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 <linux/scatterlist.h> 17 #include <rdma/ib_verbs.h> 18 19 #include "smc_pnet.h" 20 #include "smc_ib.h" 21 #include "smc_core.h" 22 #include "smc_wr.h" 23 #include "smc.h" 24 25 #define SMC_QP_MIN_RNR_TIMER 5 26 #define SMC_QP_TIMEOUT 15 /* 4096 * 2 ** timeout usec */ 27 #define SMC_QP_RETRY_CNT 7 /* 7: infinite */ 28 #define SMC_QP_RNR_RETRY 7 /* 7: infinite */ 29 30 struct smc_ib_devices smc_ib_devices = { /* smc-registered ib devices */ 31 .lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock), 32 .list = LIST_HEAD_INIT(smc_ib_devices.list), 33 }; 34 35 #define SMC_LOCAL_SYSTEMID_RESET "%%%%%%%" 36 37 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET; /* unique system 38 * identifier 39 */ 40 41 static int smc_ib_modify_qp_init(struct smc_link *lnk) 42 { 43 struct ib_qp_attr qp_attr; 44 45 memset(&qp_attr, 0, sizeof(qp_attr)); 46 qp_attr.qp_state = IB_QPS_INIT; 47 qp_attr.pkey_index = 0; 48 qp_attr.port_num = lnk->ibport; 49 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE 50 | IB_ACCESS_REMOTE_WRITE; 51 return ib_modify_qp(lnk->roce_qp, &qp_attr, 52 IB_QP_STATE | IB_QP_PKEY_INDEX | 53 IB_QP_ACCESS_FLAGS | IB_QP_PORT); 54 } 55 56 static int smc_ib_modify_qp_rtr(struct smc_link *lnk) 57 { 58 enum ib_qp_attr_mask qp_attr_mask = 59 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN | 60 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER; 61 struct ib_qp_attr qp_attr; 62 63 memset(&qp_attr, 0, sizeof(qp_attr)); 64 qp_attr.qp_state = IB_QPS_RTR; 65 qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu); 66 qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE; 67 rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport); 68 rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, 0, 1, 0); 69 rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid); 70 memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac, 71 sizeof(lnk->peer_mac)); 72 qp_attr.dest_qp_num = lnk->peer_qpn; 73 qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */ 74 qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming 75 * requests 76 */ 77 qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER; 78 79 return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask); 80 } 81 82 int smc_ib_modify_qp_rts(struct smc_link *lnk) 83 { 84 struct ib_qp_attr qp_attr; 85 86 memset(&qp_attr, 0, sizeof(qp_attr)); 87 qp_attr.qp_state = IB_QPS_RTS; 88 qp_attr.timeout = SMC_QP_TIMEOUT; /* local ack timeout */ 89 qp_attr.retry_cnt = SMC_QP_RETRY_CNT; /* retry count */ 90 qp_attr.rnr_retry = SMC_QP_RNR_RETRY; /* RNR retries, 7=infinite */ 91 qp_attr.sq_psn = lnk->psn_initial; /* starting send packet seq # */ 92 qp_attr.max_rd_atomic = 1; /* # of outstanding RDMA reads and 93 * atomic ops allowed 94 */ 95 return ib_modify_qp(lnk->roce_qp, &qp_attr, 96 IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT | 97 IB_QP_SQ_PSN | IB_QP_RNR_RETRY | 98 IB_QP_MAX_QP_RD_ATOMIC); 99 } 100 101 int smc_ib_modify_qp_reset(struct smc_link *lnk) 102 { 103 struct ib_qp_attr qp_attr; 104 105 memset(&qp_attr, 0, sizeof(qp_attr)); 106 qp_attr.qp_state = IB_QPS_RESET; 107 return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE); 108 } 109 110 int smc_ib_ready_link(struct smc_link *lnk) 111 { 112 struct smc_link_group *lgr = 113 container_of(lnk, struct smc_link_group, lnk[0]); 114 int rc = 0; 115 116 rc = smc_ib_modify_qp_init(lnk); 117 if (rc) 118 goto out; 119 120 rc = smc_ib_modify_qp_rtr(lnk); 121 if (rc) 122 goto out; 123 smc_wr_remember_qp_attr(lnk); 124 rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv, 125 IB_CQ_SOLICITED_MASK); 126 if (rc) 127 goto out; 128 rc = smc_wr_rx_post_init(lnk); 129 if (rc) 130 goto out; 131 smc_wr_remember_qp_attr(lnk); 132 133 if (lgr->role == SMC_SERV) { 134 rc = smc_ib_modify_qp_rts(lnk); 