lnk->roce_pd = NULL;
 296 }
 297
 298 int smc_ib_create_protection_domain(struct smc_link *lnk)
 299 {
 300         int rc;
 301
 302         lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
 303         rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
 304         if (IS_ERR(lnk->roce_pd))
 305                 lnk->roce_pd = NULL;
 306         return rc;
 307 }
 308
 309 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
 310 {
 311         struct smc_link *lnk = (struct smc_link *)priv;
 312         struct smc_ib_device *smcibdev = lnk->smcibdev;
 313         u8 port_idx;
 314
 315         switch (ibevent->event) {
 316         case IB_EVENT_QP_FATAL:
 317         case IB_EVENT_QP_ACCESS_ERR:
 318                 port_idx = ibevent->element.qp->port - 1;
 319                 if (port_idx < SMC_MAX_PORTS) {
 320                         set_bit(port_idx, &smcibdev->port_event_mask);
 321                         set_bit(port_idx, smcibdev->ports_going_away);
 322                         schedule_work(&smcibdev->port_event_work);
 323                 }
 324                 break;
 325         default:
 326                 break;
 327         }
 328 }
 329
 330 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
 331 {
 332         if (lnk->roce_qp)
 333                 ib_destroy_qp(lnk->roce_qp);
 334         lnk->roce_qp = NULL;
 335 }
 336
 337 /* create a queue pair within the protection domain for a link */
 338 int smc_ib_create_queue_pair(struct smc_link *lnk)
 339 {
 340         struct ib_qp_init_attr qp_attr = {
 341                 .event_handler = smc_ib_qp_event_handler,
 342                 .qp_context = lnk,
 343                 .send_cq = lnk->smcibdev->roce_cq_send,
 344                 .recv_cq = lnk->smcibdev->roce_cq_recv,
 345                 .srq = NULL,
 346                 .cap = {
 347                                 /* include unsolicited rdma_writes as well,
 348                                  * there are max. 2 RDMA_WRITE per 1 WR_SEND
 349                                  */
 350                         .max_send_wr = SMC_WR_BUF_CNT * 3,
 351                         .max_recv_wr = SMC_WR_BUF_CNT * 3,
 352                         .max_send_sge = SMC_IB_MAX_SEND_SGE,
 353                         .max_recv_sge = 1,
 354                 },
 355                 .sq_sig_type = IB_SIGNAL_REQ_WR,
 356                 .qp_type = IB_QPT_RC,
 357         };
 358         int rc;
 359
 360         lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
 361         rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
 362         if (IS_ERR(lnk->roce_qp))
 363                 lnk->roce_qp = NULL;
 364         else
 365                 smc_wr_remember_qp_attr(lnk);
 366         return rc;
 367 }
 368
 369 void smc_ib_put_memory_region(struct ib_mr *mr)
 370 {
 371         ib_dereg_mr(mr);
 372 }
 373
 374 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot)
 375 {
 376         unsigned int offset = 0;
 377         int sg_num;
 378
 379         /* map the largest prefix of a dma mapped SG list */
 380         sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK],
 381                               buf_slot->sgt[SMC_SINGLE_LINK].sgl,
 382                               buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
 383                               &offset, PAGE_SIZE);
 384
 385         return sg_num;
 386 }
 387
 388 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
 389 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
 390                              struct smc_buf_desc *buf_slot)
 391 {
 392         if (buf_slot->mr_rx[SMC_SINGLE_LINK])
 393                 return 0; /* already done */
 394
 395         buf_slot->mr_rx[SMC_SINGLE_LINK] =
 396                 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
 397         if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) {
 398                 int rc;
 399
 400                 rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]);
 401                 buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL;
 402                 return rc;
 403         }
 404
 405         if (smc_ib_map_mr_sg(buf_slot) != 1)
 406                 return -EINVAL;
 407
 408         return 0;
 409 }
 410
 411 /* synchronize buffer usage for cpu access */
 412 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev,
 413                             struct smc_buf_desc *buf_slot,
 414                             enum dma_data_direction data_direction)
 415 {
 416         struct scatterlist *sg;
 417         unsigned int i;
 418
 419         /* for now there is just one DMA address */
 420         for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
 421                     buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
 422                 if (!sg_dma_len(sg))
 423                         break;
 424                 ib_dma_sync_single_for_cpu(smcibdev->ibdev,
 425                                            sg_dma_address(sg),
 426                                            sg_dma_len(sg),
 427                                            data_direction);
 428         }
 429 }
 430
 431 /* synchronize buffer usage for device access */
 432 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev,
 433                                struct smc_buf_desc *buf_slot,
 434                                enum dma_data_direction data_direction)
 435 {
 436         struct scatterlist *sg;
 437         unsigned int i;
 438
 439         /* for now there is just one DMA address */
 440         for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg,
 441                     buf_slot->sgt[SMC_SINGLE_LINK].nents, i) {
 442                 if (!sg_dma_len(sg))
 443                         break;
 444                 ib_dma_sync_single_for_device(smcibdev->ibdev,
 445                                               sg_dma_address(sg),
 446                                               sg_dma_len(sg),
 447                                               data_direction);
 448         }
 449 }
 450
