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 <linux/mutex.h>
  20 #include <rdma/ib_verbs.h>
  21 #include <rdma/ib_cache.h>
  22
  23 #include "smc_pnet.h"
  24 #include "smc_ib.h"
  25 #include "smc_core.h"
  26 #include "smc_wr.h"
  27 #include "smc.h"
  28
  29 #define SMC_MAX_CQE 32766       /* max. # of completion queue elements */
  30
  31 #define SMC_QP_MIN_RNR_TIMER            5
  32 #define SMC_QP_TIMEOUT                  15 /* 4096 * 2 ** timeout usec */
  33 #define SMC_QP_RETRY_CNT                        7 /* 7: infinite */
  34 #define SMC_QP_RNR_RETRY                        7 /* 7: infinite */
  35
  36 struct smc_ib_devices smc_ib_devices = {        /* smc-registered ib devices */
  37         .mutex = __MUTEX_INITIALIZER(smc_ib_devices.mutex),
  38         .list = LIST_HEAD_INIT(smc_ib_devices.list),
  39 };
  40
  41 u8 local_systemid[SMC_SYSTEMID_LEN];            /* unique system identifier */
  42
  43 static int smc_ib_modify_qp_init(struct smc_link *lnk)
  44 {
  45         struct ib_qp_attr qp_attr;
  46
  47         memset(&qp_attr, 0, sizeof(qp_attr));
  48         qp_attr.qp_state = IB_QPS_INIT;
  49         qp_attr.pkey_index = 0;
  50         qp_attr.port_num = lnk->ibport;
  51         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
  52                                 | IB_ACCESS_REMOTE_WRITE;
  53         return ib_modify_qp(lnk->roce_qp, &qp_attr,
  54                             IB_QP_STATE | IB_QP_PKEY_INDEX |
  55                             IB_QP_ACCESS_FLAGS | IB_QP_PORT);
  56 }
  57
  58 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
  59 {
  60         enum ib_qp_attr_mask qp_attr_mask =
  61                 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
  62                 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
  63         struct ib_qp_attr qp_attr;
  64
  65         memset(&qp_attr, 0, sizeof(qp_attr));
  66         qp_attr.qp_state = IB_QPS_RTR;
  67         qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
  68         qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
  69         rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport);
  70         rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, 1, 0);
  71         rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid);
  72         memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac,
  73                sizeof(lnk->peer_mac));
  74         qp_attr.dest_qp_num = lnk->peer_qpn;
  75         qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
  76         qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
  77                                          * requests
  78                                          */
  79         qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
  80
  81         return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
  82 }
  83
  84 int smc_ib_modify_qp_rts(struct smc_link *lnk)
  85 {
  86         struct ib_qp_attr qp_attr;
  87
  88         memset(&qp_attr, 0, sizeof(qp_attr));
  89         qp_attr.qp_state = IB_QPS_RTS;
  90         qp_attr.timeout = SMC_QP_TIMEOUT;       /* local ack timeout */
  91         qp_attr.retry_cnt = SMC_QP_RETRY_CNT;   /* retry count */
  92         qp_attr.rnr_retry = SMC_QP_RNR_RETRY;   /* RNR retries, 7=infinite */
  93         qp_attr.sq_psn = lnk->psn_initial;      /* starting send packet seq # */
  94         qp_attr.max_rd_atomic = 1;      /* # of outstanding RDMA reads and
  95                                          * atomic ops allowed
  96                                          */
  97         return ib_modify_qp(lnk->roce_qp, &qp_attr,
  98                             IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
  99                             IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
 100                             IB_QP_MAX_QP_RD_ATOMIC);
 101 }
 102
 103 int smc_ib_modify_qp_reset(struct smc_link *lnk)
 104 {
 105         struct ib_qp_attr qp_attr;
 106
 107         memset(&qp_attr, 0, sizeof(qp_attr));
 108         qp_attr.qp_state = IB_QPS_RESET;
 109         return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
 110 }
 111
 112 int smc_ib_ready_link(struct smc_link *lnk)
