net/smc/smc_ib.c

   1 /*
   2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   3  *
   4  *  IB infrastructure:
   5  *  Establish SMC-R as an Infiniband Client to be notified about added and
   6  *  removed IB devices of type RDMA.
   7  *  Determine device and port characteristics for these IB devices.
   8  *
   9  *  Copyright IBM Corp. 2016
  10  *
  11  *  Author(s):  Ursula Braun <[email protected]>
  12  */
  13
  14 #include <linux/random.h>
  15 #include <linux/workqueue.h>
  16 #include <rdma/ib_verbs.h>
  17
  18 #include "smc_pnet.h"
  19 #include "smc_ib.h"
  20 #include "smc_core.h"
  21 #include "smc_wr.h"
  22 #include "smc.h"
  23
  24 #define SMC_QP_MIN_RNR_TIMER            5
  25 #define SMC_QP_TIMEOUT                  15 /* 4096 * 2 ** timeout usec */
  26 #define SMC_QP_RETRY_CNT                        7 /* 7: infinite */
  27 #define SMC_QP_RNR_RETRY                        7 /* 7: infinite */
  28
  29 struct smc_ib_devices smc_ib_devices = {        /* smc-registered ib devices */
  30         .lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock),
  31         .list = LIST_HEAD_INIT(smc_ib_devices.list),
  32 };
  33
  34 #define SMC_LOCAL_SYSTEMID_RESET        "%%%%%%%"
  35
  36 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET; /* unique system
  37                                                                  * identifier
  38                                                                  */
  39
  40 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags,
  41                              struct ib_mr **mr)
  42 {
  43         int rc;
  44
  45         if (*mr)
  46                 return 0; /* already done */
  47
  48         /* obtain unique key -
  49          * next invocation of get_dma_mr returns a different key!
  50          */
  51         *mr = pd->device->get_dma_mr(pd, access_flags);
  52         rc = PTR_ERR_OR_ZERO(*mr);
  53         if (IS_ERR(*mr))
  54                 *mr = NULL;
  55         return rc;
  56 }
  57
  58 static int smc_ib_modify_qp_init(struct smc_link *lnk)
  59 {
  60         struct ib_qp_attr qp_attr;
  61
  62         memset(&qp_attr, 0, sizeof(qp_attr));
  63         qp_attr.qp_state = IB_QPS_INIT;
  64         qp_attr.pkey_index = 0;
  65         qp_attr.port_num = lnk->ibport;
  66         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE
  67                                 | IB_ACCESS_REMOTE_WRITE;
  68         return ib_modify_qp(lnk->roce_qp, &qp_attr,
  69                             IB_QP_STATE | IB_QP_PKEY_INDEX |
  70                             IB_QP_ACCESS_FLAGS | IB_QP_PORT);
  71 }
  72
  73 static int smc_ib_modify_qp_rtr(struct smc_link *lnk)
  74 {
  75         enum ib_qp_attr_mask qp_attr_mask =
  76                 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN |
  77                 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER;
  78         struct ib_qp_attr qp_attr;
  79
  80         memset(&qp_attr, 0, sizeof(qp_attr));
  81         qp_attr.qp_state = IB_QPS_RTR;
  82         qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu);
  83         qp_attr.ah_attr.port_num = lnk->ibport;
  84         qp_attr.ah_attr.ah_flags = IB_AH_GRH;
  85         qp_attr.ah_attr.grh.hop_limit = 1;
  86         memcpy(&qp_attr.ah_attr.grh.dgid, lnk->peer_gid,
  87                sizeof(lnk->peer_gid));
  88         memcpy(&qp_attr.ah_attr.dmac, lnk->peer_mac,
  89                sizeof(lnk->peer_mac));
  90         qp_attr.dest_qp_num = lnk->peer_qpn;
  91         qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */
  92         qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming
  93                                          * requests
  94                                          */
  95         qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER;
  96
  97         return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask);
  98 }
  99
 100 int smc_ib_modify_qp_rts(struct smc_link *lnk)
 101 {
 102         struct ib_qp_attr qp_attr;
 103
 104         memset(&qp_attr, 0, sizeof(qp_attr));
 105         qp_attr.qp_state = IB_QPS_RTS;
 106         qp_attr.timeout = SMC_QP_TIMEOUT;       /* local ack timeout */
 107         qp_attr.retry_cnt = SMC_QP_RETRY_CNT;   /* retry count */
 108         qp_attr.rnr_retry = SMC_QP_RNR_RETRY;   /* RNR retries, 7=infinite */
 109         qp_attr.sq_psn = lnk->psn_initial;      /* starting send packet seq # */
