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 <ubraun@linux.vnet.ibm.com>
  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 u8 local_systemid[SMC_SYSTEMID_LEN];            /* unique system identifier */
  41
  42 static int smc_ib_modify_qp_init(struct smc_link *lnk)
  43 {
  44         struct ib_qp_attr qp_attr;
  45
  46         memset(&qp_attr, 0, sizeof(qp_attr));
  47         qp_attr.qp_state = IB_QPS_INIT;
  48         qp_attr.pkey_index = 0;
  49         qp_attr.port_num = lnk->ibport;
  50         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
  51                                 | IB_ACCESS_REMOTE_WRITE;
  52         return ib_modify_qp(lnk->roce_qp, &qp_attr,
  53                             IB_QP_STATE | IB_QP_PKEY_INDEX |
  54                             IB_QP_ACCESS_FLAGS | IB_QP_PORT);
  55 }
  56
  57 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
  58 {
  59         enum ib_qp_attr_mask qp_attr_mask =
  60                 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
  61                 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
  62         struct ib_qp_attr qp_attr;
  63
  64         memset(&qp_attr, 0, sizeof(qp_attr));
  65         qp_attr.qp_state = IB_QPS_RTR;
  66         qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
  67         qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
  68         rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
  69         rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, 1, 0);
  70         rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
  71         memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
  72                sizeof(lnk->peer_mac));
  73         qp_attr.dest_qp_num = lnk->peer_qpn;
  74         qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
  75         qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
  76                                          * requests
  77                                          */
  78         qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
  79
  80         return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
  81 }
  82
  83 int smc_ib_modify_qp_rts(struct smc_link *lnk)
  84 {
  85         struct ib_qp_attr qp_attr;
  86
  87         memset(&qp_attr, 0, sizeof(qp_attr));
  88         qp_attr.qp_state = IB_QPS_RTS;
  89         qp_attr.timeout = SMC_QP_TIMEOUT;       /* local ack timeout */
  90         qp_attr.retry_cnt = SMC_QP_RETRY_CNT;   /* retry count */
  91         qp_attr.rnr_retry = SMC_QP_RNR_RETRY;   /* RNR retries, 7=infinite */
  92         qp_attr.sq_psn = lnk->psn_initial;      /* starting send packet seq # */
  93         qp_attr.max_rd_atomic = 1;      /* # of outstanding RDMA reads and
  94                                          * atomic ops allowed
  95                                          */
  96         return ib_modify_qp(lnk->roce_qp, &qp_attr,
  97                             IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
  98                             IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
  99                             IB_QP_MAX_QP_RD_ATOMIC);
 100 }
 101
 102 int smc_ib_modify_qp_reset(struct smc_link *lnk)
 103 {
 104         struct ib_qp_attr qp_attr;
 105
 106         memset(&qp_attr, 0, sizeof(qp_attr));
 107         qp_attr.qp_state = IB_QPS_RESET;
 108         return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
 109 }
 110
 111 int smc_ib_ready_link(struct smc_link *lnk)
 112 {
 113         struct smc_link_group *lgr = smc_get_lgr(lnk);
 114         int rc = 0;
 115
 116         rc = smc_ib_modify_qp_init(lnk);
 117         if (rc)
 118                 goto out;
 119
 120         rc = smc_ib_modify_qp_rtr(lnk);
 121         if (rc)
 122                 goto out;
 123         smc_wr_remember_qp_attr(lnk);
 124         rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
 125                               IB_CQ_SOLICITED_MASK);
 126         if (rc)
 127                 goto out;
 128         rc = smc_wr_rx_post_init(lnk);
 129         if (rc)
 130                 goto out;
 131         smc_wr_remember_qp_attr(lnk);
 132
 133         if (lgr->role == SMC_SERV) {
 134                 rc = smc_ib_modify_qp_rts(lnk);
 135                 if (rc)
 136                         goto out;
 137                 smc_wr_remember_qp_attr(lnk);
 138         }
 139 out:
 140         return rc;
 141 }
 142
 143 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
 144 {
 145         const struct ib_gid_attr *attr;
 146         int rc;
 147
 148         attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
 149         if (IS_ERR(attr))
 150                 return -ENODEV;
 151
 152         rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]);
 153         rdma_put_gid_attr(attr);
 154         return rc;
 155 }
 156
 157 /* Create an identifier unique for this instance of SMC-R.