135 if (rc) 136 goto out; 137 smc_wr_remember_qp_attr(lnk); 138 } 139 out: 140 return rc; 141 } 142 143 /* process context wrapper for might_sleep smc_ib_remember_port_attr */ 144 static void smc_ib_port_event_work(struct work_struct *work) 145 { 146 struct smc_ib_device *smcibdev = container_of( 147 work, struct smc_ib_device, port_event_work); 148 u8 port_idx; 149 150 for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) { 151 smc_ib_remember_port_attr(smcibdev, port_idx + 1); 152 clear_bit(port_idx, &smcibdev->port_event_mask); 153 } 154 } 155 156 /* can be called in IRQ context */ 157 static void smc_ib_global_event_handler(struct ib_event_handler *handler, 158 struct ib_event *ibevent) 159 { 160 struct smc_ib_device *smcibdev; 161 u8 port_idx; 162 163 smcibdev = container_of(handler, struct smc_ib_device, event_handler); 164 165 switch (ibevent->event) { 166 case IB_EVENT_PORT_ERR: 167 port_idx = ibevent->element.port_num - 1; 168 set_bit(port_idx, &smcibdev->port_event_mask); 169 schedule_work(&smcibdev->port_event_work); 170 /* fall through */ 171 case IB_EVENT_DEVICE_FATAL: 172 /* tbd in follow-on patch: 173 * abnormal close of corresponding connections 174 */ 175 break; 176 case IB_EVENT_PORT_ACTIVE: 177 port_idx = ibevent->element.port_num - 1; 178 set_bit(port_idx, &smcibdev->port_event_mask); 179 schedule_work(&smcibdev->port_event_work); 180 break; 181 default: 182 break; 183 } 184 } 185 186 void smc_ib_dealloc_protection_domain(struct smc_link *lnk) 187 { 188 ib_dealloc_pd(lnk->roce_pd); 189 lnk->roce_pd = NULL; 190 } 191 192 int smc_ib_create_protection_domain(struct smc_link *lnk) 193 { 194 int rc; 195 196 lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0); 197 rc = PTR_ERR_OR_ZERO(lnk->roce_pd); 198 if (IS_ERR(lnk->roce_pd)) 199 lnk->roce_pd = NULL; 200 return rc; 201 } 202 203 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv) 204 { 205 switch (ibevent->event) { 206 case IB_EVENT_DEVICE_FATAL: 207 case IB_EVENT_GID_CHANGE: 208 case IB_EVENT_PORT_ERR: 209 case IB_EVENT_QP_ACCESS_ERR: 210 /* tbd in follow-on patch: 211 * abnormal close of corresponding connections 212 */ 213 break; 214 default: 215 break; 216 } 217 } 218 219 void smc_ib_destroy_queue_pair(struct smc_link *lnk) 220 { 221 ib_destroy_qp(lnk->roce_qp); 222 lnk->roce_qp = NULL; 223 } 224 225 /* create a queue pair within the protection domain for a link */ 226 int smc_ib_create_queue_pair(struct smc_link *lnk) 227 { 228 struct ib_qp_init_attr qp_attr = { 229 .event_handler = smc_ib_qp_event_handler, 230 .qp_context = lnk, 231 .send_cq = lnk->smcibdev->roce_cq_send, 232 .recv_cq = lnk->smcibdev->roce_cq_recv, 233 .srq = NULL, 234 .cap = { 235 /* include unsolicited rdma_writes as well, 236 * there are max. 2 RDMA_WRITE per 1 WR_SEND 237 */ 238 .max_send_wr = SMC_WR_BUF_CNT * 3, 239 .max_recv_wr = SMC_WR_BUF_CNT * 3, 240 .max_send_sge = SMC_IB_MAX_SEND_SGE, 241 .max_recv_sge = 1, 242 }, 243 .sq_sig_type = IB_SIGNAL_REQ_WR, 244 .qp_type = IB_QPT_RC, 245 }; 246 int rc; 247 248 lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr); 249 rc = PTR_ERR_OR_ZERO(lnk->roce_qp); 250 if (IS_ERR(lnk->roce_qp)) 251 lnk->roce_qp = NULL; 252 else 253 smc_wr_remember_qp_attr(lnk); 254 return rc; 255 } 256 257 void smc_ib_put_memory_region(struct ib_mr *mr) 258 { 259 ib_dereg_mr(mr); 260 } 261 262 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot) 263 { 264 unsigned int offset = 0; 265 int sg_num; 266 267 /* map the largest prefix of a dma mapped SG list */ 268 sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK], 269 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 270 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 271 &offset, PAGE_SIZE); 272 273 return sg_num; 274 } 275 276 /* Allocate a memory region and map the dma mapped SG list of buf_slot */ 277 