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/random.h> 16 #include <linux/workqueue.h> 17 #include <linux/scatterlist.h> 18 #include <rdma/ib_verbs.h> 19 20 #include "smc_pnet.h" 21 #include "smc_ib.h" 22 #include "smc_core.h" 23 #include "smc_wr.h" 24 #include "smc.h" 25 26 #define SMC_MAX_CQE 32766 /* max. # of completion queue elements */ 27 28 #define SMC_QP_MIN_RNR_TIMER 5 29 #define SMC_QP_TIMEOUT 15 /* 4096 * 2 ** timeout usec */ 30 #define SMC_QP_RETRY_CNT 7 /* 7: infinite */ 31 #define SMC_QP_RNR_RETRY 7 /* 7: infinite */ 32 33 struct smc_ib_devices smc_ib_devices = { /* smc-registered ib devices */ 34 .lock = __SPIN_LOCK_UNLOCKED(smc_ib_devices.lock), 35 .list = LIST_HEAD_INIT(smc_ib_devices.list), 36 }; 37 38 #define SMC_LOCAL_SYSTEMID_RESET "%%%%%%%" 39 40 u8 local_systemid[SMC_SYSTEMID_LEN] = SMC_LOCAL_SYSTEMID_RESET; /* unique system 41 * identifier 42 */ 43 44 static int smc_ib_modify_qp_init(struct smc_link *lnk) 45 { 46 struct ib_qp_attr qp_attr; 47 48 memset(&qp_attr, 0, sizeof(qp_attr)); 49 qp_attr.qp_state = IB_QPS_INIT; 50 qp_attr.pkey_index = 0; 51 qp_attr.port_num = lnk->ibport; 52 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE 53 | IB_ACCESS_REMOTE_WRITE; 54 return ib_modify_qp(lnk->roce_qp, &qp_attr, 55 IB_QP_STATE | IB_QP_PKEY_INDEX | 56 IB_QP_ACCESS_FLAGS | IB_QP_PORT); 57 } 58 59 static int smc_ib_modify_qp_rtr(struct smc_link *lnk) 60 { 61 enum ib_qp_attr_mask qp_attr_mask = 62 IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | IB_QP_DEST_QPN | 63 IB_QP_RQ_PSN | IB_QP_MAX_DEST_RD_ATOMIC | IB_QP_MIN_RNR_TIMER; 64 struct ib_qp_attr qp_attr; 65 66 memset(&qp_attr, 0, sizeof(qp_attr)); 67 qp_attr.qp_state = IB_QPS_RTR; 68 qp_attr.path_mtu = min(lnk->path_mtu, lnk->peer_mtu); 69 qp_attr.ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE; 70 rdma_ah_set_port_num(&qp_attr.ah_attr, lnk->ibport); 71 rdma_ah_set_grh(&qp_attr.ah_attr, NULL, 0, 0, 1, 0); 72 rdma_ah_set_dgid_raw(&qp_attr.ah_attr, lnk->peer_gid); 73 memcpy(&qp_attr.ah_attr.roce.dmac, lnk->peer_mac, 74 sizeof(lnk->peer_mac)); 75 qp_attr.dest_qp_num = lnk->peer_qpn; 76 qp_attr.rq_psn = lnk->peer_psn; /* starting receive packet seq # */ 77 qp_attr.max_dest_rd_atomic = 1; /* max # of resources for incoming 78 * requests 79 */ 80 qp_attr.min_rnr_timer = SMC_QP_MIN_RNR_TIMER; 81 82 return ib_modify_qp(lnk->roce_qp, &qp_attr, qp_attr_mask); 83 } 84 85 int smc_ib_modify_qp_rts(struct smc_link *lnk) 86 { 87 struct ib_qp_attr qp_attr; 88 89 memset(&qp_attr, 0, sizeof(qp_attr)); 90 qp_attr.qp_state = IB_QPS_RTS; 91 qp_attr.timeout = SMC_QP_TIMEOUT; /* local ack timeout */ 92 qp_attr.retry_cnt = SMC_QP_RETRY_CNT; /* retry count */ 93 qp_attr.rnr_retry = SMC_QP_RNR_RETRY; /* RNR retries, 7=infinite */ 94 qp_attr.sq_psn = lnk->psn_initial; /* starting send packet seq # */ 95 qp_attr.max_rd_atomic = 1; /* # of outstanding RDMA reads and 96 * atomic ops allowed 97 */ 98 return ib_modify_qp(lnk->roce_qp, &qp_attr, 99 IB_QP_STATE | IB_QP_TIMEOUT | IB_QP_RETRY_CNT | 100 IB_QP_SQ_PSN | IB_QP_RNR_RETRY | 101 IB_QP_MAX_QP_RD_ATOMIC); 102 } 103 104 int smc_ib_modify_qp_reset(struct smc_link *lnk) 105 { 106 struct ib_qp_attr qp_attr; 107 108 memset(&qp_attr, 0, sizeof(qp_attr)); 109 qp_attr.qp_state = IB_QPS_RESET; 110 return ib_modify_qp(lnk->roce_qp, &qp_attr, IB_QP_STATE); 111 } 112 113 int smc_ib_ready_link(struct smc_link *lnk) 114 { 115 struct smc_link_group *lgr = 116 container_of(lnk, struct smc_link_group, lnk[0]); 117 int rc = 0; 118 119 rc = smc_ib_modify_qp_init(lnk); 120 if (rc) 121 goto out; 122 123 rc = smc_ib_modify_qp_rtr(lnk); 124 if (rc) 125 goto out; 126 smc_wr_remember_qp_attr(lnk); 127 rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv, 128 IB_CQ_SOLICITED_MASK); 129 if (rc) 130 goto out; 131 rc = smc_wr_rx_post_init(lnk); 132 if (rc) 133 goto out; 134 smc_wr_remember_qp_attr(lnk); 135 136 if (lgr->role == SMC_SERV) { 137 rc = smc_ib_modify_qp_rts(lnk); 138 if (rc) 139 goto out; 140 smc_wr_remember_qp_attr(lnk); 