 113 {
 114         struct smc_link_group *lgr = smc_get_lgr(lnk);
 115         int rc = 0;
 116
 117         rc = smc_ib_modify_qp_init(lnk);
 118         if (rc)
 119                 goto out;
 120
 121         rc = smc_ib_modify_qp_rtr(lnk);
 122         if (rc)
 123                 goto out;
 124         smc_wr_remember_qp_attr(lnk);
 125         rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
 126                               IB_CQ_SOLICITED_MASK);
 127         if (rc)
 128                 goto out;
 129         rc = smc_wr_rx_post_init(lnk);
 130         if (rc)
 131                 goto out;
 132         smc_wr_remember_qp_attr(lnk);
 133
 134         if (lgr->role == SMC_SERV) {
 135                 rc = smc_ib_modify_qp_rts(lnk);
 136                 if (rc)
 137                         goto out;
 138                 smc_wr_remember_qp_attr(lnk);
 139         }
 140 out:
 141         return rc;
 142 }
 143
 144 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport)
 145 {
 146         const struct ib_gid_attr *attr;
 147         int rc;
 148
 149         attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0);
 150         if (IS_ERR(attr))
 151                 return -ENODEV;
 152
 153         rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]);
 154         rdma_put_gid_attr(attr);
 155         return rc;
 156 }
 157
 158 /* Create an identifier unique for this instance of SMC-R.
 159  * The MAC-address of the first active registered IB device
 160  * plus a random 2-byte number is used to create this identifier.
 161  * This name is delivered to the peer during connection initialization.
 162  */
 163 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
 164                                                 u8 ibport)
 165 {
 166         memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
 167                sizeof(smcibdev->mac[ibport - 1]));
 168 }
 169
 170 bool smc_ib_is_valid_local_systemid(void)
 171 {
 172         return !is_zero_ether_addr(&local_systemid[2]);
 173 }
 174
 175 static void smc_ib_init_local_systemid(void)
 176 {
 177         get_random_bytes(&local_systemid[0], 2);
 178 }
 179
 180 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
 181 {
 182         return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
 183 }
 184
 185 /* determine the gid for an ib-device port and vlan id */
 186 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport,
 187                          unsigned short vlan_id, u8 gid[], u8 *sgid_index)
 188 {
 189         const struct ib_gid_attr *attr;
 190         const struct net_device *ndev;
 191         int i;
 192
 193         for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) {
 194                 attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i);
 195                 if (IS_ERR(attr))
 196                         continue;
 197
 198                 rcu_read_lock();
 199                 ndev = rdma_read_gid_attr_ndev_rcu(attr);
 200                 if (!IS_ERR(ndev) &&
 201                     ((!vlan_id && !is_vlan_dev(attr->ndev)) ||
 202                      (vlan_id && is_vlan_dev(attr->ndev) &&
 203                       vlan_dev_vlan_id(attr->ndev) == vlan_id)) &&
 204                     attr->gid_type == IB_GID_TYPE_ROCE) {
 205                         rcu_read_unlock();
 206                         if (gid)
 207                                 memcpy(gid, &attr->gid, SMC_GID_SIZE);
 208                         if (sgid_index)
 209                                 *sgid_index = attr->index;
 210                         rdma_put_gid_attr(attr);
 211                         return 0;
 212                 }
 213                 rcu_read_unlock();
 214                 rdma_put_gid_attr(attr);
 215         }
 216         return -ENODEV;
 217 }
 218
 219 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
 220 {
 221         int rc;
 222
 223         memset(&smcibdev->pattr[ibport - 1], 0,
 224                sizeof(smcibdev->pattr[ibport - 1]));
 225         rc = ib_query_port(smcibdev->ibdev, ibport,
 226                            &smcibdev->pattr[ibport - 1]);
 227         if (rc)
 228                 goto out;