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags, 278 struct smc_buf_desc *buf_slot) 279 { 280 if (buf_slot->mr_rx[SMC_SINGLE_LINK]) 281 return 0; /* already done */ 282 283 buf_slot->mr_rx[SMC_SINGLE_LINK] = 284 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order); 285 if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) { 286 int rc; 287 288 rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]); 289 buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL; 290 return rc; 291 } 292 293 if (smc_ib_map_mr_sg(buf_slot) != 1) 294 return -EINVAL; 295 296 return 0; 297 } 298 299 /* synchronize buffer usage for cpu access */ 300 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev, 301 struct smc_buf_desc *buf_slot, 302 enum dma_data_direction data_direction) 303 { 304 struct scatterlist *sg; 305 unsigned int i; 306 307 /* for now there is just one DMA address */ 308 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 309 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 310 if (!sg_dma_len(sg)) 311 break; 312 ib_dma_sync_single_for_cpu(smcibdev->ibdev, 313 sg_dma_address(sg), 314 sg_dma_len(sg), 315 data_direction); 316 } 317 } 318 319 /* synchronize buffer usage for device access */ 320 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev, 321 struct smc_buf_desc *buf_slot, 322 enum dma_data_direction data_direction) 323 { 324 struct scatterlist *sg; 325 unsigned int i; 326 327 /* for now there is just one DMA address */ 328 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 329 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 330 if (!sg_dma_len(sg)) 331 break; 332 ib_dma_sync_single_for_device(smcibdev->ibdev, 333 sg_dma_address(sg), 334 sg_dma_len(sg), 335 data_direction); 336 } 337 } 338 339 /* Map a new TX or RX buffer SG-table to DMA */ 340 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev, 341 struct smc_buf_desc *buf_slot, 342 enum dma_data_direction data_direction) 343 { 344 int mapped_nents; 345 346 mapped_nents = ib_dma_map_sg(smcibdev->ibdev, 347 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 348 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 349 data_direction); 350 if (!mapped_nents) 351 return -ENOMEM; 352 353 return mapped_nents; 354 } 355 356 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev, 357 struct smc_buf_desc *buf_slot, 358 enum dma_data_direction data_direction) 359 { 360 if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address) 361 return; /* already unmapped */ 362 363 ib_dma_unmap_sg(smcibdev->ibdev, 364 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 365 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 366 data_direction); 367 buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0; 368 } 369 370 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport) 371 { 372 struct net_device *ndev; 373 int rc; 374 375 rc = ib_query_gid(smcibdev->ibdev, ibport, 0, 376 &smcibdev->gid[ibport - 1], NULL); 377 /* the SMC protocol requires specification of the roce MAC address; 378 * if net_device cannot be determined, it can be derived from gid 0 379 */ 380 ndev = smcibdev->ibdev->get_netdev(smcibdev->ibdev, ibport); 381 if (ndev) { 382 memcpy(&smcibdev->mac, ndev->dev_addr, ETH_ALEN); 383 } else if (!rc) { 384 memcpy(&smcibdev->mac[ibport - 1][0], 385 &smcibdev->gid[ibport - 1].raw[8], 3); 386 memcpy(&smcibdev->mac[ibport - 1][3], 387 &smcibdev->gid[ibport - 1].raw[13], 3); 388 smcibdev->mac[ibport - 1][0] &= ~0x02; 389 } 390 return rc; 391 } 392 393 /* Create an identifier unique for this instance of SMC-R. 394 * The MAC-address of the first active registered IB device 395 * plus a random 2-byte number is used to create this identifier. 396 * This name is delivered to the peer during connection initialization. 