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/etherdevice.h> 16 #include <linux/if_vlan.h> 17 #include <linux/random.h> 18 #include <linux/workqueue.h> 19 #include <linux/scatterlist.h> 20 #include <linux/wait.h> 21 #include <linux/mutex.h> 22 #include <linux/inetdevice.h> 23 #include <rdma/ib_verbs.h> 24 #include <rdma/ib_cache.h> 25 26 #include "smc_pnet.h" 27 #include "smc_ib.h" 28 #include "smc_core.h" 29 #include "smc_wr.h" 30 #include "smc.h" 31 #include "smc_netlink.h" 32 33 #define SMC_MAX_CQE 32766 /* max. # of completion queue elements */ 34 35 #define SMC_QP_MIN_RNR_TIMER 5 36 #define SMC_QP_TIMEOUT 15 /* 4096 * 2 ** timeout usec */ 37 #define SMC_QP_RETRY_CNT 7 /* 7: infinite */ 38 #define SMC_QP_RNR_RETRY 7 /* 7: infinite */ 39 40 struct smc_ib_devices smc_ib_devices = { /* smc-registered ib devices */ 41 .mutex = __MUTEX_INITIALIZER(smc_ib_devices.mutex), 42 .list = LIST_HEAD_INIT(smc_ib_devices.list), 43 }; 44 45 u8 local_systemid[SMC_SYSTEMID_LEN]; /* unique system identifier */ 46 47 static int smc_ib_modify_qp_init(struct smc_link *lnk) 48 { 49 struct ib_qp_attr qp_attr; 50 51 memset(&qp_attr, 0, sizeof(qp_attr)); 52 qp_attr.qp_state = IB_QPS_INIT; 53 qp_attr.pkey_index = 0; 54 qp_attr.port_num = lnk->ibport; 55 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE 56 | IB_ACCESS_REMOTE_WRITE; 57 return ib_modify_qp(lnk->roce_qp, &qp_attr, 58 IB_QP_STATE | IB_QP_PKEY_INDEX | 59 IB_QP_ACCESS_FLAGS | IB_QP_PORT); 60 } 61 62 static int smc_ib_modify_qp_rtr(struct smc_link *lnk) 63 { 64 enum ib_qp_attr_mask qp_attr_mask = 65 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN | 66 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER; 67 struct ib_qp_attr qp_attr; 68 u8 hop_lim = 1; 69 70 memset(&qp_attr, 0, sizeof(qp_attr)); 71 qp_attr.qp_state = IB_QPS_RTR; 72 qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu); 73 qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE; 74 rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport); 75 if (lnk->lgr->smc_version == SMC_V2 && lnk->lgr->uses_gateway) 76 hop_lim = IPV6_DEFAULT_HOPLIMIT; 77 rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, lnk->sgid_index, hop_lim, 0); 78 rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid); 79 if (lnk->lgr->smc_version == SMC_V2 && lnk->lgr->uses_gateway) 80 memcpy(&qp_attr.ah_attr.roce.dmac, lnk->lgr->nexthop_mac, 81 sizeof(lnk->lgr->nexthop_mac)); 82 else 83 memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac, 84 sizeof(lnk->peer_mac)); 85 qp_attr.dest_qp_num = lnk->peer_qpn; 86 qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */ 87 qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming 88 * requests 89 */ 90 qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER; 91 92 return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask); 93 } 94 95 int smc_ib_modify_qp_rts(struct smc_link *lnk) 96 { 97 struct ib_qp_attr qp_attr; 98 99 memset(&qp_attr, 0, sizeof(qp_attr)); 100 qp_attr.qp_state = IB_QPS_RTS; 101 qp_attr.timeout = SMC_QP_TIMEOUT; /* local ack timeout */ 102 qp_attr.retry_cnt = SMC_QP_RETRY_CNT; /* retry count */ 103 qp_attr.rnr_retry = SMC_QP_RNR_RETRY; /* RNR retries, 7=infinite */ 104 qp_attr.sq_psn = lnk->psn_initial; /* starting send packet seq # */ 105 qp_attr.max_rd_atomic = 1; /* # of outstanding RDMA reads and 106 * atomic ops allowed 107 */ 108 return ib_modify_qp(lnk->roce_qp, &qp_attr, 109 IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT | 110 IB_QP_SQ_PSN | IB_QP_RNR_RETRY | 111 IB_QP_MAX_QP_RD_ATOMIC); 112 } 113 114 int smc_ib_modify_qp_error(struct smc_link *lnk) 115 { 116 struct ib_qp_attr qp_attr; 117 118 memset(&qp_attr, 0, sizeof(qp_attr)); 119 qp_attr.qp_state = IB_QPS_ERR; 120 