 229         /* the SMC protocol requires specification of the RoCE MAC address */
 230         rc = smc_ib_fill_mac(smcibdev, ibport);
 231         if (rc)
 232                 goto out;
 233         if (!smc_ib_is_valid_local_systemid() &&
 234             smc_ib_port_active(smcibdev, ibport))
 235                 /* create unique system identifier */
 236                 smc_ib_define_local_systemid(smcibdev, ibport);
 237 out:
 238         return rc;
 239 }
 240
 241 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
 242 static void smc_ib_port_event_work(struct work_struct *work)
 243 {
 244         struct smc_ib_device *smcibdev = container_of(
 245                 work, struct smc_ib_device, port_event_work);
 246         u8 port_idx;
 247
 248         for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
 249                 smc_ib_remember_port_attr(smcibdev, port_idx + 1);
 250                 clear_bit(port_idx, &smcibdev->port_event_mask);
 251                 if (!smc_ib_port_active(smcibdev, port_idx + 1)) {
 252                         set_bit(port_idx, smcibdev->ports_going_away);
 253                         smcr_port_err(smcibdev, port_idx + 1);
 254                 } else {
 255                         clear_bit(port_idx, smcibdev->ports_going_away);
 256                         smcr_port_add(smcibdev, port_idx + 1);
 257                 }
 258         }
 259 }
 260
 261 /* can be called in IRQ context */
 262 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
 263                                         struct ib_event *ibevent)
 264 {
 265         struct smc_ib_device *smcibdev;
 266         bool schedule = false;
 267         u8 port_idx;
 268
 269         smcibdev = container_of(handler, struct smc_ib_device, event_handler);
 270
 271         switch (ibevent->event) {
 272         case IB_EVENT_DEVICE_FATAL:
 273                 /* terminate all ports on device */
 274                 for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) {
 275                         set_bit(port_idx, &smcibdev->port_event_mask);
 276                         if (!test_and_set_bit(port_idx,
 277                                               smcibdev->ports_going_away))
 278                                 schedule = true;
 279                 }
 280                 if (schedule)
 281                         schedule_work(&smcibdev->port_event_work);
 282                 break;
 283         case IB_EVENT_PORT_ACTIVE:
 284                 port_idx = ibevent->element.port_num - 1;
 285                 if (port_idx >= SMC_MAX_PORTS)
 286                         break;
 287                 set_bit(port_idx, &smcibdev->port_event_mask);
 288                 if (test_and_clear_bit(port_idx, smcibdev->ports_going_away))
 289                         schedule_work(&smcibdev->port_event_work);
 290                 break;
 291         case IB_EVENT_PORT_ERR:
 292                 port_idx = ibevent->element.port_num - 1;
 293                 if (port_idx >= SMC_MAX_PORTS)
 294                         break;
 295                 set_bit(port_idx, &smcibdev->port_event_mask);
 296                 if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
 297                         schedule_work(&smcibdev->port_event_work);
 298                 break;
 299         case IB_EVENT_GID_CHANGE:
 300                 port_idx = ibevent->element.port_num - 1;
 301                 if (port_idx >= SMC_MAX_PORTS)
 302                         break;
 303                 set_bit(port_idx, &smcibdev->port_event_mask);
 304                 schedule_work(&smcibdev->port_event_work);
 305                 break;
 306         default:
 307                 break;
 308         }
 309 }
 310
 311 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
 312 {
 313         if (lnk->roce_pd)
 314                 ib_dealloc_pd(lnk->roce_pd);
 315         lnk->roce_pd = NULL;
 316 }
 317
 318 int smc_ib_create_protection_domain(struct smc_link *lnk)
 319 {
 320         int rc;
 321
 322         lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
 323         rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
 324         if (IS_ERR(lnk->roce_pd))
 325                 lnk->roce_pd = NULL;