397 */ 398 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev, 399 u8 ibport) 400 { 401 memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1], 402 sizeof(smcibdev->mac[ibport - 1])); 403 get_random_bytes(&local_systemid[0], 2); 404 } 405 406 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport) 407 { 408 return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE; 409 } 410 411 int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport) 412 { 413 int rc; 414 415 memset(&smcibdev->pattr[ibport - 1], 0, 416 sizeof(smcibdev->pattr[ibport - 1])); 417 rc = ib_query_port(smcibdev->ibdev, ibport, 418 &smcibdev->pattr[ibport - 1]); 419 if (rc) 420 goto out; 421 rc = smc_ib_fill_gid_and_mac(smcibdev, ibport); 422 if (rc) 423 goto out; 424 if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET, 425 sizeof(local_systemid)) && 426 smc_ib_port_active(smcibdev, ibport)) 427 /* create unique system identifier */ 428 smc_ib_define_local_systemid(smcibdev, ibport); 429 out: 430 return rc; 431 } 432 433 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev) 434 { 435 struct ib_cq_init_attr cqattr = { 436 .cqe = SMC_WR_MAX_CQE, .comp_vector = 0 }; 437 long rc; 438 439 smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev, 440 smc_wr_tx_cq_handler, NULL, 441 smcibdev, &cqattr); 442 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send); 443 if (IS_ERR(smcibdev->roce_cq_send)) { 444 smcibdev->roce_cq_send = NULL; 445 return rc; 446 } 447 smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev, 448 smc_wr_rx_cq_handler, NULL, 449 smcibdev, &cqattr); 450 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv); 451 if (IS_ERR(smcibdev->roce_cq_recv)) { 452 smcibdev->roce_cq_recv = NULL; 453 goto err; 454 } 455 INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev, 456 smc_ib_global_event_handler); 457 ib_register_event_handler(&smcibdev->event_handler); 458 smc_wr_add_dev(smcibdev); 459 smcibdev->initialized = 1; 460 return rc; 461 462 err: 463 ib_destroy_cq(smcibdev->roce_cq_send); 464 return rc; 465 } 466 467 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev) 468 { 469 if (!smcibdev->initialized) 470 return; 471 smc_wr_remove_dev(smcibdev); 472 ib_unregister_event_handler(&smcibdev->event_handler); 473 ib_destroy_cq(smcibdev->roce_cq_recv); 474 ib_destroy_cq(smcibdev->roce_cq_send); 475 } 476 477 static struct ib_client smc_ib_client; 478 479 /* callback function for ib_register_client() */ 480 static void smc_ib_add_dev(struct ib_device *ibdev) 481 { 482 struct smc_ib_device *smcibdev; 483 484 if (ibdev->node_type != RDMA_NODE_IB_CA) 485 return; 486 487 smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL); 488 if (!smcibdev) 489 return; 490 491 smcibdev->ibdev = ibdev; 492 INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work); 493 494 spin_lock(&smc_ib_devices.lock); 495 list_add_tail(&smcibdev->list, &smc_ib_devices.list); 496 spin_unlock(&smc_ib_devices.lock); 497 ib_set_client_data(ibdev, &smc_ib_client, smcibdev); 498 } 499 500 /* callback function for ib_register_client() */ 501 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data) 502 { 503 struct smc_ib_device *smcibdev; 504 505 smcibdev = ib_get_client_data(ibdev, &smc_ib_client); 506 ib_set_client_data(ibdev, &smc_ib_client, NULL); 507 spin_lock(&smc_ib_devices.lock); 508 list_del_init(&smcibdev->list); /* remove from smc_ib_devices */ 509 spin_unlock(&smc_ib_devices.lock); 510 smc_pnet_remove_by_ibdev(smcibdev); 511 smc_ib_cleanup_per_ibdev(smcibdev); 512 kfree(smcibdev); 513 } 514 515 static struct ib_client smc_ib_client = { 516 .name = "smc_ib", 517 .add = smc_ib_add_dev, 518 .remove = smc_ib_remove_dev, 519 }; 520 521 int __init smc_ib_register_client(void) 522 { 523 return ib_register_client(&smc_ib_client); 524 } 525 526 void smc_ib_unregister_client(void) 527 { 528 ib_unregister_client(&smc_ib_client); 529 } 530