return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE); 121 } 122 123 int smc_ib_ready_link(struct smc_link *lnk) 124 { 125 struct smc_link_group *lgr = smc_get_lgr(lnk); 126 int rc = 0; 127 128 rc = smc_ib_modify_qp_init(lnk); 129 if (rc) 130 goto out; 131 132 rc = smc_ib_modify_qp_rtr(lnk); 133 if (rc) 134 goto out; 135 smc_wr_remember_qp_attr(lnk); 136 rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv, 137 IB_CQ_SOLICITED_MASK); 138 if (rc) 139 goto out; 140 rc = smc_wr_rx_post_init(lnk); 141 if (rc) 142 goto out; 143 smc_wr_remember_qp_attr(lnk); 144 145 if (lgr->role == SMC_SERV) { 146 rc = smc_ib_modify_qp_rts(lnk); 147 if (rc) 148 goto out; 149 smc_wr_remember_qp_attr(lnk); 150 } 151 out: 152 return rc; 153 } 154 155 static int smc_ib_fill_mac(struct smc_ib_device *smcibdev, u8 ibport) 156 { 157 const struct ib_gid_attr *attr; 158 int rc; 159 160 attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, 0); 161 if (IS_ERR(attr)) 162 return -ENODEV; 163 164 rc = rdma_read_gid_l2_fields(attr, NULL, smcibdev->mac[ibport - 1]); 165 rdma_put_gid_attr(attr); 166 return rc; 167 } 168 169 /* Create an identifier unique for this instance of SMC-R. 170 * The MAC-address of the first active registered IB device 171 * plus a random 2-byte number is used to create this identifier. 172 * This name is delivered to the peer during connection initialization. 173 */ 174 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev, 175 u8 ibport) 176 { 177 memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1], 178 sizeof(smcibdev->mac[ibport - 1])); 179 } 180 181 bool smc_ib_is_valid_local_systemid(void) 182 { 183 return !is_zero_ether_addr(&local_systemid[2]); 184 } 185 186 static void smc_ib_init_local_systemid(void) 187 { 188 get_random_bytes(&local_systemid[0], 2); 189 } 190 191 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport) 192 { 193 return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE; 194 } 195 196 int smc_ib_find_route(struct net *net, __be32 saddr, __be32 daddr, 197 u8 nexthop_mac[], u8 *uses_gateway) 198 { 199 struct neighbour *neigh = NULL; 200 struct rtable *rt = NULL; 201 struct flowi4 fl4 = { 202 .saddr = saddr, 203 .daddr = daddr 204 }; 205 206 if (daddr == cpu_to_be32(INADDR_NONE)) 207 goto out; 208 rt = ip_route_output_flow(net, &fl4, NULL); 209 if (IS_ERR(rt)) 210 goto out; 211 if (rt->rt_uses_gateway && rt->rt_gw_family != AF_INET) 212 goto out_rt; 213 neigh = dst_neigh_lookup(&rt->dst, &fl4.daddr); 214 if (!neigh) 215 goto out_rt; 216 memcpy(nexthop_mac, neigh->ha, ETH_ALEN); 217 *uses_gateway = rt->rt_uses_gateway; 218 neigh_release(neigh); 219 ip_rt_put(rt); 220 return 0; 221 222 out_rt: 223 ip_rt_put(rt); 224 out: 225 return -ENOENT; 226 } 227 228 static int smc_ib_determine_gid_rcu(const struct net_device *ndev, 229 const struct ib_gid_attr *attr, 230 u8 gid[], u8 *sgid_index, 231 struct smc_init_info_smcrv2 *smcrv2) 232 { 233 if (!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) { 234 if (gid) 235 memcpy(gid, &attr->gid, SMC_GID_SIZE); 236 if (sgid_index) 237 *sgid_index = attr->index; 238 return 0; 239 } 240 if (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP && 241 smc_ib_gid_to_ipv4((u8 *)&attr->gid) != cpu_to_be32(INADDR_NONE)) { 242 struct in_device *in_dev = __in_dev_get_rcu(ndev); 243 struct net *net = dev_net(ndev); 244 const struct in_ifaddr *ifa; 245 bool subnet_match = false; 246 247 if (!in_dev) 248 goto out; 249 in_dev_for_each_ifa_rcu(ifa, in_dev) { 250 if (!inet_ifa_match(smcrv2->saddr, ifa)) 251 continue; 252 subnet_match = true; 253 break; 254 } 255 if (!subnet_match) 256 goto out; 257 if (smcrv2->daddr && smc_ib_find_route(net, smcrv2->saddr, 258 smcrv2->daddr, 259 smcrv2->nexthop_mac, 260 &smcrv2->uses_gateway)) 261 goto