 158  * The MAC-address of the first active registered IB device
 159  * plus a random 2-byte number is used to create this identifier.
 160  * This name is delivered to the peer during connection initialization.
 161  */
 162 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
 163                                                 u8 ibport)
 164 {
 165         memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
 166                sizeof(smcibdev->mac[ibport - 1]));
 167 }
 168
 169 bool smc_ib_is_valid_local_systemid(void)
 170 {
 171         return !is_zero_ether_addr(&local_systemid[2]);
 172 }
 173
 174 static void smc_ib_init_local_systemid(void)
 175 {
 176         get_random_bytes(&local_systemid[0], 2);
 177 }
 178
 179 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
 180 {
 181         return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
 182 }
 183
 184 /* determine the gid for an ib-device port and vlan id */
 185 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
 186                          unsigned short vlan_id, u8 gid[], u8 *sgid_index)
 187 {
 188         const struct ib_gid_attr *attr;
 189         const struct net_device *ndev;
 190         int i;
 191
 192         for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
 193                 attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
 194                 if (IS_ERR(attr))
 195                         continue;
 196
 197                 rcu_read_lock();
 198                 ndev = rdma_read_gid_attr_ndev_rcu(attr);
 199                 if (!IS_ERR(ndev) &&
 200                     ((!vlan_id && !is_vlan_dev(attr->ndev)) ||
 201                      (vlan_id && is_vlan_dev(attr->ndev) &&
 202                       vlan_dev_vlan_id(attr->ndev) == vlan_id)) &&
 203                     attr->gid_type == IB_GID_TYPE_ROCE) {
 204                         rcu_read_unlock();
 205                         if (gid)
 206                                 memcpy(gid, &attr->gid, SMC_GID_SIZE);
 207                         if (sgid_index)
 208                                 *sgid_index = attr->index;
 209                         rdma_put_gid_attr(attr);
 210                         return 0;
 211                 }
 212                 rcu_read_unlock();
 213                 rdma_put_gid_attr(attr);
 214         }
 215         return -ENODEV;
 216 }
 217
 218 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
 219 {
 220         int rc;
 221
 222         memset(&smcibdev->pattr[ibport - 1], 0,
 223                sizeof(smcibdev->pattr[ibport - 1]));
 224         rc = ib_query_port(smcibdev->ibdev, ibport,
 225                            &smcibdev->pattr[ibport - 1]);
 226         if (rc)
 227                 goto out;
 228         /* the SMC protocol requires specification of the RoCE MAC address */
 229         rc = smc_ib_fill_mac(smcibdev, ibport);
 230         if (rc)
 231                 goto out;
 232         if (!smc_ib_is_valid_local_systemid() &&
 233             smc_ib_port_active(smcibdev, ibport))
 234                 /* create unique system identifier */
 235                 smc_ib_define_local_systemid(smcibdev, ibport);
 236 out:
 237         return rc;
 238 }
 239
 240 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
 241 static void smc_ib_port_event_work(struct work_struct *work)
 242 {
 243         struct smc_ib_device *smcibdev = container_of(
 244                 work, struct smc_ib_device, port_event_work);
 245         u8 port_idx;
 246
 247         for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
 248                 smc_ib_remember_port_attr(smcibdev, port_idx + 1);
 249                 clear_bit(port_idx, &smcibdev->port_event_mask);
 250                 if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
 251                         set_bit(port_idx, smcibdev->ports_going_away);
 252                         smcr_port_err(smcibdev, port_idx + 1);
 253                 } else {
 254                         clear_bit(port_idx, smcibdev->ports_going_away);
 255                         smcr_port_add(smcibdev, port_idx + 1);
 256                 }
 257         }
 258 }
 259
 260 /* can be called in IRQ context */
 261 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
 262                                         struct ib_event *ibevent)
 263 {
 264         struct smc_ib_device *smcibdev;
 265         bool schedule = false;
 266         u8 port_idx;
 267
 268         smcibdev = container_of(handler, struct smc_ib_device, event_handler);
 269
 270         switch (ibevent->event) {
 271         case IB_EVENT_DEVICE_FATAL:
 272                 /* terminate all ports on device */
 273                 for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
 274                         set_bit(port_idx, &smcibdev->port_event_mask);
