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 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); 182 183 switch (ibevent->event) { 184 case IB_EVENT_PORT_ERR: 185 port_idx = ibevent->element.port_num - 1; 186 set_bit(port_idx, &smcibdev->port_event_mask); 187 schedule_work(&smcibdev->port_event_work); 188 /* fall through */ 189 case IB_EVENT_DEVICE_FATAL: 190 /* tbd in follow-on patch: 191 * abnormal close of corresponding connections 192 */ 193 break; 194 case IB_EVENT_PORT_ACTIVE: 195 port_idx = ibevent->element.port_num - 1; 196 set_bit(port_idx, &smcibdev->port_event_mask); 197 schedule_work(&smcibdev->port_event_work); 198 break; 199 default: 200 break; 201 } 202 } 203 204 void smc_ib_dealloc_protection_domain(struct smc_link *lnk) 205 { 206 ib_dealloc_pd(lnk->roce_pd); 207 lnk->roce_pd = NULL; 208 } 209 210 int smc_ib_create_protection_domain(struct smc_link *lnk) 211 { 212 int rc; 213 214 lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0); 215 rc = PTR_ERR_OR_ZERO(lnk->roce_pd); 216 if (IS_ERR(lnk->roce_pd)) 217 lnk->roce_pd = NULL; 218 return rc; 219 } 220 221 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv) 222 { 223 switch (ibevent->event) { 224 case IB_EVENT_DEVICE_FATAL: 225 case IB_EVENT_GID_CHANGE: 226 case IB_EVENT_PORT_ERR: 227 case IB_EVENT_QP_ACCESS_ERR: 228 /* tbd in follow-on patch: 229 * abnormal close of corresponding connections 230 */ 231 break; 232 default: 233 break; 234 } 235 } 236 237 void smc_ib_destroy_queue_pair(struct smc_link *lnk) 238 { 239 ib_destroy_qp(lnk->roce_qp); 240 lnk->roce_qp = NULL; 241 } 242 243 /* create a queue pair within the protection domain for a link */ 244 int smc_ib_create_queue_pair(struct smc_link *lnk) 245 { 246 struct ib_qp_init_attr qp_attr = { 247 .event_handler = smc_ib_qp_event_handler, 248 .qp_context = lnk, 249 .send_cq = lnk->smcibdev->roce_cq_send, 250 .recv_cq = lnk->smcibdev->roce_cq_recv, 251 .srq = NULL, 252 .cap = { 253 .max_send_wr = SMC_WR_BUF_CNT, 254 /* include unsolicited rdma_writes as well, 255 * there are max. 2 RDMA_WRITE per 1 WR_SEND 256 */ 257 .max_recv_wr = SMC_WR_BUF_CNT * 3, 258 .max_send_sge = SMC_IB_MAX_SEND_SGE, 259 .max_recv_sge = 1, 260 }, 261 .sq_sig_type = IB_SIGNAL_REQ_WR, 262 .qp_type = IB_QPT_RC, 263 }; 264 int rc; 265 266 lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr); 267 rc = PTR_ERR_OR_ZERO(lnk->roce_qp); 268 if (IS_ERR(lnk->roce_qp)) 269 lnk->roce_qp = NULL; 270 else 271 smc_wr_remember_qp_attr(lnk); 272 return rc; 273 } 274 275 /* map a new TX or RX buffer to DMA */ 276 int smc_ib_buf_map(struct smc_ib_device *smcibdev, int buf_size, 277 struct smc_buf_desc *buf_slot, 278 enum dma_data_direction data_direction) 279 { 280 int rc = 0; 281 282 if (buf_slot->dma_addr[SMC_SINGLE_LINK]) 283 return rc; /* already mapped */ 284 buf_slot->dma_addr[SMC_SINGLE_LINK] = 285 ib_dma_map_single(smcibdev->ibdev, buf_slot->cpu_addr, 286 buf_size, data_direction); 287 if (ib_dma_mapping_error(smcibdev->ibdev, 288 buf_slot->dma_addr[SMC_SINGLE_LINK])) 289 rc = -EIO; 290 return rc; 291 } 292 293 void smc_ib_buf_unmap(struct smc_ib_device *smcibdev, int buf_size, 294 struct smc_buf_desc *buf_slot, 295 enum dma_data_direction data_direction) 296 { 297 if (!buf_slot->dma_addr[SMC_SINGLE_LINK]) 298 return; /* already unmapped */ 299 ib_dma_unmap_single(smcibdev->ibdev, *buf_slot->dma_addr, buf_size, 300 data_direction); 301 buf_slot->dma_addr[SMC_SINGLE_LINK] = 0; 302 } 303 304 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport) 305 { 306 struct net_device *ndev; 307 int rc; 308 309 rc = ib_query_gid(smcibdev->ibdev, ibport, 0, 310 &smcibdev->gid[ibport - 1], NULL); 311 /* the SMC protocol requires specification of the roce MAC address; 312 * if net_device cannot be determined, it can be derived from gid 0 313 */ 314 ndev = smcibdev->ibdev->get_netdev(smcibdev->ibdev, ibport); 315 if (ndev) { 316 memcpy(&smcibdev->mac, ndev->dev_addr, ETH_ALEN); 317 } else if (!rc) { 318 memcpy(&smcibdev->mac[ibport - 1][0], 319 &smcibdev->gid[ibport - 1].raw[8], 3); 320 memcpy(&smcibdev->mac[ibport - 1][3], 321 &smcibdev->gid[ibport - 1].raw[13], 3); 322 smcibdev->mac[ibport - 1][0] &= ~0x02; 323 } 324 return rc; 325 } 326 327 /* Create an identifier unique for this instance of SMC-R. 328 * The MAC-address of the first active registered IB device 329 * plus a random 2-byte number is used to create this identifier. 