out; 262 263 if (gid) 264 memcpy(gid, &attr->gid, SMC_GID_SIZE); 265 if (sgid_index) 266 *sgid_index = attr->index; 267 return 0; 268 } 269 out: 270 return -ENODEV; 271 } 272 273 /* determine the gid for an ib-device port and vlan id */ 274 int smc_ib_determine_gid(struct smc_ib_device *smcibdev, u8 ibport, 275 unsigned short vlan_id, u8 gid[], u8 *sgid_index, 276 struct smc_init_info_smcrv2 *smcrv2) 277 { 278 const struct ib_gid_attr *attr; 279 const struct net_device *ndev; 280 int i; 281 282 for (i = 0; i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) { 283 attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i); 284 if (IS_ERR(attr)) 285 continue; 286 287 rcu_read_lock(); 288 ndev = rdma_read_gid_attr_ndev_rcu(attr); 289 if (!IS_ERR(ndev) && 290 ((!vlan_id && !is_vlan_dev(ndev)) || 291 (vlan_id && is_vlan_dev(ndev) && 292 vlan_dev_vlan_id(ndev) == vlan_id))) { 293 if (!smc_ib_determine_gid_rcu(ndev, attr, gid, 294 sgid_index, smcrv2)) { 295 rcu_read_unlock(); 296 rdma_put_gid_attr(attr); 297 return 0; 298 } 299 } 300 rcu_read_unlock(); 301 rdma_put_gid_attr(attr); 302 } 303 return -ENODEV; 304 } 305 306 /* check if gid is still defined on smcibdev */ 307 static bool smc_ib_check_link_gid(u8 gid[SMC_GID_SIZE], bool smcrv2, 308 struct smc_ib_device *smcibdev, u8 ibport) 309 { 310 const struct ib_gid_attr *attr; 311 bool rc = false; 312 int i; 313 314 for (i = 0; !rc && i < smcibdev->pattr[ibport - 1].gid_tbl_len; i++) { 315 attr = rdma_get_gid_attr(smcibdev->ibdev, ibport, i); 316 if (IS_ERR(attr)) 317 continue; 318 319 rcu_read_lock(); 320 if ((!smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE) || 321 (smcrv2 && attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP && 322 !(ipv6_addr_type((const struct in6_addr *)&attr->gid) 323 & IPV6_ADDR_LINKLOCAL))) 324 if (!memcmp(gid, &attr->gid, SMC_GID_SIZE)) 325 rc = true; 326 rcu_read_unlock(); 327 rdma_put_gid_attr(attr); 328 } 329 return rc; 330 } 331 332 /* check all links if the gid is still defined on smcibdev */ 333 static void smc_ib_gid_check(struct smc_ib_device *smcibdev, u8 ibport) 334 { 335 struct smc_link_group *lgr; 336 bool stale_gid = false; 337 int i; 338 339 spin_lock_bh(&smc_lgr_list.lock); 340 list_for_each_entry(lgr, &smc_lgr_list.list, list) { 341 if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id, 342 SMC_MAX_PNETID_LEN)) 343 continue; /* lgr is not affected */ 344 if (list_empty(&lgr->list)) 345 continue; 346 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 347 if (lgr->lnk[i].state == SMC_LNK_UNUSED || 348 lgr->lnk[i].smcibdev != smcibdev) 349 continue; 350 if (!smc_ib_check_link_gid(lgr->lnk[i].gid, 351 lgr->smc_version == SMC_V2, 352 smcibdev, ibport)) { 353 stale_gid = true; 354 goto out; 355 } 356 } 357 } 358 out: 359 spin_unlock_bh(&smc_lgr_list.lock); 360 if (stale_gid) 361 smcr_port_err(smcibdev, ibport); 362 } 363 364 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport) 365 { 366 int rc; 367 368 memset(&smcibdev->pattr[ibport - 1], 0, 369 sizeof(smcibdev->pattr[ibport - 1])); 370 rc = ib_query_port(smcibdev->ibdev, ibport, 371 &smcibdev->pattr[ibport - 1]); 372 if (rc) 373 goto out; 374 /* the SMC protocol requires specification of the RoCE MAC address */ 375 rc = smc_ib_fill_mac(smcibdev, ibport); 376 if (rc) 377 goto out; 378 if (!smc_ib_is_valid_local_systemid() && 379 smc_ib_port_active(smcibdev, ibport)) 380 /* create unique system identifier */ 381 smc_ib_define_local_systemid(smcibdev, ibport); 382 out: 383 return rc; 384 } 385 386 /* process context wrapper for might_sleep smc_ib_remember_port_attr */ 387 static void smc_ib_port_event_work(struct work_struct *work) 388 { 389 struct smc_ib_device *smcibdev = container_of( 390 work, struct