 275                         if (!test_and_set_bit(port_idx,
 276                                               smcibdev->ports_going_away))
 277                                 schedule = true;
 278                 }
 279                 if (schedule)
 280                         schedule_work(&smcibdev->port_event_work);
 281                 break;
 282         case IB_EVENT_PORT_ACTIVE:
 283                 port_idx = ibevent->element.port_num - 1;
 284                 if (port_idx >= SMC_MAX_PORTS)
 285                         break;
 286                 set_bit(port_idx, &smcibdev->port_event_mask);
 287                 if (test_and_clear_bit(port_idx, smcibdev->ports_going_away))
 288                         schedule_work(&smcibdev->port_event_work);
 289                 break;
 290         case IB_EVENT_PORT_ERR:
 291                 port_idx = ibevent->element.port_num - 1;
 292                 if (port_idx >= SMC_MAX_PORTS)
 293                         break;
 294                 set_bit(port_idx, &smcibdev->port_event_mask);
 295                 if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
 296                         schedule_work(&smcibdev->port_event_work);
 297                 break;
 298         case IB_EVENT_GID_CHANGE:
 299                 port_idx = ibevent->element.port_num - 1;
 300                 if (port_idx >= SMC_MAX_PORTS)
 301                         break;
 302                 set_bit(port_idx, &smcibdev->port_event_mask);
 303                 schedule_work(&smcibdev->port_event_work);
 304                 break;
 305         default:
 306                 break;
 307         }
 308 }
 309
 310 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
 311 {
 312         if (lnk->roce_pd)
 313                 ib_dealloc_pd(lnk->roce_pd);
 314         lnk->roce_pd = NULL;
 315 }
 316
 317 int smc_ib_create_protection_domain(struct smc_link *lnk)
 318 {
 319         int rc;
 320
 321         lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
 322         rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
 323         if (IS_ERR(lnk->roce_pd))
 324                 lnk->roce_pd = NULL;
 325         return rc;
 326 }
 327
 328 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
 329 {
 330         struct smc_link *lnk = (struct smc_link *)priv;
 331         struct smc_ib_device *smcibdev = lnk->smcibdev;
 332         u8 port_idx;
 333
 334         switch (ibevent->event) {
 335         case IB_EVENT_QP_FATAL:
 336         case IB_EVENT_QP_ACCESS_ERR:
 337                 port_idx = ibevent->element.qp->port - 1;
 338                 if (port_idx >= SMC_MAX_PORTS)
 339                         break;
 340                 set_bit(port_idx, &smcibdev->port_event_mask);
 341                 if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
 342                         schedule_work(&smcibdev->port_event_work);
 343                 break;
 344         default:
 345                 break;
 346         }
 347 }
 348
 349 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
 350 {
 351         if (lnk->roce_qp)
 352                 ib_destroy_qp(lnk->roce_qp);
 353         lnk->roce_qp = NULL;
 354 }
 355
 356 /* create a queue pair within the protection domain for a link */
 357 int smc_ib_create_queue_pair(struct smc_link *lnk)
 358 {
 359         struct ib_qp_init_attr qp_attr = {
 360                 .event_handler = smc_ib_qp_event_handler,
 361                 .qp_context = lnk,
 362                 .send_cq = lnk->smcibdev->roce_cq_send,
 363                 .recv_cq = lnk->smcibdev->roce_cq_recv,
 364                 .srq = NULL,
 365                 .cap = {
 366                                 /* include unsolicited rdma_writes as well,
 367                                  * there are max. 2 RDMA_WRITE per 1 WR_SEND
 368                                  */
 369                         .max_send_wr = SMC_WR_BUF_CNT * 3,
 370                         .max_recv_wr = SMC_WR_BUF_CNT * 3,
 371                         .max_send_sge = SMC_IB_MAX_SEND_SGE,
 372                         .max_recv_sge = 1,
 373                 },
 374                 .sq_sig_type = IB_SIGNAL_REQ_WR,
 375                 .qp_type = IB_QPT_RC,
 376         };
 377         int rc;
 378
 379         lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
 380         rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
 381         if (IS_ERR(lnk->roce_qp))
 382                 lnk->roce_qp = NULL;
 383         else
 384                 smc_wr_remember_qp_attr(lnk);
 385         return rc;
 386 }
 387
 388 void smc_ib_put_memory_region(struct ib_mr *mr)
 389 {
 390         ib_dereg_mr(mr);
 391 }
 392
 393 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx)
 394 {
 395         unsigned int offset = 0;
 396         int sg_num;
 397
 398         /* map the largest prefix of a dma mapped SG list */
