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 void smc_ib_port_terminate(struct smc_ib_device *smcibdev, u8 ibport) 147 { 148 struct smc_link_group *lgr, *l; 149 150 list_for_each_entry_safe(lgr, l, &smc_lgr_list.list, list) { 151 if (lgr->lnk[SMC_SINGLE_LINK].smcibdev == smcibdev && 152 lgr->lnk[SMC_SINGLE_LINK].ibport == ibport) 153 smc_lgr_terminate(lgr); 154 } 155 } 156 157 /* process context wrapper for might_sleep smc_ib_remember_port_attr */ 158 static void smc_ib_port_event_work(struct work_struct *work) 159 { 160 struct smc_ib_device *smcibdev = container_of( 161 work, struct smc_ib_device, port_event_work); 162 u8 port_idx; 163 164 for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) { 165 smc_ib_remember_port_attr(smcibdev, port_idx + 1); 166 clear_bit(port_idx, &smcibdev->port_event_mask); 167 if (!smc_ib_port_active(smcibdev, port_idx + 1)) 168 smc_ib_port_terminate(smcibdev, port_idx + 1); 169 } 170 } 171 172 /* can be called in IRQ context */ 173 static void smc_ib_global_event_handler(struct ib_event_handler *handler, 174 struct ib_event *ibevent) 175 { 176 struct smc_ib_device *smcibdev; 177 u8 port_idx; 178 179 smcibdev = container_of(handler, struct smc_ib_device, event_handler); 180 181 switch (ibevent->event) { 182 case IB_EVENT_PORT_ERR: 183 case IB_EVENT_DEVICE_FATAL: 184 case IB_EVENT_PORT_ACTIVE: 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 break; 189 default: 190 break; 191 } 192 } 193 194 void smc_ib_dealloc_protection_domain(struct smc_link *lnk) 195 { 196 if (lnk->roce_pd) 197 ib_dealloc_pd(lnk->roce_pd); 198 lnk->roce_pd = NULL; 199 } 200 201 int smc_ib_create_protection_domain(struct smc_link *lnk) 202 { 203 int rc; 204 205 lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0); 206 rc = PTR_ERR_OR_ZERO(lnk->roce_pd); 207 if (IS_ERR(lnk->roce_pd)) 208 lnk->roce_pd = NULL; 209 return rc; 210 } 211 212 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv) 213 { 214 struct smc_ib_device *smcibdev = 215 (struct smc_ib_device *)ibevent->device; 216 u8 port_idx; 217 218 switch (ibevent->event) { 219 case IB_EVENT_DEVICE_FATAL: 220 case IB_EVENT_GID_CHANGE: 221 case IB_EVENT_PORT_ERR: 222 case IB_EVENT_QP_ACCESS_ERR: 223 port_idx = ibevent->element.port_num - 1; 224 set_bit(port_idx, &smcibdev->port_event_mask); 225 schedule_work(&smcibdev->port_event_work); 226 break; 227 default: 228 break; 229 } 230 } 231 232 void smc_ib_destroy_queue_pair(struct smc_link *lnk) 233 { 234 if (lnk->roce_qp) 235 ib_destroy_qp(lnk->roce_qp); 236 lnk->roce_qp = NULL; 237 } 238 239 /* create a queue pair within the protection domain for a link */ 240 int smc_ib_create_queue_pair(struct smc_link *lnk) 241 { 242 struct ib_qp_init_attr qp_attr = { 243 .event_handler = smc_ib_qp_event_handler, 244 .qp_context = lnk, 245 .send_cq = lnk->smcibdev->roce_cq_send, 246 .recv_cq = lnk->smcibdev->roce_cq_recv, 247 .srq = NULL, 248 .cap = { 249 /* include unsolicited rdma_writes as well, 250 * there are max. 2 RDMA_WRITE per 1 WR_SEND 251 */ 252 .max_send_wr = SMC_WR_BUF_CNT * 3, 253 .max_recv_wr = SMC_WR_BUF_CNT * 3, 254 .max_send_sge = SMC_IB_MAX_SEND_SGE, 255 .max_recv_sge = 1, 256 }, 257 .sq_sig_type = IB_SIGNAL_REQ_WR, 258 .qp_type = IB_QPT_RC, 259 }; 260 int rc; 261 262 lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr); 263 rc = PTR_ERR_OR_ZERO(lnk->roce_qp); 264 if (IS_ERR(lnk->roce_qp)) 265 lnk->roce_qp = NULL; 266 else 267 smc_wr_remember_qp_attr(lnk); 268 return rc; 269 } 270 271 void smc_ib_put_memory_region(struct ib_mr *mr) 272 { 273 ib_dereg_mr(mr); 274 } 275 276 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot) 277 { 278 unsigned int offset = 0; 279 int sg_num; 280 281 /* map the largest prefix of a dma mapped SG list */ 282 sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK], 283 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 284 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 285 &offset, PAGE_SIZE); 286 287 return sg_num; 288 } 289 290 /* Allocate a memory region and map the dma mapped SG list of buf_slot */ 291 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags, 292 struct smc_buf_desc *buf_slot) 293 { 294 if (buf_slot->mr_rx[SMC_SINGLE_LINK]) 295 return 0; /* already done */ 296 297 buf_slot->mr_rx[SMC_SINGLE_LINK] = 298 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order); 299 if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) { 300 int rc; 301 302 rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]); 303 buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL; 304 return rc; 305 } 306 307 if (smc_ib_map_mr_sg(buf_slot) != 1) 308 return -EINVAL; 309 310 return 0; 311 } 312 313 /* synchronize buffer usage for cpu access */ 314 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev, 315 struct smc_buf_desc *buf_slot, 316 enum dma_data_direction data_direction) 317 { 318 struct scatterlist *sg; 319 unsigned int i; 320 321 /* for now there is just one DMA address */ 322 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 323 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 324 if (!sg_dma_len(sg)) 325 break; 326 ib_dma_sync_single_for_cpu(smcibdev->ibdev, 327 sg_dma_address(sg), 328 sg_dma_len(sg), 329 data_direction); 330 } 331 } 332 333 /* synchronize buffer usage for device access */ 334 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev, 335 struct smc_buf_desc *buf_slot, 336 enum dma_data_direction data_direction) 337 { 338 struct scatterlist *sg; 339 unsigned int i; 340 341 /* for now there is just one DMA address */ 342 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 343 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 344 if (!sg_dma_len(sg)) 345 break; 346 ib_dma_sync_single_for_device(smcibdev->ibdev, 347 sg_dma_address(sg), 348 sg_dma_len(sg), 349 data_direction); 350 } 351 } 352 353 /* Map a new TX or RX buffer SG-table to DMA */ 354 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev, 355 struct smc_buf_desc *buf_slot, 356 enum dma_data_direction data_direction) 357 { 358 int mapped_nents; 359 360 mapped_nents = ib_dma_map_sg(smcibdev->ibdev, 361 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 362 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 363 data_direction); 364 if (!mapped_nents) 365 return -ENOMEM; 366 367 return mapped_nents; 368 } 369 370 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev, 371 struct smc_buf_desc *buf_slot, 372 enum dma_data_direction data_direction) 373 { 374 if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address) 375 return; /* already unmapped */ 376 377 ib_dma_unmap_sg(smcibdev->ibdev, 378 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 379 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 380 data_direction); 381 buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0; 382 } 383 384 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport) 385 { 386 struct ib_gid_attr gattr; 387 int rc; 388 389 rc = ib_query_gid(smcibdev->ibdev, ibport, 0, 390 &smcibdev->gid[ibport - 1], &gattr); 391 if (rc || !gattr.ndev) 392 return -ENODEV; 393 394 memcpy(smcibdev->mac[ibport - 1], gattr.ndev->dev_addr, ETH_ALEN); 395 dev_put(gattr.ndev); 396 return 0; 397 } 398 399 /* Create an identifier unique for this instance of SMC-R. 400 * The MAC-address of the first active registered IB device 401 * plus a random 2-byte number is used to create this identifier. 402 * This name is delivered to the peer during connection initialization. 