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 /* process context wrapper for might_sleep smc_ib_remember_port_attr */ 147 static void smc_ib_port_event_work(struct work_struct *work) 148 { 149 struct smc_ib_device *smcibdev = container_of( 150 work, struct smc_ib_device, port_event_work); 151 u8 port_idx; 152 153 for_each_set_bit(port_idx, &smcibdev->port_event_mask, SMC_MAX_PORTS) { 154 smc_ib_remember_port_attr(smcibdev, port_idx + 1); 155 clear_bit(port_idx, &smcibdev->port_event_mask); 156 if (!smc_ib_port_active(smcibdev, port_idx + 1)) 157 smc_port_terminate(smcibdev, port_idx + 1); 158 } 159 } 160 161 /* can be called in IRQ context */ 162 static void smc_ib_global_event_handler(struct ib_event_handler *handler, 163 struct ib_event *ibevent) 164 { 165 struct smc_ib_device *smcibdev; 166 u8 port_idx; 167 168 smcibdev = container_of(handler, struct smc_ib_device, event_handler); 169 170 switch (ibevent->event) { 171 case IB_EVENT_PORT_ERR: 172 case IB_EVENT_DEVICE_FATAL: 173 case IB_EVENT_PORT_ACTIVE: 174 port_idx = ibevent->element.port_num - 1; 175 set_bit(port_idx, &smcibdev->port_event_mask); 176 schedule_work(&smcibdev->port_event_work); 177 break; 178 default: 179 break; 180 } 181 } 182 183 void smc_ib_dealloc_protection_domain(struct smc_link *lnk) 184 { 185 if (lnk->roce_pd) 186 ib_dealloc_pd(lnk->roce_pd); 187 lnk->roce_pd = NULL; 188 } 189 190 int smc_ib_create_protection_domain(struct smc_link *lnk) 191 { 192 int rc; 193 194 lnk->roce_pd = ib_alloc_pd(lnk->smcibdev->ibdev, 0); 195 rc = PTR_ERR_OR_ZERO(lnk->roce_pd); 196 if (IS_ERR(lnk->roce_pd)) 197 lnk->roce_pd = NULL; 198 return rc; 199 } 200 201 static void smc_ib_qp_event_handler(struct ib_event *ibevent, void *priv) 202 { 203 struct smc_ib_device *smcibdev = 204 (struct smc_ib_device *)ibevent->device; 205 u8 port_idx; 206 207 switch (ibevent->event) { 208 case IB_EVENT_DEVICE_FATAL: 209 case IB_EVENT_GID_CHANGE: 210 case IB_EVENT_PORT_ERR: 211 case IB_EVENT_QP_ACCESS_ERR: 212 port_idx = ibevent->element.port_num - 1; 213 set_bit(port_idx, &smcibdev->port_event_mask); 214 schedule_work(&smcibdev->port_event_work); 215 break; 216 default: 217 break; 218 } 219 } 220 221 void smc_ib_destroy_queue_pair(struct smc_link *lnk) 222 { 223 if (lnk->roce_qp) 224 ib_destroy_qp(lnk->roce_qp); 225 lnk->roce_qp = NULL; 226 } 227 228 /* create a queue pair within the protection domain for a link */ 229 int smc_ib_create_queue_pair(struct smc_link *lnk) 230 { 231 struct ib_qp_init_attr qp_attr = { 232 .event_handler = smc_ib_qp_event_handler, 233 .qp_context = lnk, 234 .send_cq = lnk->smcibdev->roce_cq_send, 235 .recv_cq = lnk->smcibdev->roce_cq_recv, 236 .srq = NULL, 237 .cap = { 238 /* include unsolicited rdma_writes as well, 239 * there are max. 2 RDMA_WRITE per 1 WR_SEND 240 */ 241 .max_send_wr = SMC_WR_BUF_CNT * 3, 242 .max_recv_wr = SMC_WR_BUF_CNT * 3, 243 .max_send_sge = SMC_IB_MAX_SEND_SGE, 244 .max_recv_sge = 1, 245 }, 246 .sq_sig_type = IB_SIGNAL_REQ_WR, 247 .qp_type = IB_QPT_RC, 248 }; 249 int rc; 250 251 lnk->roce_qp = ib_create_qp(lnk->roce_pd, &qp_attr); 252 rc = PTR_ERR_OR_ZERO(lnk->roce_qp); 253 if (IS_ERR(lnk->roce_qp)) 254 lnk->roce_qp = NULL; 255 else 256 smc_wr_remember_qp_attr(lnk); 257 return rc; 258 } 259 260 void smc_ib_put_memory_region(struct ib_mr *mr) 261 { 262 ib_dereg_mr(mr); 263 } 264 265 static int smc_ib_map_mr_sg(struct smc_buf_desc *buf_slot) 266 { 267 unsigned int offset = 0; 268 int sg_num; 269 270 /* map the largest prefix of a dma mapped SG list */ 271 sg_num = ib_map_mr_sg(buf_slot->mr_rx[SMC_SINGLE_LINK], 272 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 273 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 274 &offset, PAGE_SIZE); 275 276 return sg_num; 277 } 278 279 /* Allocate a memory region and map the dma mapped SG list of buf_slot */ 280 int smc_ib_get_memory_region(struct ib_pd *pd, int access_flags, 281 struct smc_buf_desc *buf_slot) 282 { 283 if (buf_slot->mr_rx[SMC_SINGLE_LINK]) 284 return 0; /* already done */ 285 286 buf_slot->mr_rx[SMC_SINGLE_LINK] = 287 ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, 1 << buf_slot->order); 288 if (IS_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK])) { 289 int rc; 290 291 rc = PTR_ERR(buf_slot->mr_rx[SMC_SINGLE_LINK]); 292 buf_slot->mr_rx[SMC_SINGLE_LINK] = NULL; 293 return rc; 294 } 295 296 if (smc_ib_map_mr_sg(buf_slot) != 1) 297 return -EINVAL; 298 299 return 0; 300 } 301 302 /* synchronize buffer usage for cpu access */ 303 void smc_ib_sync_sg_for_cpu(struct smc_ib_device *smcibdev, 304 struct smc_buf_desc *buf_slot, 305 enum dma_data_direction data_direction) 306 { 307 struct scatterlist *sg; 308 unsigned int i; 309 310 /* for now there is just one DMA address */ 311 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 312 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 313 if (!sg_dma_len(sg)) 314 break; 315 ib_dma_sync_single_for_cpu(smcibdev->ibdev, 316 sg_dma_address(sg), 317 sg_dma_len(sg), 318 data_direction); 319 } 320 } 321 322 /* synchronize buffer usage for device access */ 323 void smc_ib_sync_sg_for_device(struct smc_ib_device *smcibdev, 324 struct smc_buf_desc *buf_slot, 325 enum dma_data_direction data_direction) 326 { 327 struct scatterlist *sg; 328 unsigned int i; 329 330 /* for now there is just one DMA address */ 331 for_each_sg(buf_slot->sgt[SMC_SINGLE_LINK].sgl, sg, 332 buf_slot->sgt[SMC_SINGLE_LINK].nents, i) { 333 if (!sg_dma_len(sg)) 334 break; 335 ib_dma_sync_single_for_device(smcibdev->ibdev, 336 sg_dma_address(sg), 337 sg_dma_len(sg), 338 data_direction); 339 } 340 } 341 342 /* Map a new TX or RX buffer SG-table to DMA */ 343 int smc_ib_buf_map_sg(struct smc_ib_device *smcibdev, 344 struct smc_buf_desc *buf_slot, 345 enum dma_data_direction data_direction) 346 { 347 int mapped_nents; 348 349 mapped_nents = ib_dma_map_sg(smcibdev->ibdev, 350 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 351 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 352 data_direction); 353 if (!mapped_nents) 354 return -ENOMEM; 355 356 return mapped_nents; 357 } 358 359 void smc_ib_buf_unmap_sg(struct smc_ib_device *smcibdev, 360 struct smc_buf_desc *buf_slot, 361 enum dma_data_direction data_direction) 362 { 363 if (!buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address) 364 return; /* already unmapped */ 365 366 ib_dma_unmap_sg(smcibdev->ibdev, 367 buf_slot->sgt[SMC_SINGLE_LINK].sgl, 368 buf_slot->sgt[SMC_SINGLE_LINK].orig_nents, 369 data_direction); 370 buf_slot->sgt[SMC_SINGLE_LINK].sgl->dma_address = 0; 371 } 372 373 static int smc_ib_fill_gid_and_mac(struct smc_ib_device *smcibdev, u8 ibport) 374 { 375 struct ib_gid_attr gattr; 376 int rc; 377 378 rc = ib_query_gid(smcibdev->ibdev, ibport, 0, 379 &smcibdev->gid[ibport - 1], &gattr); 380 if (rc || !gattr.ndev) 381 return -ENODEV; 382 383 memcpy(smcibdev->mac[ibport - 1], gattr.ndev->dev_addr, ETH_ALEN); 384 dev_put(gattr.ndev); 385 return 0; 386 } 387 388 /* Create an identifier unique for this instance of SMC-R. 389 * The MAC-address of the first active registered IB device 390 * plus a random 2-byte number is used to create this identifier. 