 326         return rc;
 327 }
 328
 329 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
 330 {
 331         struct smc_link *lnk = (struct smc_link *)priv;
 332         struct smc_ib_device *smcibdev = lnk->smcibdev;
 333         u8 port_idx;
 334
 335         switch (ibevent->event) {
 336         case IB_EVENT_QP_FATAL:
 337         case IB_EVENT_QP_ACCESS_ERR:
 338                 port_idx = ibevent->element.qp->port - 1;
 339                 if (port_idx >= SMC_MAX_PORTS)
 340                         break;
 341                 set_bit(port_idx, &smcibdev->port_event_mask);
 342                 if (!test_and_set_bit(port_idx, smcibdev->ports_going_away))
 343                         schedule_work(&smcibdev->port_event_work);
 344                 break;
 345         default:
 346                 break;
 347         }
 348 }
 349
 350 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
 351 {
 352         if (lnk->roce_qp)
 353                 ib_destroy_qp(lnk->roce_qp);
 354         lnk->roce_qp = NULL;
 355 }
 356
 357 /* create a queue pair within the protection domain for a link */
 358 int smc_ib_create_queue_pair(struct smc_link *lnk)
 359 {
 360         struct ib_qp_init_attr qp_attr = {
 361                 .event_handler = smc_ib_qp_event_handler,
 362                 .qp_context = lnk,
 363                 .send_cq = lnk->smcibdev->roce_cq_send,
 364                 .recv_cq = lnk->smcibdev->roce_cq_recv,
 365                 .srq = NULL,
 366                 .cap = {
 367                                 /* include unsolicited rdma_writes as well,
 368                                  * there are max. 2 RDMA_WRITE per 1 WR_SEND
 369                                  */
 370                         .max_send_wr = SMC_WR_BUF_CNT * 3,
 371                         .max_recv_wr = SMC_WR_BUF_CNT * 3,
 372                         .max_send_sge = SMC_IB_MAX_SEND_SGE,
 373                         .max_recv_sge = 1,
 374                 },
 375                 .sq_sig_type = IB_SIGNAL_REQ_WR,
 376                 .qp_type = IB_QPT_RC,
 377         };
 378         int rc;
 379
 380         lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
 381         rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
 382         if (IS_ERR(lnk->roce_qp))
 383                 lnk->roce_qp = NULL;
 384         else
 385                 smc_wr_remember_qp_attr(lnk);
 386         return rc;
 387 }
 388
 389 void smc_ib_put_memory_region(struct ib_mr *mr)
 390 {
 391         ib_dereg_mr(mr);
 392 }
 393
 394 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx)
 395 {
 396         unsigned int offset = 0;
 397         int sg_num;
 398
 399         /* map the largest prefix of a dma mapped SG list */
 400         sg_num = ib_map_mr_sg(buf_slot->mr_rx[link_idx],
 401                               buf_slot->sgt[link_idx].sgl,
 402                               buf_slot->sgt[link_idx].orig_nents,
 403                               &offset, PAGE_SIZE);
 404
 405         return sg_num;
 406 }
 407
 408 /* Allocate a memory region and map the dma mapped SG list of buf_slot */
 409 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
 410                              struct smc_buf_desc *buf_slot, u8 link_idx)
 411 {
 412         if (buf_slot->mr_rx[link_idx])
 413                 return 0; /* already done */
 414
 415         buf_slot->mr_rx[link_idx] =
 416                 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order);
 417         if (IS_ERR(buf_slot->mr_rx[link_idx])) {
 418                 int rc;
 419
 420                 rc = PTR_ERR(buf_slot->mr_rx[link_idx]);
 421                 buf_slot->mr_rx[link_idx] = NULL;
 422                 return rc;
 423         }
 424
 425         if (smc_ib_map_mr_sg(buf_slot, link_idx) != 1)
 426                 return -EINVAL;
 427
 428         return 0;
 429 }
 430
 431 /* synchronize buffer usage for cpu access */
 432 void smc_ib_sync_sg_for_cpu(struct smc_link *lnk,
 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[lnk->link_idx].sgl, sg,
 441                     buf_slot->sgt[lnk->link_idx].nents, i) {
 442                 if (!sg_dma_len(sg))
 443                         break;