 110         qp_attr.max_rd_atomic = 1;      /* # of outstanding RDMA reads and
 111                                          * atomic ops allowed
 112                                          */
 113         return ib_modify_qp(lnk->roce_qp, &qp_attr,
 114                             IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
 115                             IB_QP_SQ_PSN | IB_QP_RNR_RETRY |
 116                             IB_QP_MAX_QP_RD_ATOMIC);
 117 }
 118
 119 int smc_ib_modify_qp_reset(struct smc_link *lnk)
 120 {
 121         struct ib_qp_attr qp_attr;
 122
 123         memset(&qp_attr, 0, sizeof(qp_attr));
 124         qp_attr.qp_state = IB_QPS_RESET;
 125         return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE);
 126 }
 127
 128 int smc_ib_ready_link(struct smc_link *lnk)
 129 {
 130         struct smc_link_group *lgr =
 131                 container_of(lnk, struct smc_link_group, lnk[0]);
 132         int rc = 0;
 133
 134         rc = smc_ib_modify_qp_init(lnk);
 135         if (rc)
 136                 goto out;
 137
 138         rc = smc_ib_modify_qp_rtr(lnk);
 139         if (rc)
 140                 goto out;
 141         smc_wr_remember_qp_attr(lnk);
 142         rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv,
 143                               IB_CQ_SOLICITED_MASK);
 144         if (rc)
 145                 goto out;
 146         rc = smc_wr_rx_post_init(lnk);
 147         if (rc)
 148                 goto out;
 149         smc_wr_remember_qp_attr(lnk);
 150
 151         if (lgr->role == SMC_SERV) {
 152                 rc = smc_ib_modify_qp_rts(lnk);
 153                 if (rc)
 154                         goto out;
 155                 smc_wr_remember_qp_attr(lnk);
 156         }
 157 out:
 158         return rc;
 159 }
 160
 161 /* process context wrapper for might_sleep smc_ib_remember_port_attr */
 162 static void smc_ib_port_event_work(struct work_struct *work)
 163 {
 164         struct smc_ib_device *smcibdev = container_of(
 165                 work, struct smc_ib_device, port_event_work);
 166         u8 port_idx;
 167
 168         for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) {
 169                 smc_ib_remember_port_attr(smcibdev, port_idx + 1);
 170                 clear_bit(port_idx, &smcibdev->port_event_mask);
 171         }
 172 }
 173
 174 /* can be called in IRQ context */
 175 static void smc_ib_global_event_handler(struct ib_event_handler *handler,
 176                                         struct ib_event *ibevent)
 177 {
 178         struct smc_ib_device *smcibdev;
 179         u8 port_idx;
 180
 181         smcibdev = container_of(handler, struct smc_ib_device, event_handler);
 182         if (!smc_pnet_find_ib(smcibdev->ibdev->name))
 183                 return;
 184
 185         switch (ibevent->event) {
 186         case IB_EVENT_PORT_ERR:
 187                 port_idx = ibevent->element.port_num - 1;
 188                 set_bit(port_idx, &smcibdev->port_event_mask);
 189                 schedule_work(&smcibdev->port_event_work);
 190                 /* fall through */
 191         case IB_EVENT_DEVICE_FATAL:
 192                 /* tbd in follow-on patch:
 193                  * abnormal close of corresponding connections
 194                  */
 195                 break;
 196         case IB_EVENT_PORT_ACTIVE:
 197                 port_idx = ibevent->element.port_num - 1;
 198                 set_bit(port_idx, &smcibdev->port_event_mask);
 199                 schedule_work(&smcibdev->port_event_work);
 200                 break;
 201         default:
 202                 break;
 203         }
 204 }
 205
 206 void smc_ib_dealloc_protection_domain(struct smc_link *lnk)
 207 {
 208         ib_dealloc_pd(lnk->roce_pd);
 209         lnk->roce_pd = NULL;
 210 }
 211
 212 int smc_ib_create_protection_domain(struct smc_link *lnk)
 213 {
 214         int rc;
 215
 216         lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0);
 217         rc = PTR_ERR_OR_ZERO(lnk->roce_pd);
 218         if (IS_ERR(lnk->roce_pd))
 219                 lnk->roce_pd = NULL;
 220         return rc;
 221 }
 222
 223 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv)
 224 {
 225         switch (ibevent->event) {
 226         case IB_EVENT_DEVICE_FATAL:
 227         case IB_EVENT_GID_CHANGE:
 228         case IB_EVENT_PORT_ERR:
 229         case IB_EVENT_QP_ACCESS_ERR:
 230                 /* tbd in follow-on patch:
 231                  * abnormal close of corresponding connections
 232                  */
 233                 break;
 234         default:
 235                 break;
 236         }
 237 }
 238
 239 void smc_ib_destroy_queue_pair(struct smc_link *lnk)
 240 {
 241         ib_destroy_qp(lnk->roce_qp);
 242         lnk->roce_qp = NULL;
 243 }
 244
 245 /* create a queue pair within the protection domain for a link */
 246 int smc_ib_create_queue_pair(struct smc_link *lnk)
 247 {
 248         struct ib_qp_init_attr qp_attr = {
 249                 .event_handler = smc_ib_qp_event_handler,
 250                 .qp_context = lnk,
 251                 .send_cq = lnk->smcibdev->roce_cq_send,
 252                 .recv_cq = lnk->smcibdev->roce_cq_recv,
 253                 .srq = NULL,
 254                 .cap = {
 255                         .max_send_wr = SMC_WR_BUF_CNT,
 256                                 /* include unsolicited rdma_writes as well,
 257                                  * there are max. 2 RDMA_WRITE per 1 WR_SEND
 258                                  */
 259                         .max_recv_wr = SMC_WR_BUF_CNT * 3,
 260                         .max_send_sge = SMC_IB_MAX_SEND_SGE,
 261                         .max_recv_sge = 1,
 262                         .max_inline_data = SMC_WR_TX_SIZE,
 263                 },
 264                 .sq_sig_type = IB_SIGNAL_REQ_WR,
 265                 .qp_type = IB_QPT_RC,
 266         };
 267         int rc;
 268
 269         lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr);
 270         rc = PTR_ERR_OR_ZERO(lnk->roce_qp);
 271         if (IS_ERR(lnk->roce_qp))
 272                 lnk->roce_qp = NULL;
 273         else
 274                 smc_wr_remember_qp_attr(lnk);
 275         return rc;
 276 }
 277
 278 /* map a new TX or RX buffer to DMA */
 279 int smc_ib_buf_map(struct smc_ib_device *smcibdev, int buf_size,
 280                    struct smc_buf_desc *buf_slot,
 281                    enum dma_data_direction data_direction)
 282 {
 283         int rc = 0;
 284
 285         if (buf_slot->dma_addr[SMC_SINGLE_LINK])
 286                 return rc; /* already mapped */
 287         buf_slot->dma_addr[SMC_SINGLE_LINK] =
 288                 ib_dma_map_single(smcibdev->ibdev, buf_slot->cpu_addr,
 289                                   buf_size, data_direction);
 290         if (ib_dma_mapping_error(smcibdev->ibdev,
 291                                  buf_slot->dma_addr[SMC_SINGLE_LINK]))
 292                 rc = -EIO;
 293         return rc;
 294 }
 295
 296 void smc_ib_buf_unmap(struct smc_ib_device *smcibdev, int buf_size,
 297                       struct smc_buf_desc *buf_slot,
 298                       enum dma_data_direction data_direction)
 299 {
 300         if (!buf_slot->dma_addr[SMC_SINGLE_LINK])
 301                 return; /* already unmapped */
 302         ib_dma_unmap_single(smcibdev->ibdev, *buf_slot->dma_addr, buf_size,
 303                             data_direction);
 304         buf_slot->dma_addr[SMC_SINGLE_LINK] = 0;
 305 }
 306
 307 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport)
 308 {
 309         struct net_device *ndev;
 310         int rc;
 311
 312         rc = ib_query_gid(smcibdev->ibdev, ibport, 0,
 313                           &smcibdev->gid[ibport - 1], NULL);
 314         /* the SMC protocol requires specification of the roce MAC address;
 315          * if net_device cannot be determined, it can be derived from gid 0
 316          */
 317         ndev = smcibdev->ibdev->get_netdev(smcibdev->ibdev, ibport);
 318         if (ndev) {
 319                 memcpy(&smcibdev->mac, ndev->dev_addr, ETH_ALEN);
 320         } else if (!rc) {
 321                 memcpy(&smcibdev->mac[ibport - 1][0],
 322                        &smcibdev->gid[ibport - 1].raw[8], 3);
 323                 memcpy(&smcibdev->mac[ibport - 1][3],
 324                        &smcibdev->gid[ibport - 1].raw[13], 3);
 325                 smcibdev->mac[ibport - 1][0] &= ~0x02;
 326         }
 327         return rc;
 328 }
 329
 330 /* Create an identifier unique for this instance of SMC-R.
 331  * The MAC-address of the first active registered IB device
 332  * plus a random 2-byte number is used to create this identifier.
 333  * This name is delivered to the peer during connection initialization.