141 } 142 out: 143 return rc; 144 } 145 146 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport) 147 { 148 struct ib_gid_attr gattr; 149 int rc; 150 151 rc = ib_query_gid(smcibdev->ibdev, ibport, 0, 152 &smcibdev->gid[ibport - 1], &gattr); 153 if (rc || !gattr.ndev) 154 return -ENODEV; 155 156 memcpy(smcibdev->mac[ibport - 1], gattr.ndev->dev_addr, ETH_ALEN); 157 dev_put(gattr.ndev); 158 return 0; 159 } 160 161 /* Create an identifier unique for this instance of SMC-R. 162 * The MAC-address of the first active registered IB device 163 * plus a random 2-byte number is used to create this identifier. 164 * This name is delivered to the peer during connection initialization. 165 */ 166 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev, 167 u8 ibport) 168 { 169 memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1], 170 sizeof(smcibdev->mac[ibport - 1])); 171 get_random_bytes(&local_systemid[0], 2); 172 } 173 174 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport) 175 { 176 return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE; 177 } 178 179 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport) 180 { 181 int rc; 182 183 memset(&smcibdev->pattr[ibport - 1], 0, 184 sizeof(smcibdev->pattr[ibport - 1])); 185 rc = ib_query_port(smcibdev->ibdev, ibport, 186 &smcibdev->pattr[ibport - 1]); 187 if (rc) 188 goto out; 189 /* the SMC protocol requires specification of the RoCE MAC address */ 190 rc = smc_ib_fill_gid_and_mac(smcibdev, ibport); 191 if (rc) 192 goto out; 193 if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET, 194 sizeof(local_systemid)) && 195 smc_ib_port_active(smcibdev, ibport)) 196 /* create unique system identifier */ 197 smc_ib_define_local_systemid(smcibdev, ibport); 198 out: 199 return rc; 200 } 201 202 /* process context wrapper for might_sleep smc_ib_remember_port_attr */ 203 static void smc_ib_port_event_work(struct work_struct *work) 204 { 205 struct smc_ib_device *smcibdev = container_of( 206 work, struct smc_ib_device, port_event_work); 207 u8 port_idx; 208 209 for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) { 210 smc_ib_remember_port_attr(smcibdev, port_idx + 1); 211 clear_bit(port_idx, &smcibdev->port_event_mask); 212 if (!smc_ib_port_active(smcibdev, port_idx + 1)) 213 smc_port_terminate(smcibdev, port_idx + 1); 214 } 215 } 216 217 /* can be called in IRQ context */ 218 static void smc_ib_global_event_handler(struct ib_event_handler *handler, 219 struct ib_event *ibevent) 220 { 221 struct smc_ib_device *smcibdev; 222 u8 port_idx; 223 224 smcibdev = container_of(handler, struct smc_ib_device, event_handler); 225 226 switch (ibevent->event) { 227 case IB_EVENT_PORT_ERR: 228 case IB_EVENT_DEVICE_FATAL: 229 case IB_EVENT_PORT_ACTIVE: 230 port_idx = ibevent->element.port_num - 1; 231 set_bit(port_idx, &smcibdev->port_event_mask); 232 schedule_work(&smcibdev->port_event_work); 233 break; 234 default: 235 break; 236 } 237 } 238 239 void smc_ib_dealloc_protection_domain(struct smc_link *lnk) 240 { 241 if (lnk->roce_pd) 242 ib_dealloc_pd(lnk->roce_pd); 243 lnk->roce_pd = NULL; 244 } 245 246 int smc_ib_create_protection_domain(struct smc_link *lnk) 247 { 248 int rc; 249 250 lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0); 251 rc = PTR_ERR_OR_ZERO(lnk->roce_pd); 252 if (IS_ERR(lnk->roce_pd)) 253 lnk->roce_pd = NULL; 254 return rc; 255 } 256 257 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv) 258 { 259 struct smc_ib_device *smcibdev = 260 (struct smc_ib_device *)ibevent->device; 261 u8 port_idx; 262 263 switch (ibevent->event) { 264 case IB_EVENT_DEVICE_FATAL: 265 case IB_EVENT_GID_CHANGE: 266 case IB_EVENT_PORT_ERR: 267 case IB_EVENT_QP_ACCESS_ERR: 268 port_idx = ibevent->element.port_num - 1; 269 set_bit(port_idx, &smcibdev->port_event_mask); 270 schedule_work(&smcibdev->port_event_work); 271 break; 272 default: 273 break; 274 } 275 } 276 277 void smc_ib_destroy_queue_pair(struct smc_link *lnk) 278 { 279 if (lnk->roce_qp) 