403 */ 404 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev, 405 u8 ibport) 406 { 407 memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1], 408 sizeof(smcibdev->mac[ibport - 1])); 409 get_random_bytes(&local_systemid[0], 2); 410 } 411 412 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport) 413 { 414 return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE; 415 } 416 417 int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport) 418 { 419 int rc; 420 421 memset(&smcibdev->pattr[ibport - 1], 0, 422 sizeof(smcibdev->pattr[ibport - 1])); 423 rc = ib_query_port(smcibdev->ibdev, ibport, 424 &smcibdev->pattr[ibport - 1]); 425 if (rc) 426 goto out; 427 /* the SMC protocol requires specification of the RoCE MAC address */ 428 rc = smc_ib_fill_gid_and_mac(smcibdev, ibport); 429 if (rc) 430 goto out; 431 if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET, 432 sizeof(local_systemid)) && 433 smc_ib_port_active(smcibdev, ibport)) 434 /* create unique system identifier */ 435 smc_ib_define_local_systemid(smcibdev, ibport); 436 out: 437 return rc; 438 } 439 440 long smc_ib_setup_per_ibdev(struct smc_ib_device *smcibdev) 441 { 442 struct ib_cq_init_attr cqattr = { 443 .cqe = SMC_MAX_CQE, .comp_vector = 0 }; 444 int cqe_size_order, smc_order; 445 long rc; 446 447 /* the calculated number of cq entries fits to mlx5 cq allocation */ 448 cqe_size_order = cache_line_size() == 128 ? 7 : 6; 449 smc_order = MAX_ORDER - cqe_size_order - 1; 450 if (SMC_MAX_CQE + 2 > (0x00000001 << smc_order) * PAGE_SIZE) 451 cqattr.cqe = (0x00000001 << smc_order) * PAGE_SIZE - 2; 452 smcibdev->roce_cq_send = ib_create_cq(smcibdev->ibdev, 453 smc_wr_tx_cq_handler, NULL, 454 smcibdev, &cqattr); 455 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_send); 456 if (IS_ERR(smcibdev->roce_cq_send)) { 457 smcibdev->roce_cq_send = NULL; 458 return rc; 459 } 460 smcibdev->roce_cq_recv = ib_create_cq(smcibdev->ibdev, 461 smc_wr_rx_cq_handler, NULL, 462 smcibdev, &cqattr); 463 rc = PTR_ERR_OR_ZERO(smcibdev->roce_cq_recv); 464 if (IS_ERR(smcibdev->roce_cq_recv)) { 465 smcibdev->roce_cq_recv = NULL; 466 goto err; 467 } 468 INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev, 469 smc_ib_global_event_handler); 470 ib_register_event_handler(&smcibdev->event_handler); 471 smc_wr_add_dev(smcibdev); 472 smcibdev->initialized = 1; 473 return rc; 474 475 err: 476 ib_destroy_cq(smcibdev->roce_cq_send); 477 return rc; 478 } 479 480 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev) 481 { 482 if (!smcibdev->initialized) 483 return; 484 smcibdev->initialized = 0; 485 smc_wr_remove_dev(smcibdev); 486 ib_unregister_event_handler(&smcibdev->event_handler); 487 ib_destroy_cq(smcibdev->roce_cq_recv); 488 ib_destroy_cq(smcibdev->roce_cq_send); 489 } 490 491 static struct ib_client smc_ib_client; 492 493 /* callback function for ib_register_client() */ 494 static void smc_ib_add_dev(struct ib_device *ibdev) 495 { 496 struct smc_ib_device *smcibdev; 497 498 if (ibdev->node_type != RDMA_NODE_IB_CA) 499 return; 500 501 smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL); 502 if (!smcibdev) 503 return; 504 505 smcibdev->ibdev = ibdev; 506 INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work); 507 508 spin_lock(&smc_ib_devices.lock); 509 list_add_tail(&smcibdev->list, &smc_ib_devices.list); 510 spin_unlock(&smc_ib_devices.lock); 511 ib_set_client_data(ibdev, &smc_ib_client, smcibdev); 512 } 513 514 /* callback function for ib_register_client() */ 515 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data) 516 { 517 struct smc_ib_device *smcibdev; 518 519 smcibdev = ib_get_client_data(ibdev, &smc_ib_client); 520 ib_set_client_data(ibdev, &smc_ib_client, NULL); 521 spin_lock(&smc_ib_devices.lock); 522 list_del_init(&smcibdev->list); /* remove from smc_ib_devices */ 523 spin_unlock(&smc_ib_devices.lock); 524 smc_pnet_remove_by_ibdev(smcibdev); 525 smc_ib_cleanup_per_ibdev(smcibdev); 526 kfree(smcibdev); 527 } 528 529 static struct ib_client smc_ib_client = { 530 .name = "smc_ib", 531 .add = smc_ib_add_dev, 532 .remove = smc_ib_remove_dev, 533 }; 534 535 int __init smc_ib_register_client(void) 536 { 537 return ib_register_client(&smc_ib_client); 538 } 539 540 void smc_ib_unregister_client(void) 541 { 542 ib_unregister_client(&smc_ib_client); 543 } 544