330 * This name is delivered to the peer during connection initialization. 331 */ 332 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev, 333 u8 ibport) 334 { 335 memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1], 336 sizeof(smcibdev->mac[ibport - 1])); 337 get_random_bytes(&local_systemid[0], 2); 338 } 339 340 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport) 341 { 342 return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE; 343 } 344 345 int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport) 346 { 347 int rc; 348 349 memset(&smcibdev->pattr[ibport - 1], 0, 350 sizeof(smcibdev->pattr[ibport - 1])); 351 rc = ib_query_port(smcibdev->ibdev, ibport, 352 &smcibdev->pattr[ibport - 1]); 353 if (rc) 354 goto out; 355 rc = smc_ib_fill_gid_and_mac(smcibdev, ibport); 356 if (rc) 357 goto out; 358 if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET, 359 sizeof(local_systemid)) && 360 smc_ib_port_active(smcibdev, ibport)) 361 /* create unique system identifier */ 362 smc_ib_define_local_systemid(smcibdev, ibport); 363 out: 364 return rc; 365 } 366 367 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev) 368 { 369 struct ib_cq_init_attr cqattr = { 370 .cqe = SMC_WR_MAX_CQE, .comp_vector = 0 }; 371 long rc; 372 373 smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev, 374 smc_wr_tx_cq_handler, NULL, 375 smcibdev, &cqattr); 376 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send); 377 if (IS_ERR(smcibdev->roce_cq_send)) { 378 smcibdev->roce_cq_send = NULL; 379 return rc; 380 } 381 smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev, 382 smc_wr_rx_cq_handler, NULL, 383 smcibdev, &cqattr); 384 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv); 385 if (IS_ERR(smcibdev->roce_cq_recv)) { 386 smcibdev->roce_cq_recv = NULL; 387 goto err; 388 } 389 INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev, 390 smc_ib_global_event_handler); 391 ib_register_event_handler(&smcibdev->event_handler); 392 smc_wr_add_dev(smcibdev); 393 smcibdev->initialized = 1; 394 return rc; 395 396 err: 397 ib_destroy_cq(smcibdev->roce_cq_send); 398 return rc; 399 } 400 401 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev) 402 { 403 if (!smcibdev->initialized) 404 return; 405 smc_wr_remove_dev(smcibdev); 406 ib_unregister_event_handler(&smcibdev->event_handler); 407 ib_destroy_cq(smcibdev->roce_cq_recv); 408 ib_destroy_cq(smcibdev->roce_cq_send); 409 } 410 411 static struct ib_client smc_ib_client; 412 413 /* callback function for ib_register_client() */ 414 static void smc_ib_add_dev(struct ib_device *ibdev) 415 { 416 struct smc_ib_device *smcibdev; 417 418 if (ibdev->node_type != RDMA_NODE_IB_CA) 419 return; 420 421 smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL); 422 if (!smcibdev) 423 return; 424 425 smcibdev->ibdev = ibdev; 426 INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work); 427 428 spin_lock(&smc_ib_devices.lock); 429 list_add_tail(&smcibdev->list, &smc_ib_devices.list); 430 spin_unlock(&smc_ib_devices.lock); 431 ib_set_client_data(ibdev, &smc_ib_client, smcibdev); 432 } 433 434 /* callback function for ib_register_client() */ 435 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data) 436 { 437 struct smc_ib_device *smcibdev; 438 439 smcibdev = ib_get_client_data(ibdev, &smc_ib_client); 440 ib_set_client_data(ibdev, &smc_ib_client, NULL); 441 spin_lock(&smc_ib_devices.lock); 442 list_del_init(&smcibdev->list); /* remove from smc_ib_devices */ 443 spin_unlock(&smc_ib_devices.lock); 444 smc_pnet_remove_by_ibdev(smcibdev); 445 smc_ib_cleanup_per_ibdev(smcibdev); 446 kfree(smcibdev); 447 } 448 449 static struct ib_client smc_ib_client = { 450 .name = "smc_ib", 451 .add = smc_ib_add_dev, 452 .remove = smc_ib_remove_dev, 453 }; 454 455 int __init smc_ib_register_client(void) 456 { 457 return ib_register_client(&smc_ib_client); 458 } 459 460 void smc_ib_unregister_client(void) 461 { 462 ib_unregister_client(&smc_ib_client); 463 } 464