smc_ib_device, port_event_work); 391 u8 port_idx; 392 393 for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) { 394 smc_ib_remember_port_attr(smcibdev, port_idx + 1); 395 clear_bit(port_idx, &smcibdev->port_event_mask); 396 if (!smc_ib_port_active(smcibdev, port_idx + 1)) { 397 set_bit(port_idx, smcibdev->ports_going_away); 398 smcr_port_err(smcibdev, port_idx + 1); 399 } else { 400 clear_bit(port_idx, smcibdev->ports_going_away); 401 smcr_port_add(smcibdev, port_idx + 1); 402 smc_ib_gid_check(smcibdev, port_idx + 1); 403 } 404 } 405 } 406 407 /* can be called in IRQ context */ 408 static void smc_ib_global_event_handler(struct ib_event_handler *handler, 409 struct ib_event *ibevent) 410 { 411 struct smc_ib_device *smcibdev; 412 bool schedule = false; 413 u8 port_idx; 414 415 smcibdev = container_of(handler, struct smc_ib_device, event_handler); 416 417 switch (ibevent->event) { 418 case IB_EVENT_DEVICE_FATAL: 419 /* terminate all ports on device */ 420 for (port_idx = 0; port_idx < SMC_MAX_PORTS; port_idx++) { 421 set_bit(port_idx, &smcibdev->port_event_mask); 422 if (!test_and_set_bit(port_idx, 423 smcibdev->ports_going_away)) 424 schedule = true; 425 } 426 if (schedule) 427 schedule_work(&smcibdev->port_event_work); 428 break; 429 case IB_EVENT_PORT_ACTIVE: 430 port_idx = ibevent->element.port_num - 1; 431 if (port_idx >= SMC_MAX_PORTS) 432 break; 433 set_bit(port_idx, &smcibdev->port_event_mask); 434 if (test_and_clear_bit(port_idx, smcibdev->ports_going_away)) 435 schedule_work(&smcibdev->port_event_work); 436 break; 437 case IB_EVENT_PORT_ERR: 438 port_idx = ibevent->element.port_num - 1; 439 if (port_idx >= SMC_MAX_PORTS) 440 break; 441 set_bit(port_idx, &smcibdev->port_event_mask); 442 if (!test_and_set_bit(port_idx, smcibdev->ports_going_away)) 443 schedule_work(&smcibdev->port_event_work); 444 break; 445 case IB_EVENT_GID_CHANGE: 446 port_idx = ibevent->element.port_num - 1; 447 if (port_idx >= SMC_MAX_PORTS) 448 break; 449 set_bit(port_idx, &smcibdev->port_event_mask); 450 schedule_work(&smcibdev->port_event_work); 451 break; 452 default: 453 break; 454 } 455 } 456 457 void smc_ib_dealloc_protection_domain(struct smc_link *lnk) 458 { 459 if (lnk->roce_pd) 460 ib_dealloc_pd(lnk->roce_pd); 461 lnk->roce_pd = NULL; 462 } 463 464 int smc_ib_create_protection_domain(struct smc_link *lnk) 465 { 466 int rc; 467 468 lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0); 469 rc = PTR_ERR_OR_ZERO(lnk->roce_pd); 470 if (IS_ERR(lnk->roce_pd)) 471 lnk->roce_pd = NULL; 472 return rc; 473 } 474 475 static bool smcr_diag_is_dev_critical(struct smc_lgr_list *smc_lgr, 476 struct smc_ib_device *smcibdev) 477 { 478 struct smc_link_group *lgr; 479 bool rc = false; 480 int i; 481 482 spin_lock_bh(&smc_lgr->lock); 483 list_for_each_entry(lgr, &smc_lgr->list, list) { 484 if (lgr->is_smcd) 485 continue; 486 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) { 487 if (lgr->lnk[i].state == SMC_LNK_UNUSED || 488 lgr->lnk[i].smcibdev != smcibdev) 489 continue; 490 if (lgr->type == SMC_LGR_SINGLE || 491 lgr->type == SMC_LGR_ASYMMETRIC_LOCAL) { 492 rc = true; 493 goto out; 494 } 495 } 496 } 497 out: 498 spin_unlock_bh(&smc_lgr->lock); 499 return rc; 500 } 501 502 static int smc_nl_handle_dev_port(struct sk_buff *skb, 503 struct ib_device *ibdev, 504 struct smc_ib_device *smcibdev, 505 int port) 506 { 507 char smc_pnet[SMC_MAX_PNETID_LEN + 1]; 508 struct nlattr *port_attrs; 509 unsigned char port_state; 510 int lnk_count = 0; 511 512 port_attrs = nla_nest_start(skb, SMC_NLA_DEV_PORT + port); 513 if (!port_attrs) 514 goto errout; 515 516 if (nla_put_u8(skb, SMC_NLA_DEV_PORT_PNET_USR, 517 smcibdev->pnetid_by_user[port])) 518 goto errattr; 519 memcpy(smc_pnet, &smcibdev->pnetid[port], SMC_MAX_PNETID_LEN); 