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 = smc_get_lgr(lnk); 116 int rc = 0; 117 118 rc = smc_ib_modify_qp_init(lnk); 119 if (rc) 120 goto out; 121 122 rc = smc_ib_modify_qp_rtr(lnk); 123 if (rc) 124 goto out; 125 smc_wr_remember_qp_attr(lnk); 126 rc = ib_req_notify_cq(lnk->smcibdev->roce_cq_recv, 127 IB_CQ_SOLICITED_MASK); 128 if (rc) 129 goto out; 130 rc = smc_wr_rx_post_init(lnk); 131 if (rc) 132 goto out; 133 smc_wr_remember_qp_attr(lnk); 134 135 if (lgr->role == SMC_SERV) { 136 rc = smc_ib_modify_qp_rts(lnk); 137 if (rc) 138 goto out; 139 smc_wr_remember_qp_attr(lnk); 140 } 141 out: 142 return rc; 143 } 144 145 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport) 146 { 147 struct ib_gid_attr gattr; 148 int rc; 149 150 rc = ib_query_gid(smcibdev->ibdev, ibport, 0, 151 &smcibdev->gid[ibport - 1], &gattr); 152 if (rc || !gattr.ndev) 153 return -ENODEV; 154 155 memcpy(smcibdev->mac[ibport - 1], gattr.ndev->dev_addr, ETH_ALEN); 156 dev_put(gattr.ndev); 157 return 0; 158 } 159 160 /* Create an identifier unique for this instance of SMC-R. 161 * The MAC-address of the first active registered IB device 162 * plus a random 2-byte number is used to create this identifier. 163 * This name is delivered to the peer during connection initialization. 164 */ 165 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev, 166 u8 ibport) 167 { 168 memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1], 169 sizeof(smcibdev->mac[ibport - 1])); 170 get_random_bytes(&local_systemid[0], 2); 171 } 172 173 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport) 174 { 175 return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE; 176 } 177 178 static int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport) 179 { 180 int rc; 181 182 memset(&smcibdev->pattr[ibport - 1], 0, 183 sizeof(smcibdev->pattr[ibport - 1])); 184 rc = ib_query_port(smcibdev->ibdev, ibport, 185 &smcibdev->pattr[ibport - 1]); 186 if (rc) 187 goto out; 188 /* the SMC protocol requires specification of the RoCE MAC address */ 189 rc = smc_ib_fill_gid_and_mac(smcibdev, ibport); 190 if (rc) 191 goto out; 192 if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET, 193 sizeof(local_systemid)) && 194 smc_ib_port_active(smcibdev, ibport)) 195 /* create unique system identifier */ 196 smc_ib_define_local_systemid(smcibdev, ibport); 197 out: 198 return rc; 199 } 200 201 /* process context wrapper for might_sleep smc_ib_remember_port_attr */ 202 static void smc_ib_port_event_work(struct work_struct *work) 203 { 204 struct smc_ib_device *smcibdev = container_of( 205 work, struct smc_ib_device, port_event_work); 206 u8 port_idx; 207 208 for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) { 209 smc_ib_remember_port_attr(smcibdev, port_idx + 1); 210 clear_bit(port_idx, &smcibdev->port_event_mask); 211 if (!smc_ib_port_active(smcibdev, port_idx + 1)) 212 smc_port_terminate(smcibdev, port_idx + 1); 213 } 214 } 215 216 /* can be called in IRQ context */ 217 static void smc_ib_global_event_handler(struct ib_event_handler *handler, 218 struct ib_event *ibevent) 219 { 220 struct smc_ib_device *smcibdev; 221 u8 port_idx; 222 223 smcibdev = container_of(handler, struct smc_ib_device, event_handler); 224 225 switch (ibevent->event) { 226 case IB_EVENT_PORT_ERR: 227 case IB_EVENT_DEVICE_FATAL: 228 case IB_EVENT_PORT_ACTIVE: 229 port_idx = ibevent->element.port_num - 1; 230 set_bit(port_idx, &smcibdev->port_event_mask); 231 schedule_work(&smcibdev->port_event_work); 232 break; 233 default: 234 break; 235 } 236 } 237 238 void smc_ib_dealloc_protection_domain(struct smc_link *lnk) 239 { 240 if (lnk->roce_pd) 241 ib_dealloc_pd(lnk->roce_pd); 242 lnk->roce_pd = NULL; 243 } 244 245 int smc_ib_create_protection_domain(struct smc_link *lnk) 246 { 247 int rc; 248 249 lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0); 250 rc = PTR_ERR_OR_ZERO(lnk->roce_pd); 251 if (IS_ERR(lnk->roce_pd)) 252 lnk->roce_pd = NULL; 253 return rc; 254 } 255 256 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv) 257 { 258 struct