280 ib_destroy_qp(lnk->roce_qp); 281 lnk->roce_qp = NULL; 282 } 283 284 /* create a queue pair within the protection domain for a link */ 285 int smc_ib_create_queue_pair(struct smc_link *lnk) 286 { 287 struct ib_qp_init_attr qp_attr = { 288 .event_handler = smc_ib_qp_event_handler, 289 .qp_context = lnk, 290 .send_cq = lnk->smcibdev->roce_cq_send, 291 .recv_cq = lnk->smcibdev->roce_cq_recv, 292 .srq = NULL, 293 .cap = { 294 /* include unsolicited rdma_writes as well, 295 * there are max. 2 RDMA_WRITE per 1 WR_SEND 296 */ 297 .max_send_wr = SMC_WR_BUF_CNT * 3, 298 .max_recv_wr = SMC_WR_BUF_CNT * 3, 299 .max_send_sge = SMC_IB_MAX_SEND_SGE, 300 .max_recv_sge = 1, 301 }, 302 .sq_sig_type = IB_SIGNAL_REQ_WR, 303 .qp_type = IB_QPT_RC, 304 }; 305 int rc; 306 307 lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr); 308 rc = PTR_ERR_OR_ZERO(lnk->roce_qp); 309 if (IS_ERR(lnk->roce_qp)) 310 lnk->roce_qp = NULL; 311 else 312 smc_wr_remember_qp_attr(lnk); 313 return rc; 314 } 315 316 void smc_ib_put_memory_region(struct ib_mr *mr) 317 { 318 ib_dereg_mr(mr); 319 } 320 321 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot) 322 { 323 unsigned int offset = 0; 324 int sg_num; 325 326 /* map the largest prefix of a dma mapped SG list */ 327 sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK], 328 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 329 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 330 &offset, PAGE_SIZE); 331 332 return sg_num; 333 } 334 335 /* Allocate a memory region and map the dma mapped SG list of buf_slot */ 336 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags, 337 struct smc_buf_desc *buf_slot) 338 { 339 if (buf_slot->mr_rx[SMC_SINGLE_LINK]) 340 return 0; /* already done */ 341 342 buf_slot->mr_rx[SMC_SINGLE_LINK] = 343 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order); 344 if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) { 345 int rc; 346 347 rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]); 348 buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL; 349 return rc; 350 } 351 352 if (smc_ib_map_mr_sg(buf_slot) != 1) 353 return -EINVAL; 354 355 return 0; 356 } 357 358 /* synchronize buffer usage for cpu access */ 359 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev, 360 struct smc_buf_desc *buf_slot, 361 enum dma_data_direction data_direction) 362 { 363 struct scatterlist *sg; 364 unsigned int i; 365 366 /* for now there is just one DMA address */ 367 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 368 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 369 if (!sg_dma_len(sg)) 370 break; 371 ib_dma_sync_single_for_cpu(smcibdev->ibdev, 372 sg_dma_address(sg), 373 sg_dma_len(sg), 374 data_direction); 375 } 376 } 377 378 /* synchronize buffer usage for device access */ 379 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev, 380 struct smc_buf_desc *buf_slot, 381 enum dma_data_direction data_direction) 382 { 383 struct scatterlist *sg; 384 unsigned int i; 385 386 /* for now there is just one DMA address */ 387 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 388 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 389 if (!sg_dma_len(sg)) 390 break; 391 ib_dma_sync_single_for_device(smcibdev->ibdev, 392 sg_dma_address(sg), 393 sg_dma_len(sg), 394 data_direction); 395 } 396 } 397 398 /* Map a new TX or RX buffer SG-table to DMA */ 399 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev, 400 struct smc_buf_desc *buf_slot, 401 enum dma_data_direction data_direction) 402 { 403 int mapped_nents; 404 405 mapped_nents = ib_dma_map_sg(smcibdev->ibdev, 406 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 407 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 408 data_direction); 409 if (!mapped_nents) 410 return -ENOMEM; 411 412 return mapped_nents; 413 } 414 415 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev, 416 struct smc_buf_desc *buf_slot, 417 enum dma_data_direction data_direction) 418 { 419 