 451 /* Map a new TX or RX buffer SG-table to DMA */
 452 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev,
 453                       struct smc_buf_desc *buf_slot,
 454                       enum dma_data_direction data_direction)
 455 {
 456         int mapped_nents;
 457
 458         mapped_nents = ib_dma_map_sg(smcibdev->ibdev,
 459                                      buf_slot->sgt[SMC_SINGLE_LINK].sgl,
 460                                      buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
 461                                      data_direction);
 462         if (!mapped_nents)
 463                 return -ENOMEM;
 464
 465         return mapped_nents;
 466 }
 467
 468 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev,
 469                          struct smc_buf_desc *buf_slot,
 470                          enum dma_data_direction data_direction)
 471 {
 472         if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address)
 473                 return; /* already unmapped */
 474
 475         ib_dma_unmap_sg(smcibdev->ibdev,
 476                         buf_slot->sgt[SMC_SINGLE_LINK].sgl,
 477                         buf_slot->sgt[SMC_SINGLE_LINK].orig_nents,
 478                         data_direction);
 479         buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0;
 480 }
 481
 482 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
 483 {
 484         struct ib_cq_init_attr cqattr = {
 485                 .cqe = SMC_MAX_CQE, .comp_vector = 0 };
 486         int cqe_size_order, smc_order;
 487         long rc;
 488
 489         /* the calculated number of cq entries fits to mlx5 cq allocation */
 490         cqe_size_order = cache_line_size() == 128 ? 7 : 6;
 491         smc_order = MAX_ORDER - cqe_size_order - 1;
 492         if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
 493                 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
 494         smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
 495                                               smc_wr_tx_cq_handler, NULL,
 496                                               smcibdev, &cqattr);
 497         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
 498         if (IS_ERR(smcibdev->roce_cq_send)) {
 499                 smcibdev->roce_cq_send = NULL;
 500                 return rc;
 501         }
 502         smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
 503                                               smc_wr_rx_cq_handler, NULL,
 504                                               smcibdev, &cqattr);
 505         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
 506         if (IS_ERR(smcibdev->roce_cq_recv)) {
 507                 smcibdev->roce_cq_recv = NULL;
 508                 goto err;
 509         }
 510         smc_wr_add_dev(smcibdev);
 511         smcibdev->initialized = 1;
 512         return rc;
 513
 514 err:
 515         ib_destroy_cq(smcibdev->roce_cq_send);
 516         return rc;
 517 }
 518
 519 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
 520 {
 521         if (!smcibdev->initialized)
 522                 return;
 523         smcibdev->initialized = 0;
 524         ib_destroy_cq(smcibdev->roce_cq_recv);
 525         ib_destroy_cq(smcibdev->roce_cq_send);
 526         smc_wr_remove_dev(smcibdev);
 527 }
 528
 529 static struct ib_client smc_ib_client;
 530
 531 /* callback function for ib_register_client() */
 532 static void smc_ib_add_dev(struct ib_device *ibdev)
 533 {
 534         struct smc_ib_device *smcibdev;
 535         u8 port_cnt;
 536         int i;
 537
 538         if (ibdev->node_type != RDMA_NODE_IB_CA)
 539                 return;
 540
 541         smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
 542         if (!smcibdev)
 543                 return;
 544
 545         smcibdev->ibdev = ibdev;
 546         INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
 547         atomic_set(&smcibdev->lnk_cnt, 0);
 548         init_waitqueue_head(&smcibdev->lnks_deleted);
 549         spin_lock(&smc_ib_devices.lock);
 550         list_add_tail(&smcibdev->list, &smc_ib_devices.list);
 551         spin_unlock(&smc_ib_devices.lock);
 552         ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
 553         INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
 554                               smc_ib_global_event_handler);
 555         ib_register_event_handler(&smcibdev->event_handler);
 556
 557         /* trigger reading of the port attributes */
 558         port_cnt = smcibdev->ibdev->phys_port_cnt;
 559         for (i = 0;
 560              i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
 561              i++) {
 562                 set_bit(i, &smcibdev->port_event_mask);
 563                 /* determine pnetids of the port */
 564                 smc_pnetid_by_dev_port(ibdev->dev.parent, i,
 565                                        smcibdev->pnetid[i]);
 566         }
 567         schedule_work(&smcibdev->port_event_work);
 568 }
 569
 570 /* callback function for ib_unregister_client() */
 571 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
 572 {
 573         struct smc_ib_device *smcibdev;
 574
 575         smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
 576         ib_set_client_data(ibdev, &smc_ib_client, NULL);
 577         spin_lock(&smc_ib_devices.lock);
 578         list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
 579         spin_unlock(&smc_ib_devices.lock);
 580         smc_smcr_terminate_all(smcibdev);
 581         smc_ib_cleanup_per_ibdev(smcibdev);
 582         ib_unregister_event_handler(&smcibdev->event_handler);
 583         kfree(smcibdev);
 584 }
 585
 586 static struct ib_client smc_ib_client = {
 587         .name   = "smc_ib",
 588         .add    = smc_ib_add_dev,
 589         .remove = smc_ib_remove_dev,
 590 };
 591
 592 int __init smc_ib_register_client(void)
 593 {
 594         return ib_register_client(&smc_ib_client);
 595 }
 596
 597 void smc_ib_unregister_client(void)
 598 {
 599         ib_unregister_client(&smc_ib_client);
 600 }