smc_ib_device *smcibdev = 259 (struct smc_ib_device *)ibevent->device; 260 u8 port_idx; 261 262 switch (ibevent->event) { 263 case IB_EVENT_DEVICE_FATAL: 264 case IB_EVENT_GID_CHANGE: 265 case IB_EVENT_PORT_ERR: 266 case IB_EVENT_QP_ACCESS_ERR: 267 port_idx = ibevent->element.port_num - 1; 268 set_bit(port_idx, &smcibdev->port_event_mask); 269 schedule_work(&smcibdev->port_event_work); 270 break; 271 default: 272 break; 273 } 274 } 275 276 void smc_ib_destroy_queue_pair(struct smc_link *lnk) 277 { 278 if (lnk->roce_qp) 279 ib_destroy_qp(lnk->roce_qp); 280 lnk->roce_qp = NULL; 281 } 282 283 /* create a queue pair within the protection domain for a link */ 284 int smc_ib_create_queue_pair(struct smc_link *lnk) 285 { 286 struct ib_qp_init_attr qp_attr = { 287 .event_handler = smc_ib_qp_event_handler, 288 .qp_context = lnk, 289 .send_cq = lnk->smcibdev->roce_cq_send, 290 .recv_cq = lnk->smcibdev->roce_cq_recv, 291 .srq = NULL, 292 .cap = { 293 /* include unsolicited rdma_writes as well, 294 * there are max. 2 RDMA_WRITE per 1 WR_SEND 295 */ 296 .max_send_wr = SMC_WR_BUF_CNT * 3, 297 .max_recv_wr = SMC_WR_BUF_CNT * 3, 298 .max_send_sge = SMC_IB_MAX_SEND_SGE, 299 .max_recv_sge = 1, 300 }, 301 .sq_sig_type = IB_SIGNAL_REQ_WR, 302 .qp_type = IB_QPT_RC, 303 }; 304 int rc; 305 306 lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr); 307 rc = PTR_ERR_OR_ZERO(lnk->roce_qp); 308 if (IS_ERR(lnk->roce_qp)) 309 lnk->roce_qp = NULL; 310 else 311 smc_wr_remember_qp_attr(lnk); 312 return rc; 313 } 314 315 void smc_ib_put_memory_region(struct ib_mr *mr) 316 { 317 ib_dereg_mr(mr); 318 } 319 320 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot) 321 { 322 unsigned int offset = 0; 323 int sg_num; 324 325 /* map the largest prefix of a dma mapped SG list */ 326 sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK], 327 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 328 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 329 &offset, PAGE_SIZE); 330 331 return sg_num; 332 } 333 334 /* Allocate a memory region and map the dma mapped SG list of buf_slot */ 335 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags, 336 struct smc_buf_desc *buf_slot) 337 { 338 if (buf_slot->mr_rx[SMC_SINGLE_LINK]) 339 return 0; /* already done */ 340 341 buf_slot->mr_rx[SMC_SINGLE_LINK] = 342 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order); 343 if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) { 344 int rc; 345 346 rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]); 347 buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL; 348 return rc; 349 } 350 351 if (smc_ib_map_mr_sg(buf_slot) != 1) 352 return -EINVAL; 353 354 return 0; 355 } 356 357 /* synchronize buffer usage for cpu access */ 358 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev, 359 struct smc_buf_desc *buf_slot, 360 enum dma_data_direction data_direction) 361 { 362 struct scatterlist *sg; 363 unsigned int i; 364 365 /* for now there is just one DMA address */ 366 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 367 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 368 if (!sg_dma_len(sg)) 369 break; 370 ib_dma_sync_single_for_cpu(smcibdev->ibdev, 371 sg_dma_address(sg), 372 sg_dma_len(sg), 373 data_direction); 374 } 375 } 376 377 /* synchronize buffer usage for device access */ 378 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev, 379 struct smc_buf_desc *buf_slot, 380 enum dma_data_direction data_direction) 381 { 382 struct scatterlist *sg; 383 unsigned int i; 384 385 /* for now there is just one DMA address */ 386 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 387 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 388 if (!sg_dma_len(sg)) 389 break; 390 ib_dma_sync_single_for_device(smcibdev->ibdev, 391 sg_dma_address(sg), 392 sg_dma_len(sg), 393 data_direction); 394 } 395 } 396 397 /* Map a new TX or RX buffer SG-table to DMA */ 398 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev, 399 struct smc_buf_desc *buf_slot, 400 enum dma_data_direction data_direction) 