 399         sg_num = ib_map_mr_sg(buf_slot->mr_rx[link_idx],
 400                               buf_slot->sgt[link_idx].sgl,
 401                               buf_slot->sgt[link_idx].orig_nents,
 402                               &offset, PAGE_SIZE);
 403
 404         return sg_num;
 405 }
 406
 407 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
 408 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
 409                              struct smc_buf_desc *buf_slot, u8 link_idx)
 410 {
 411         if (buf_slot->mr_rx[link_idx])
 412                 return 0; /* already done */
 413
 414         buf_slot->mr_rx[link_idx] =
 415                 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
 416         if (IS_ERR(buf_slot->mr_rx[link_idx])) {
 417                 int rc;
 418
 419                 rc = PTR_ERR(buf_slot->mr_rx[link_idx]);
 420                 buf_slot->mr_rx[link_idx] = NULL;
 421                 return rc;
 422         }
 423
 424         if (smc_ib_map_mr_sg(buf_slot, link_idx) != 1)
 425                 return -EINVAL;
 426
 427         return 0;
 428 }
 429
 430 /* synchronize buffer usage for cpu access */
 431 void smc_ib_sync_sg_for_cpu(struct smc_link *lnk,
 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[lnk->link_idx].sgl, sg,
 440                     buf_slot->sgt[lnk->link_idx].nents, i) {
 441                 if (!sg_dma_len(sg))
 442                         break;
 443                 ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev,
 444                                            sg_dma_address(sg),
 445                                            sg_dma_len(sg),
 446                                            data_direction);
 447         }
 448 }
 449
 450 /* synchronize buffer usage for device access */
 451 void smc_ib_sync_sg_for_device(struct smc_link *lnk,
 452                                struct smc_buf_desc *buf_slot,
 453                                enum dma_data_direction data_direction)
 454 {
 455         struct scatterlist *sg;
 456         unsigned int i;
 457
 458         /* for now there is just one DMA address */
 459         for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
 460                     buf_slot->sgt[lnk->link_idx].nents, i) {
 461                 if (!sg_dma_len(sg))
 462                         break;
 463                 ib_dma_sync_single_for_device(lnk->smcibdev->ibdev,
 464                                               sg_dma_address(sg),
 465                                               sg_dma_len(sg),
 466                                               data_direction);
 467         }
 468 }
 469
 470 /* Map a new TX or RX buffer SG-table to DMA */
 471 int smc_ib_buf_map_sg(struct smc_link *lnk,
 472                       struct smc_buf_desc *buf_slot,
 473                       enum dma_data_direction data_direction)
 474 {
 475         int mapped_nents;
 476
 477         mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev,
 478                                      buf_slot->sgt[lnk->link_idx].sgl,
 479                                      buf_slot->sgt[lnk->link_idx].orig_nents,
 480                                      data_direction);
 481         if (!mapped_nents)
 482                 return -ENOMEM;
 483
 484         return mapped_nents;
 485 }
 486
 487 void smc_ib_buf_unmap_sg(struct smc_link *lnk,
 488                          struct smc_buf_desc *buf_slot,
 489                          enum dma_data_direction data_direction)
 490 {
 491         if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address)
 492                 return; /* already unmapped */
 493
 494         ib_dma_unmap_sg(lnk->smcibdev->ibdev,
 495                         buf_slot->sgt[lnk->link_idx].sgl,
 496                         buf_slot->sgt[lnk->link_idx].orig_nents,
 497                         data_direction);
 498         buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0;
 499 }
 500
 501 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
 502 {
 503         struct ib_cq_init_attr cqattr = {
 504                 .cqe = SMC_MAX_CQE, .comp_vector = 0 };
 505         int cqe_size_order, smc_order;
 506         long rc;
 507
 508         /* the calculated number of cq entries fits to mlx5 cq allocation */
 509         cqe_size_order = cache_line_size() == 128 ? 7 : 6;
 510         smc_order = MAX_ORDER - cqe_size_order - 1;
 511         if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
 512                 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
 513         smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
 514                                               smc_wr_tx_cq_handler, NULL,
 515                                               smcibdev, &cqattr);
 516         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
