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