520 smc_pnet[SMC_MAX_PNETID_LEN] = 0; 521 if (nla_put_string(skb, SMC_NLA_DEV_PORT_PNETID, smc_pnet)) 522 goto errattr; 523 if (nla_put_u32(skb, SMC_NLA_DEV_PORT_NETDEV, 524 smcibdev->ndev_ifidx[port])) 525 goto errattr; 526 if (nla_put_u8(skb, SMC_NLA_DEV_PORT_VALID, 1)) 527 goto errattr; 528 port_state = smc_ib_port_active(smcibdev, port + 1); 529 if (nla_put_u8(skb, SMC_NLA_DEV_PORT_STATE, port_state)) 530 goto errattr; 531 lnk_count = atomic_read(&smcibdev->lnk_cnt_by_port[port]); 532 if (nla_put_u32(skb, SMC_NLA_DEV_PORT_LNK_CNT, lnk_count)) 533 goto errattr; 534 nla_nest_end(skb, port_attrs); 535 return 0; 536 errattr: 537 nla_nest_cancel(skb, port_attrs); 538 errout: 539 return -EMSGSIZE; 540 } 541 542 static bool smc_nl_handle_pci_values(const struct smc_pci_dev *smc_pci_dev, 543 struct sk_buff *skb) 544 { 545 if (nla_put_u32(skb, SMC_NLA_DEV_PCI_FID, smc_pci_dev->pci_fid)) 546 return false; 547 if (nla_put_u16(skb, SMC_NLA_DEV_PCI_CHID, smc_pci_dev->pci_pchid)) 548 return false; 549 if (nla_put_u16(skb, SMC_NLA_DEV_PCI_VENDOR, smc_pci_dev->pci_vendor)) 550 return false; 551 if (nla_put_u16(skb, SMC_NLA_DEV_PCI_DEVICE, smc_pci_dev->pci_device)) 552 return false; 553 if (nla_put_string(skb, SMC_NLA_DEV_PCI_ID, smc_pci_dev->pci_id)) 554 return false; 555 return true; 556 } 557 558 static int smc_nl_handle_smcr_dev(struct smc_ib_device *smcibdev, 559 struct sk_buff *skb, 560 struct netlink_callback *cb) 561 { 562 char smc_ibname[IB_DEVICE_NAME_MAX]; 563 struct smc_pci_dev smc_pci_dev; 564 struct pci_dev *pci_dev; 565 unsigned char is_crit; 566 struct nlattr *attrs; 567 void *nlh; 568 int i; 569 570 nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 571 &smc_gen_nl_family, NLM_F_MULTI, 572 SMC_NETLINK_GET_DEV_SMCR); 573 if (!nlh) 574 goto errmsg; 575 attrs = nla_nest_start(skb, SMC_GEN_DEV_SMCR); 576 if (!attrs) 577 goto errout; 578 is_crit = smcr_diag_is_dev_critical(&smc_lgr_list, smcibdev); 579 if (nla_put_u8(skb, SMC_NLA_DEV_IS_CRIT, is_crit)) 580 goto errattr; 581 if (smcibdev->ibdev->dev.parent) { 582 memset(&smc_pci_dev, 0, sizeof(smc_pci_dev)); 583 pci_dev = to_pci_dev(smcibdev->ibdev->dev.parent); 584 smc_set_pci_values(pci_dev, &smc_pci_dev); 585 if (!smc_nl_handle_pci_values(&smc_pci_dev, skb)) 586 goto errattr; 587 } 588 snprintf(smc_ibname, sizeof(smc_ibname), "%s", smcibdev->ibdev->name); 589 if (nla_put_string(skb, SMC_NLA_DEV_IB_NAME, smc_ibname)) 590 goto errattr; 591 for (i = 1; i <= SMC_MAX_PORTS; i++) { 592 if (!rdma_is_port_valid(smcibdev->ibdev, i)) 593 continue; 594 if (smc_nl_handle_dev_port(skb, smcibdev->ibdev, 595 smcibdev, i - 1)) 596 goto errattr; 597 } 598 599 nla_nest_end(skb, attrs); 600 genlmsg_end(skb, nlh); 601 return 0; 602 603 errattr: 604 nla_nest_cancel(skb, attrs); 605 errout: 606 genlmsg_cancel(skb, nlh); 607 errmsg: 608 return -EMSGSIZE; 609 } 610 611 static void smc_nl_prep_smcr_dev(struct smc_ib_devices *dev_list, 612 struct sk_buff *skb, 613 struct netlink_callback *cb) 614 { 615 struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb); 616 struct smc_ib_device *smcibdev; 617 int snum = cb_ctx->pos[0]; 618 int num = 0; 619 620 mutex_lock(&dev_list->mutex); 621 list_for_each_entry(smcibdev, &dev_list->list, list) { 622 if (num < snum) 623 goto next; 624 if (smc_nl_handle_smcr_dev(smcibdev, skb, cb)) 625 goto errout; 626 next: 627 num++; 628 } 629 errout: 630 mutex_unlock(&dev_list->mutex); 631 cb_ctx->pos[0] = num; 632 } 633 634 int smcr_nl_get_device(struct sk_buff *skb, struct netlink_callback *cb) 635 { 636 smc_nl_prep_smcr_dev(&smc_ib_devices, skb, cb); 637 return skb->len; 638 } 639 640 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv) 641 { 642 struct smc_link *lnk = (struct smc_link *)priv; 643 struct smc_ib_device *smcibdev = lnk->smcibdev; 644 u8 port_idx; 645 646 switch (ibevent->event) { 647 case IB_EVENT_QP_FATAL: 648 case IB_EVENT_QP_ACCESS_ERR: 649 port_idx = ibevent->element.qp->port - 1; 650 if (port_idx >= SMC_MAX_PORTS) 651 break; 652 set_bit(port_idx, &smcibdev->port_event_mask); 653 if (!test_and_set_bit(port_idx, smcibdev->ports_going_away)) 654 schedule_work(&smcibdev->port_event_work); 655 break; 656 default: 657 break; 658 } 659 } 660 661 void smc_ib_destroy_queue_pair(struct smc_link *lnk) 662 { 663 if (lnk->roce_qp) 664 ib_destroy_qp(lnk->roce_qp); 665 lnk->roce_qp = NULL; 666 } 667 668 /* create a queue pair within the protection domain for a link */ 669 int smc_ib_create_queue_pair(struct smc_link *lnk) 670 { 671 struct ib_qp_init_attr qp_attr = { 672 .event_handler = smc_ib_qp_event_handler, 673 .qp_context = lnk, 674 .send_cq = lnk->smcibdev->roce_cq_send, 675 .recv_cq = lnk->smcibdev->roce_cq_recv, 676 .srq = NULL, 677 .cap = { 678 .max_send_sge = SMC_IB_MAX_SEND_SGE, 679 .max_recv_sge = lnk->wr_rx_sge_cnt, 680 .max_inline_data = 0, 681 }, 682 .sq_sig_type = IB_SIGNAL_REQ_WR, 683 .qp_type = IB_QPT_RC, 684 }; 685 int rc; 686 687 /* include unsolicited rdma_writes as well, 688 * there are max. 2 RDMA_WRITE per 1 WR_SEND 689 */ 690 qp_attr.cap.max_send_wr = 3 * lnk->lgr->max_send_wr; 691 qp_attr.cap.max_recv_wr = lnk->lgr->max_recv_wr; 692 lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr); 693 rc = PTR_ERR_OR_ZERO(lnk->roce_qp); 694 if (IS_ERR(lnk->roce_qp)) 695 lnk->roce_qp = NULL; 696 else 697 smc_wr_remember_qp_attr(lnk); 698 return rc; 699 } 700 701 void smc_ib_put_memory_region(struct ib_mr *mr) 702 { 703 ib_dereg_mr(mr); 704 } 705 706 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot, u8 link_idx) 707 { 708 unsigned int offset = 0; 709 int sg_num; 710 711 /* map the largest prefix of a dma mapped SG list */ 712 sg_num = ib_map_mr_sg(buf_slot->mr[link_idx], 713 buf_slot->sgt[link_idx].sgl, 714 buf_slot->sgt[link_idx].orig_nents, 715 &offset, PAGE_SIZE); 716 717 return sg_num; 718 } 719 720 /* Allocate a memory region and map the dma mapped SG list of buf_slot */ 721 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags, 722 struct smc_buf_desc *buf_slot, u8 link_idx) 723 { 724 if (buf_slot->mr[link_idx]) 725 return 0; /* already done */ 726 727 buf_slot->mr[link_idx] = 728 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order); 729 if (IS_ERR(buf_slot->mr[link_idx])) { 730 int rc; 731 732 rc = PTR_ERR(buf_slot->mr[link_idx]); 733 buf_slot->mr[link_idx] = NULL; 734 return rc; 735 } 736 737 if (smc_ib_map_mr_sg(buf_slot, link_idx) != 738 buf_slot->sgt[link_idx].orig_nents) 739 return -EINVAL; 740 741 return 0; 742 } 743 744 bool smc_ib_is_sg_need_sync(struct smc_link *lnk, 745 struct smc_buf_desc *buf_slot) 746 { 747 struct scatterlist *sg; 748 unsigned int i; 749 bool ret = false; 750 751 if (!lnk->smcibdev->ibdev->dma_device) 752 return ret; 753 754 /* for now there is just one DMA address */ 755 for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg, 756 buf_slot->sgt[lnk->link_idx].nents, i) { 757 if (!sg_dma_len(sg)) 758 break; 759 if (dma_need_sync(lnk->smcibdev->ibdev->dma_device, 760 sg_dma_address(sg))) { 761 ret = true; 762 goto out; 763 } 764 } 765 766 out: 767 return ret; 768 } 769 770 /* synchronize buffer usage for cpu access */ 771 void smc_ib_sync_sg_for_cpu(struct smc_link *lnk, 772 struct smc_buf_desc *buf_slot, 773 enum dma_data_direction data_direction) 774 { 775 struct scatterlist *sg; 776 unsigned int i; 777 778 if (!