 444                 ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev,
 445                                            sg_dma_address(sg),
 446                                            sg_dma_len(sg),
 447                                            data_direction);
 448         }
 449 }
 450
 451 /* synchronize buffer usage for device access */
 452 void smc_ib_sync_sg_for_device(struct smc_link *lnk,
 453                                struct smc_buf_desc *buf_slot,
 454                                enum dma_data_direction data_direction)
 455 {
 456         struct scatterlist *sg;
 457         unsigned int i;
 458
 459         /* for now there is just one DMA address */
 460         for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg,
 461                     buf_slot->sgt[lnk->link_idx].nents, i) {
 462                 if (!sg_dma_len(sg))
 463                         break;
 464                 ib_dma_sync_single_for_device(lnk->smcibdev->ibdev,
 465                                               sg_dma_address(sg),
 466                                               sg_dma_len(sg),
 467                                               data_direction);
 468         }
 469 }
 470
 471 /* Map a new TX or RX buffer SG-table to DMA */
 472 int smc_ib_buf_map_sg(struct smc_link *lnk,
 473                       struct smc_buf_desc *buf_slot,
 474                       enum dma_data_direction data_direction)
 475 {
 476         int mapped_nents;
 477
 478         mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev,
 479                                      buf_slot->sgt[lnk->link_idx].sgl,
 480                                      buf_slot->sgt[lnk->link_idx].orig_nents,
 481                                      data_direction);
 482         if (!mapped_nents)
 483                 return -ENOMEM;
 484
 485         return mapped_nents;
 486 }
 487
 488 void smc_ib_buf_unmap_sg(struct smc_link *lnk,
 489                          struct smc_buf_desc *buf_slot,
 490                          enum dma_data_direction data_direction)
 491 {
 492         if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address)
 493                 return; /* already unmapped */
 494
 495         ib_dma_unmap_sg(lnk->smcibdev->ibdev,
 496                         buf_slot->sgt[lnk->link_idx].sgl,
 497                         buf_slot->sgt[lnk->link_idx].orig_nents,
 498                         data_direction);
 499         buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0;
 500 }
 501
 502 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
 503 {
 504         struct ib_cq_init_attr cqattr = {
 505                 .cqe = SMC_MAX_CQE, .comp_vector = 0 };
 506         int cqe_size_order, smc_order;
 507         long rc;
 508
 509         mutex_lock(&smcibdev->mutex);
 510         rc = 0;
 511         if (smcibdev->initialized)
 512                 goto out;
 513         /* the calculated number of cq entries fits to mlx5 cq allocation */
 514         cqe_size_order = cache_line_size() == 128 ? 7 : 6;
 515         smc_order = MAX_ORDER - cqe_size_order - 1;
 516         if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE)
 517                 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2;
 518         smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
 519                                               smc_wr_tx_cq_handler, NULL,
 520                                               smcibdev, &cqattr);
 521         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
 522         if (IS_ERR(smcibdev->roce_cq_send)) {
 523                 smcibdev->roce_cq_send = NULL;
 524                 goto out;
 525         }
 526         smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
 527                                               smc_wr_rx_cq_handler, NULL,
 528                                               smcibdev, &cqattr);
 529         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
 530         if (IS_ERR(smcibdev->roce_cq_recv)) {
 531                 smcibdev->roce_cq_recv = NULL;
 532                 goto err;
 533         }
 534         smc_wr_add_dev(smcibdev);
 535         smcibdev->initialized = 1;
 536         goto out;