if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address) 420 return; /* already unmapped */ 421 422 ib_dma_unmap_sg(smcibdev->ibdev, 423 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 424 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 425 data_direction); 426 buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0; 427 } 428 429 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev) 430 { 431 struct ib_cq_init_attr cqattr = { 432 .cqe = SMC_MAX_CQE, .comp_vector = 0 }; 433 int cqe_size_order, smc_order; 434 long rc; 435 436 /* the calculated number of cq entries fits to mlx5 cq allocation */ 437 cqe_size_order = cache_line_size() == 128 ? 7 : 6; 438 smc_order = MAX_ORDER - cqe_size_order - 1; 439 if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE) 440 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2; 441 smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev, 442 smc_wr_tx_cq_handler, NULL, 443 smcibdev, &cqattr); 444 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send); 445 if (IS_ERR(smcibdev->roce_cq_send)) { 446 smcibdev->roce_cq_send = NULL; 447 return rc; 448 } 449 smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev, 450 smc_wr_rx_cq_handler, NULL, 451 smcibdev, &cqattr); 452 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv); 453 if (IS_ERR(smcibdev->roce_cq_recv)) { 454 smcibdev->roce_cq_recv = NULL; 455 goto err; 456 } 457 smc_wr_add_dev(smcibdev); 458 smcibdev->initialized = 1; 459 return rc; 460 461 err: 462 ib_destroy_cq(smcibdev->roce_cq_send); 463 return rc; 464 } 465 466 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev) 467 { 468 if (!smcibdev->initialized) 469 return; 470 smcibdev->initialized = 0; 471 smc_wr_remove_dev(smcibdev); 472 ib_destroy_cq(smcibdev->roce_cq_recv); 473 ib_destroy_cq(smcibdev->roce_cq_send); 474 } 475 476 static struct ib_client smc_ib_client; 477 478 /* callback function for ib_register_client() */ 479 static void smc_ib_add_dev(struct ib_device *ibdev) 480 { 481 struct smc_ib_device *smcibdev; 482 u8 port_cnt; 483 int i; 484 485 if (ibdev->node_type != RDMA_NODE_IB_CA) 486 return; 487 488 smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL); 489 if (!smcibdev) 490 return; 491 492 smcibdev->ibdev = ibdev; 493 INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work); 494 495 spin_lock(&smc_ib_devices.lock); 496 list_add_tail(&smcibdev->list, &smc_ib_devices.list); 497 spin_unlock(&smc_ib_devices.lock); 498 ib_set_client_data(ibdev, &smc_ib_client, smcibdev); 499 INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev, 500 smc_ib_global_event_handler); 501 ib_register_event_handler(&smcibdev->event_handler); 502 503 /* trigger reading of the port attributes */ 504 port_cnt = smcibdev->ibdev->phys_port_cnt; 505 for (i = 0; 506 i < min_t(size_t, port_cnt, SMC_MAX_PORTS); 507 i++) { 508 set_bit(i, &smcibdev->port_event_mask); 509 /* determine pnetids of the port */ 510 smc_pnetid_by_dev_port(ibdev->dev.parent, i, 511 smcibdev->pnetid[i]); 512 } 513 schedule_work(&smcibdev->port_event_work); 514 } 515 516 /* callback function for ib_register_client() */ 517 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data) 518 { 519 struct smc_ib_device *smcibdev; 520 521 smcibdev = ib_get_client_data(ibdev, &smc_ib_client); 522 ib_set_client_data(ibdev, &smc_ib_client, NULL); 523 spin_lock(&smc_ib_devices.lock); 524 list_del_init(&smcibdev->list); /* remove from smc_ib_devices */ 525 spin_unlock(&smc_ib_devices.lock); 526 smc_pnet_remove_by_ibdev(smcibdev); 527 smc_ib_cleanup_per_ibdev(smcibdev); 528 ib_unregister_event_handler(&smcibdev->event_handler); 529 kfree(smcibdev); 530 } 531 532 static struct ib_client smc_ib_client = { 533 .name = "smc_ib", 534 .add = smc_ib_add_dev, 535 .remove = smc_ib_remove_dev, 536 }; 537 538 int __init smc_ib_register_client(void) 539 { 540 return ib_register_client(&smc_ib_client); 541 } 542 543 void smc_ib_unregister_client(void) 544 { 545 ib_unregister_client(&smc_ib_client); 546 } 547