(buf_slot->is_dma_need_sync & (1U << lnk->link_idx))) 779 return; 780 781 /* for now there is just one DMA address */ 782 for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg, 783 buf_slot->sgt[lnk->link_idx].nents, i) { 784 if (!sg_dma_len(sg)) 785 break; 786 ib_dma_sync_single_for_cpu(lnk->smcibdev->ibdev, 787 sg_dma_address(sg), 788 sg_dma_len(sg), 789 data_direction); 790 } 791 } 792 793 /* synchronize buffer usage for device access */ 794 void smc_ib_sync_sg_for_device(struct smc_link *lnk, 795 struct smc_buf_desc *buf_slot, 796 enum dma_data_direction data_direction) 797 { 798 struct scatterlist *sg; 799 unsigned int i; 800 801 if (!(buf_slot->is_dma_need_sync & (1U << lnk->link_idx))) 802 return; 803 804 /* for now there is just one DMA address */ 805 for_each_sg(buf_slot->sgt[lnk->link_idx].sgl, sg, 806 buf_slot->sgt[lnk->link_idx].nents, i) { 807 if (!sg_dma_len(sg)) 808 break; 809 ib_dma_sync_single_for_device(lnk->smcibdev->ibdev, 810 sg_dma_address(sg), 811 sg_dma_len(sg), 812 data_direction); 813 } 814 } 815 816 /* Map a new TX or RX buffer SG-table to DMA */ 817 int smc_ib_buf_map_sg(struct smc_link *lnk, 818 struct smc_buf_desc *buf_slot, 819 enum dma_data_direction data_direction) 820 { 821 int mapped_nents; 822 823 mapped_nents = ib_dma_map_sg(lnk->smcibdev->ibdev, 824 buf_slot->sgt[lnk->link_idx].sgl, 825 buf_slot->sgt[lnk->link_idx].orig_nents, 826 data_direction); 827 if (!mapped_nents) 828 return -ENOMEM; 829 830 return mapped_nents; 831 } 832 833 void smc_ib_buf_unmap_sg(struct smc_link *lnk, 834 struct smc_buf_desc *buf_slot, 835 enum dma_data_direction data_direction) 836 { 837 if (!buf_slot->sgt[lnk->link_idx].sgl->dma_address) 838 return; /* already unmapped */ 839 840 ib_dma_unmap_sg(lnk->smcibdev->ibdev, 841 buf_slot->sgt[lnk->link_idx].sgl, 842 buf_slot->sgt[lnk->link_idx].orig_nents, 843 data_direction); 844 buf_slot->sgt[lnk->link_idx].sgl->dma_address = 0; 845 } 846 847 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev) 848 { 849 struct ib_cq_init_attr cqattr = { 850 .cqe = SMC_MAX_CQE, .comp_vector = 0 }; 851 int cqe_size_order, smc_order; 852 long rc; 853 854 mutex_lock(&smcibdev->mutex); 855 rc = 0; 856 if (smcibdev->initialized) 857 goto out; 858 /* the calculated number of cq entries fits to mlx5 cq allocation */ 859 cqe_size_order = cache_line_size() == 128 ? 7 : 6; 860 smc_order = MAX_PAGE_ORDER - cqe_size_order; 861 if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE) 862 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2; 863 smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev, 864 smc_wr_tx_cq_handler, NULL, 865 smcibdev, &cqattr); 866 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send); 867 if (IS_ERR(smcibdev->roce_cq_send)) { 868 smcibdev->roce_cq_send = NULL; 869 goto out; 870 } 871 smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev, 872 smc_wr_rx_cq_handler, NULL, 873 smcibdev, &cqattr); 874 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv); 875 if (IS_ERR(smcibdev->roce_cq_recv)) { 876 smcibdev->roce_cq_recv = NULL; 877 goto err; 878 } 879 smc_wr_add_dev(smcibdev); 880 smcibdev->initialized = 1; 881 goto out; 882 883 err: 884 ib_destroy_cq(smcibdev->roce_cq_send); 885 out: 886 mutex_unlock(&smcibdev->mutex); 887 return rc; 888 } 889 890 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev) 891 { 892 mutex_lock(&smcibdev->mutex); 893 if (!smcibdev->initialized) 894 goto out; 895 smcibdev->initialized = 0; 896 ib_destroy_cq(smcibdev->roce_cq_recv); 897 ib_destroy_cq(smcibdev->roce_cq_send); 898 smc_wr_remove_dev(smcibdev); 899 out: 900 mutex_unlock(&smcibdev->mutex); 901 } 902 903 static struct ib_client smc_ib_client; 904 905 static void smc_copy_netdev_ifindex(struct smc_ib_device *smcibdev, int port) 906 { 907 struct ib_device *ibdev = smcibdev->ibdev; 908 struct net_device *ndev; 909 910 ndev = ib_device_get_netdev(ibdev, port + 1); 911 if (ndev) { 912 smcibdev->ndev_ifidx[port] = ndev->ifindex; 