 537
 538 err:
 539         ib_destroy_cq(smcibdev->roce_cq_send);
 540 out:
 541         mutex_unlock(&smcibdev->mutex);
 542         return rc;
 543 }
 544
 545 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
 546 {
 547         mutex_lock(&smcibdev->mutex);
 548         if (!smcibdev->initialized)
 549                 goto out;
 550         smcibdev->initialized = 0;
 551         ib_destroy_cq(smcibdev->roce_cq_recv);
 552         ib_destroy_cq(smcibdev->roce_cq_send);
 553         smc_wr_remove_dev(smcibdev);
 554 out:
 555         mutex_unlock(&smcibdev->mutex);
 556 }
 557
 558 static struct ib_client smc_ib_client;
 559
 560 /* callback function for ib_register_client() */
 561 static int smc_ib_add_dev(struct ib_device *ibdev)
 562 {
 563         struct smc_ib_device *smcibdev;
 564         u8 port_cnt;
 565         int i;
 566
 567         if (ibdev->node_type != RDMA_NODE_IB_CA)
 568                 return -EOPNOTSUPP;
 569
 570         smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
 571         if (!smcibdev)
 572                 return -ENOMEM;
 573
 574         smcibdev->ibdev = ibdev;
 575         INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
 576         atomic_set(&smcibdev->lnk_cnt, 0);
 577         init_waitqueue_head(&smcibdev->lnks_deleted);
 578         mutex_init(&smcibdev->mutex);
 579         mutex_lock(&smc_ib_devices.mutex);
 580         list_add_tail(&smcibdev->list, &smc_ib_devices.list);
 581         mutex_unlock(&smc_ib_devices.mutex);
 582         ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
 583         INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
 584                               smc_ib_global_event_handler);
 585         ib_register_event_handler(&smcibdev->event_handler);
 586
 587         /* trigger reading of the port attributes */
 588         port_cnt = smcibdev->ibdev->phys_port_cnt;
 589         pr_warn_ratelimited("smc: adding ib device %s with port count %d\n",
 590                             smcibdev->ibdev->name, port_cnt);
 591         for (i = 0;
 592              i < min_t(size_t, port_cnt, SMC_MAX_PORTS);
 593              i++) {
 594                 set_bit(i, &smcibdev->port_event_mask);
 595                 /* determine pnetids of the port */
 596                 if (smc_pnetid_by_dev_port(ibdev->dev.parent, i,
 597                                            smcibdev->pnetid[i]))
 598                         smc_pnetid_by_table_ib(smcibdev, i + 1);
 599                 pr_warn_ratelimited("smc:    ib device %s port %d has pnetid "
 600                                     "%.16s%s\n",
 601                                     smcibdev->ibdev->name, i + 1,
 602                                     smcibdev->pnetid[i],
 603                                     smcibdev->pnetid_by_user[i] ?
 604                                      " (user defined)" :
 605                                      "");
 606         }
 607         schedule_work(&smcibdev->port_event_work);
 608         return 0;
 609 }
 610
 611 /* callback function for ib_unregister_client() */
 612 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
 613 {
 614         struct smc_ib_device *smcibdev = client_data;
 615
 616         mutex_lock(&smc_ib_devices.mutex);
 617         list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
 618         mutex_unlock(&smc_ib_devices.mutex);
 619         pr_warn_ratelimited("smc: removing ib device %s\n",
 620                             smcibdev->ibdev->name);
 621         smc_smcr_terminate_all(smcibdev);
 622         smc_ib_cleanup_per_ibdev(smcibdev);
 623         ib_unregister_event_handler(&smcibdev->event_handler);
 624         cancel_work_sync(&smcibdev->port_event_work);
 625         kfree(smcibdev);
 626 }
 627
 628 static struct ib_client smc_ib_client = {
 629         .name   = "smc_ib",
 630         .add    = smc_ib_add_dev,
 631         .remove = smc_ib_remove_dev,
 632 };
 633
 634 int __init smc_ib_register_client(void)
 635 {
 636         smc_ib_init_local_systemid();
 637         return ib_register_client(&smc_ib_client);
 638 }
 639
 640 void smc_ib_unregister_client(void)
 641 {
 642         ib_unregister_client(&smc_ib_client);
 643 }