391 * This name is delivered to the peer during connection initialization. 392 */ 393 static inline void smc_ib_define_local_systemid(struct smc_ib_device *smcibdev, 394 u8 ibport) 395 { 396 memcpy(&local_systemid[2], &smcibdev->mac[ibport - 1], 397 sizeof(smcibdev->mac[ibport - 1])); 398 get_random_bytes(&local_systemid[0], 2); 399 } 400 401 bool smc_ib_port_active(struct smc_ib_device *smcibdev, u8 ibport) 402 { 403 return smcibdev->pattr[ibport - 1].state == IB_PORT_ACTIVE; 404 } 405 406 int smc_ib_remember_port_attr(struct smc_ib_device *smcibdev, u8 ibport) 407 { 408 int rc; 409 410 memset(&smcibdev->pattr[ibport - 1], 0, 411 sizeof(smcibdev->pattr[ibport - 1])); 412 rc = ib_query_port(smcibdev->ibdev, ibport, 413 &smcibdev->pattr[ibport - 1]); 414 if (rc) 415 goto out; 416 /* the SMC protocol requires specification of the RoCE MAC address */ 417 rc = smc_ib_fill_gid_and_mac(smcibdev, ibport); 418 if (rc) 419 goto out; 420 if (!strncmp(local_systemid, SMC_LOCAL_SYSTEMID_RESET, 421 sizeof(local_systemid)) && 422 smc_ib_port_active(smcibdev, ibport)) 423 /* create unique system identifier */ 424 smc_ib_define_local_systemid(smcibdev, ibport); 425 out: 426 return rc; 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 INIT_IB_EVENT_HANDLER(&smcibdev->event_handler, smcibdev->ibdev, 458 smc_ib_global_event_handler); 459 ib_register_event_handler(&smcibdev->event_handler); 460 smc_wr_add_dev(smcibdev); 461 smcibdev->initialized = 1; 462 return rc; 463 464 err: 465 ib_destroy_cq(smcibdev->roce_cq_send); 466 return rc; 467 } 468 469 static void smc_ib_cleanup_per_ibdev(struct smc_ib_device *smcibdev) 470 { 471 if (!smcibdev->initialized) 472 return; 473 smcibdev->initialized = 0; 474 smc_wr_remove_dev(smcibdev); 475 ib_unregister_event_handler(&smcibdev->event_handler); 476 ib_destroy_cq(smcibdev->roce_cq_recv); 477 ib_destroy_cq(smcibdev->roce_cq_send); 478 } 479 480 static struct ib_client smc_ib_client; 481 482 /* callback function for ib_register_client() */ 483 static void smc_ib_add_dev(struct ib_device *ibdev) 484 { 485 struct smc_ib_device *smcibdev; 486 487 if (ibdev->node_type != RDMA_NODE_IB_CA) 488 return; 489 490 smcibdev = kzalloc(sizeof(*smcibdev), GFP_KERNEL); 491 if (!smcibdev) 492 return; 493 494 smcibdev->ibdev = ibdev; 495 INIT_WORK(&smcibdev->port_event_work, smc_ib_port_event_work); 496 497 spin_lock(&smc_ib_devices.lock); 498 list_add_tail(&smcibdev->list, &smc_ib_devices.list); 499 spin_unlock(&smc_ib_devices.lock); 500 ib_set_client_data(ibdev, &smc_ib_client, smcibdev); 501 } 502 503 /* callback function for ib_register_client() */ 504 static void smc_ib_remove_dev(struct ib_device *ibdev, void *client_data) 505 { 506 struct smc_ib_device *smcibdev; 507 508 smcibdev = ib_get_client_data(ibdev, &smc_ib_client); 509 ib_set_client_data(ibdev, &smc_ib_client, NULL); 510 spin_lock(&smc_ib_devices.lock); 511 list_del_init(&smcibdev->list); /* remove from smc_ib_devices */ 512 spin_unlock(&smc_ib_devices.lock); 513 smc_pnet_remove_by_ibdev(smcibdev); 514 smc_ib_cleanup_per_ibdev(smcibdev); 515 kfree(smcibdev); 516 } 517 518 static struct ib_client smc_ib_client = { 519 .name = "smc_ib", 520 .add = smc_ib_add_dev, 521 .remove = smc_ib_remove_dev, 522 }; 523 524 int __init smc_ib_register_client(void) 525 { 526 return ib_register_client(&smc_ib_client); 527 } 528 529 void smc_ib_unregister_client(void) 530 { 531 ib_unregister_client(&smc_ib_client); 532 } 533