913 dev_put(ndev); 914 } 915 } 916 917 void smc_ib_ndev_change(struct net_device *ndev, unsigned long event) 918 { 919 struct smc_ib_device *smcibdev; 920 struct ib_device *libdev; 921 struct net_device *lndev; 922 u8 port_cnt; 923 int i; 924 925 mutex_lock(&smc_ib_devices.mutex); 926 list_for_each_entry(smcibdev, &smc_ib_devices.list, list) { 927 port_cnt = smcibdev->ibdev->phys_port_cnt; 928 for (i = 0; i < min_t(size_t, port_cnt, SMC_MAX_PORTS); i++) { 929 libdev = smcibdev->ibdev; 930 lndev = ib_device_get_netdev(libdev, i + 1); 931 dev_put(lndev); 932 if (lndev != ndev) 933 continue; 934 if (event == NETDEV_REGISTER) 935 smcibdev->ndev_ifidx[i] = ndev->ifindex; 936 if (event == NETDEV_UNREGISTER) 937 smcibdev->ndev_ifidx[i] = 0; 938 } 939 } 940 mutex_unlock(&smc_ib_devices.mutex); 941 } 942 943 /* callback function for ib_register_client() */ 944 static int smc_ib_add_dev(struct ib_device *ibdev) 945 { 946 struct smc_ib_device *smcibdev; 947 u8 port_cnt; 948 int i; 949 950 if (ibdev->node_type != RDMA_NODE_IB_CA) 951 return -EOPNOTSUPP; 952 953 smcibdev = kzalloc_obj(*smcibdev); 954 if (!smcibdev) 955 return -ENOMEM; 956 957 smcibdev->ibdev = ibdev; 958 INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work); 959 atomic_set(&smcibdev->lnk_cnt, 0); 960 init_waitqueue_head(&smcibdev->lnks_deleted); 961 mutex_init(&smcibdev->mutex); 962 mutex_lock(&smc_ib_devices.mutex); 963 list_add_tail(&smcibdev->list, &smc_ib_devices.list); 964 mutex_unlock(&smc_ib_devices.mutex); 965 ib_set_client_data(ibdev, &smc_ib_client, smcibdev); 966 INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev, 967 smc_ib_global_event_handler); 968 ib_register_event_handler(&smcibdev->event_handler); 969 970 /* trigger reading of the port attributes */ 971 port_cnt = smcibdev->ibdev->phys_port_cnt; 972 pr_warn_ratelimited("smc: adding ib device %s with port count %d\n", 973 smcibdev->ibdev->name, port_cnt); 974 for (i = 0; 975 i < min_t(size_t, port_cnt, SMC_MAX_PORTS); 976 i++) { 977 set_bit(i, &smcibdev->port_event_mask); 978 /* determine pnetids of the port */ 979 if (smc_pnetid_by_dev_port(ibdev->dev.parent, i, 980 smcibdev->pnetid[i])) 981 smc_pnetid_by_table_ib(smcibdev, i + 1); 982 smc_copy_netdev_ifindex(smcibdev, i); 983 if (smc_pnet_is_pnetid_set(smcibdev->pnetid[i])) 984 pr_warn_ratelimited("smc: ib device %s port %d has pnetid %.16s%s\n", 985 smcibdev->ibdev->name, i + 1, 986 smcibdev->pnetid[i], 987 smcibdev->pnetid_by_user[i] ? 988 " (user defined)" : 989 ""); 990 else 991 pr_warn_ratelimited("smc: ib device %s port %d has no pnetid\n", 992 smcibdev->ibdev->name, i + 1); 993 994 } 995 schedule_work(&smcibdev->port_event_work); 996 return 0; 997 } 998 999 /* callback function for ib_unregister_client() */ 1000 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data) 1001 { 1002 struct smc_ib_device *smcibdev = client_data; 1003 1004 mutex_lock(&smc_ib_devices.mutex); 1005 list_del_init(&smcibdev->list); /* remove from smc_ib_devices */ 1006 mutex_unlock(&smc_ib_devices.mutex); 1007 pr_warn_ratelimited("smc: removing ib device %s\n", 1008 smcibdev->ibdev->name); 1009 smc_smcr_terminate_all(smcibdev); 1010 smc_ib_cleanup_per_ibdev(smcibdev); 1011 ib_unregister_event_handler(&smcibdev->event_handler); 1012 cancel_work_sync(&smcibdev->port_event_work); 1013 kfree(smcibdev); 1014 } 1015 1016 static struct ib_client smc_ib_client = { 1017 .name = "smc_ib", 1018 .add = smc_ib_add_dev, 1019 .remove = smc_ib_remove_dev, 1020 }; 1021 1022 int __init smc_ib_register_client(void) 1023 { 1024 smc_ib_init_local_systemid(); 1025 return ib_register_client(&smc_ib_client); 1026 } 1027 1028 void smc_ib_unregister_client(void) 1029 { 1030 ib_unregister_client(&smc_ib_client); 1031 } 1032