401 { 402 int mapped_nents; 403 404 mapped_nents = ib_dma_map_sg(smcibdev->ibdev, 405 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 406 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 407 data_direction); 408 if (!mapped_nents) 409 return -ENOMEM; 410 411 return mapped_nents; 412 } 413 414 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev, 415 struct smc_buf_desc *buf_slot, 416 enum dma_data_direction data_direction) 417 { 418 if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address) 419 return; /* already unmapped */ 420 421 ib_dma_unmap_sg(smcibdev->ibdev, 422 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 423 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 424 data_direction); 425 buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0; 426 } 427 428 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev) 429 { 430 struct ib_cq_init_attr cqattr = { 431 .cqe = SMC_MAX_CQE, .comp_vector = 0 }; 432 int cqe_size_order, smc_order; 433 long rc; 434 435 /* the calculated number of cq entries fits to mlx5 cq allocation */ 436 cqe_size_order = cache_line_size() == 128 ? 7 : 6; 437 smc_order = MAX_ORDER - cqe_size_order - 1; 438 if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE) 439 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2; 440 smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev, 441 smc_wr_tx_cq_handler, NULL, 442 smcibdev, &cqattr); 443 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send); 444 if (IS_ERR(smcibdev->roce_cq_send)) { 445 smcibdev->roce_cq_send = NULL; 446 return rc; 447 } 448 smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev, 449 smc_wr_rx_cq_handler, NULL, 450 smcibdev, &cqattr); 451 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv); 452 if (IS_ERR(smcibdev->roce_cq_recv)) { 453 smcibdev->roce_cq_recv = NULL; 454 goto err; 455 } 456 smc_wr_add_dev(smcibdev); 457 smcibdev->initialized = 1; 458 return rc; 459 460 err: 461 ib_destroy_cq(smcibdev->roce_cq_send); 462 return rc; 463 } 464 465 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev) 466 { 467 if (!smcibdev->initialized) 468 return; 469 smcibdev->initialized = 0; 470 smc_wr_remove_dev(smcibdev); 471 ib_destroy_cq(smcibdev->roce_cq_recv); 472 ib_destroy_cq(smcibdev->roce_cq_send); 473 } 474 475 static struct ib_client smc_ib_client; 476 477 /* callback function for ib_register_client() */ 478 static void smc_ib_add_dev(struct ib_device *ibdev) 479 { 480 struct smc_ib_device *smcibdev; 481 u8 port_cnt; 482 int i; 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 INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev, 499 smc_ib_global_event_handler); 500 ib_register_event_handler(&smcibdev->event_handler); 501 502 /* trigger reading of the port attributes */ 503 port_cnt = smcibdev->ibdev->phys_port_cnt; 504 for (i = 0; 505 i < min_t(size_t, port_cnt, SMC_MAX_PORTS); 506 i++) { 507 set_bit(i, &smcibdev->port_event_mask); 508 /* determine pnetids of the port */ 509 smc_pnetid_by_dev_port(ibdev->dev.parent, i, 510 smcibdev->pnetid[i]); 511 } 512 schedule_work(&smcibdev->port_event_work); 513 } 514 515 /* callback function for ib_register_client() */ 516 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data) 517 { 518 struct smc_ib_device *smcibdev; 519 520 smcibdev = ib_get_client_data(ibdev, &smc_ib_client); 521 ib_set_client_data(ibdev, &smc_ib_client, NULL); 522 spin_lock(&smc_ib_devices.lock); 523 list_del_init(&smcibdev->list); /* remove from smc_ib_devices */ 524 spin_unlock(&smc_ib_devices.lock); 525 smc_pnet_remove_by_ibdev(smcibdev); 526 smc_ib_cleanup_per_ibdev(smcibdev); 527 ib_unregister_event_handler(&smcibdev->event_handler); 528 kfree(smcibdev); 529 } 530 531 static struct ib_client smc_ib_client = { 532 .name = "smc_ib", 533 .add = smc_ib_add_dev, 534 .remove = smc_ib_remove_dev, 535 }; 536 537 int __init smc_ib_register_client(void) 538 { 539 return ib_register_client(&smc_ib_client); 540 } 541 542 void smc_ib_unregister_client(void) 543 { 544 ib_unregister_client(&smc_ib_client); 545 } 546