 517         if (IS_ERR(smcibdev->roce_cq_send)) {
 518                 smcibdev->roce_cq_send = NULL;
 519                 return rc;
 520         }
 521         smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
 522                                               smc_wr_rx_cq_handler, NULL,
 523                                               smcibdev, &cqattr);
 524         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
 525         if (IS_ERR(smcibdev->roce_cq_recv)) {
 526                 smcibdev->roce_cq_recv = NULL;
 527                 goto err;
 528         }
 529         smc_wr_add_dev(smcibdev);
 530         smcibdev->initialized = 1;
 531         return rc;
 532
 533 err:
 534         ib_destroy_cq(smcibdev->roce_cq_send);
 535         return rc;
 536 }
 537
 538 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
 539 {
 540         if (!smcibdev->initialized)
 541                 return;
 542         smcibdev->initialized = 0;
 543         ib_destroy_cq(smcibdev->roce_cq_recv);
 544         ib_destroy_cq(smcibdev->roce_cq_send);
 545         smc_wr_remove_dev(smcibdev);
 546 }
 547
 548 static struct ib_client smc_ib_client;
 549
 550 /* callback function for ib_register_client() */
 551 static void smc_ib_add_dev(struct ib_device *ibdev)
 552 {
 553         struct smc_ib_device *smcibdev;
 554         u8 port_cnt;
 555         int i;
 556
 557         if (ibdev->node_type != RDMA_NODE_IB_CA)
 558                 return;
 559
 560         smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
 561         if (!smcibdev)
 562                 return;
 563
 564         smcibdev->ibdev = ibdev;
 565         INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
 566         atomic_set(&smcibdev->lnk_cnt, 0);
 567         init_waitqueue_head(&smcibdev->lnks_deleted);
 568         spin_lock(&smc_ib_devices.lock);
 569         list_add_tail(&smcibdev->list, &smc_ib_devices.list);
 570         spin_unlock(&smc_ib_devices.lock);
 571         ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
 572         INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
 573                               smc_ib_global_event_handler);
 574         ib_register_event_handler(&smcibdev->event_handler);
 575
 576         /* trigger reading of the port attributes */
 577         port_cnt = smcibdev->ibdev->phys_port_cnt;
 578         pr_warn_ratelimited("smc: adding ib device %s with port count %d\n",
 579                             smcibdev->ibdev->name, port_cnt);
 580         for (i = 0;
 581              i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
 582              i++) {
 583                 set_bit(i, &smcibdev->port_event_mask);
 584                 /* determine pnetids of the port */
 585                 if (smc_pnetid_by_dev_port(ibdev->dev.parent, i,
 586                                            smcibdev->pnetid[i]))
 587                         smc_pnetid_by_table_ib(smcibdev, i + 1);
 588                 pr_warn_ratelimited("smc:    ib device %s port %d has pnetid "
 589                                     "%.16s%s\n",
 590                                     smcibdev->ibdev->name, i + 1,
 591                                     smcibdev->pnetid[i],
 592                                     smcibdev->pnetid_by_user[i] ?
 593                                      " (user defined)" :
 594                                      "");
 595         }
 596         schedule_work(&smcibdev->port_event_work);
 597 }
 598
 599 /* callback function for ib_unregister_client() */
 600 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
 601 {
 602         struct smc_ib_device *smcibdev;
 603
 604         smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
 605         if (!smcibdev || smcibdev->ibdev != ibdev)
 606                 return;
 607         ib_set_client_data(ibdev, &smc_ib_client, NULL);
 608         spin_lock(&smc_ib_devices.lock);
 609         list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
 610         spin_unlock(&smc_ib_devices.lock);
 611         pr_warn_ratelimited("smc: removing ib device %s\n",
 612                             smcibdev->ibdev->name);
 613         smc_smcr_terminate_all(smcibdev);
 614         smc_ib_cleanup_per_ibdev(smcibdev);
 615         ib_unregister_event_handler(&smcibdev->event_handler);
 616         cancel_work_sync(&smcibdev->port_event_work);
 617         kfree(smcibdev);
 618 }
 619
 620 static struct ib_client smc_ib_client = {
 621         .name   = "smc_ib",
 622         .add    = smc_ib_add_dev,
 623         .remove = smc_ib_remove_dev,
 624 };
 625
 626 int __init smc_ib_register_client(void)
 627 {
 628         smc_ib_init_local_systemid();
 629         return ib_register_client(&smc_ib_client);
 630 }
 631
 632 void smc_ib_unregister_client(void)
 633 {
 634         ib_unregister_client(&smc_ib_client);
 635 }