 334  */
 335 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev,
 336                                                 u8 ibport)
 337 {
 338         memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1],
 339                sizeof(smcibdev->mac[ibport - 1]));
 340         get_random_bytes(&local_systemid[0], 2);
 341 }
 342
 343 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport)
 344 {
 345         return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE;
 346 }
 347
 348 int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport)
 349 {
 350         int rc;
 351
 352         memset(&smcibdev->pattr[ibport - 1], 0,
 353                sizeof(smcibdev->pattr[ibport - 1]));
 354         rc = ib_query_port(smcibdev->ibdev, ibport,
 355                            &smcibdev->pattr[ibport - 1]);
 356         if (rc)
 357                 goto out;
 358         rc = smc_ib_fill_gid_and_mac(smcibdev, ibport);
 359         if (rc)
 360                 goto out;
 361         if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET,
 362                      sizeof(local_systemid)) &&
 363             smc_ib_port_active(smcibdev, ibport))
 364                 /* create unique system identifier */
 365                 smc_ib_define_local_systemid(smcibdev, ibport);
 366 out:
 367         return rc;
 368 }
 369
 370 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev)
 371 {
 372         struct ib_cq_init_attr cqattr = {
 373                 .cqe = SMC_WR_MAX_CQE, .comp_vector = 0 };
 374         long rc;
 375
 376         smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev,
 377                                               smc_wr_tx_cq_handler, NULL,
 378                                               smcibdev, &cqattr);
 379         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send);
 380         if (IS_ERR(smcibdev->roce_cq_send)) {
 381                 smcibdev->roce_cq_send = NULL;
 382                 return rc;
 383         }
 384         smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev,
 385                                               smc_wr_rx_cq_handler, NULL,
 386                                               smcibdev, &cqattr);
 387         rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv);
 388         if (IS_ERR(smcibdev->roce_cq_recv)) {
 389                 smcibdev->roce_cq_recv = NULL;
 390                 goto err;
 391         }
 392         INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev,
 393                               smc_ib_global_event_handler);
 394         ib_register_event_handler(&smcibdev->event_handler);
 395         smc_wr_add_dev(smcibdev);
 396         smcibdev->initialized = 1;
 397         return rc;
 398
 399 err:
 400         ib_destroy_cq(smcibdev->roce_cq_send);
 401         return rc;
 402 }
 403
 404 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev)
 405 {
 406         if (!smcibdev->initialized)
 407                 return;
 408         smc_wr_remove_dev(smcibdev);
 409         ib_unregister_event_handler(&smcibdev->event_handler);
 410         ib_destroy_cq(smcibdev->roce_cq_recv);
 411         ib_destroy_cq(smcibdev->roce_cq_send);
 412 }
 413
 414 static struct ib_client smc_ib_client;
 415
 416 /* callback function for ib_register_client() */
 417 static void smc_ib_add_dev(struct ib_device *ibdev)
 418 {
 419         struct smc_ib_device *smcibdev;
 420
 421         if (ibdev->node_type != RDMA_NODE_IB_CA)
 422                 return;
 423
 424         smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL);
 425         if (!smcibdev)
 426                 return;
 427
 428         smcibdev->ibdev = ibdev;
 429         INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work);
 430
 431         spin_lock(&smc_ib_devices.lock);
 432         list_add_tail(&smcibdev->list, &smc_ib_devices.list);
 433         spin_unlock(&smc_ib_devices.lock);
 434         ib_set_client_data(ibdev, &smc_ib_client, smcibdev);
 435 }
 436
 437 /* callback function for ib_register_client() */
 438 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data)
 439 {
 440         struct smc_ib_device *smcibdev;
 441
 442         smcibdev = ib_get_client_data(ibdev, &smc_ib_client);
 443         ib_set_client_data(ibdev, &smc_ib_client, NULL);
 444         spin_lock(&smc_ib_devices.lock);
 445         list_del_init(&smcibdev->list); /* remove from smc_ib_devices */
 446         spin_unlock(&smc_ib_devices.lock);
 447         smc_pnet_remove_by_ibdev(smcibdev);
 448         smc_ib_cleanup_per_ibdev(smcibdev);
 449         kfree(smcibdev);
 450 }
 451
 452 static struct ib_client smc_ib_client = {
 453         .name   = "smc_ib",
 454         .add    = smc_ib_add_dev,
 455         .remove = smc_ib_remove_dev,
 456 };
 457
 458 int __init smc_ib_register_client(void)
 459 {
 460         return ib_register_client(&smc_ib_client);
 461 }
 462
 463 void smc_ib_unregister_client(void)
 464 {
 465         ib_unregister_client(&smc_ib_client);
 466 }