1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * RDMA Transport Layer 4 * 5 * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved. 6 * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved. 7 * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved. 8 */ 9 #undef pr_fmt 10 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt 11 12 #include <linux/module.h> 13 #include <linux/inet.h> 14 15 #include "rtrs-pri.h" 16 #include "rtrs-log.h" 17 18 MODULE_DESCRIPTION("RDMA Transport Core"); 19 MODULE_LICENSE("GPL"); 20 21 struct rtrs_iu *rtrs_iu_alloc(u32 iu_num, size_t size, gfp_t gfp_mask, 22 struct ib_device *dma_dev, 23 enum dma_data_direction dir, 24 void (*done)(struct ib_cq *cq, struct ib_wc *wc)) 25 { 26 struct rtrs_iu *ius, *iu; 27 int i; 28 29 ius = kcalloc(iu_num, sizeof(*ius), gfp_mask); 30 if (!ius) 31 return NULL; 32 for (i = 0; i < iu_num; i++) { 33 iu = &ius[i]; 34 iu->direction = dir; 35 iu->buf = kzalloc(size, gfp_mask); 36 if (!iu->buf) 37 goto err; 38 39 iu->dma_addr = ib_dma_map_single(dma_dev, iu->buf, size, dir); 40 if (ib_dma_mapping_error(dma_dev, iu->dma_addr)) 41 goto err; 42 43 iu->cqe.done = done; 44 iu->size = size; 45 } 46 return ius; 47 err: 48 rtrs_iu_free(ius, dma_dev, i); 49 return NULL; 50 } 51 EXPORT_SYMBOL_GPL(rtrs_iu_alloc); 52 53 void rtrs_iu_free(struct rtrs_iu *ius, struct ib_device *ibdev, u32 queue_num) 54 { 55 struct rtrs_iu *iu; 56 int i; 57 58 if (!ius) 59 return; 60 61 for (i = 0; i < queue_num; i++) { 62 iu = &ius[i]; 63 ib_dma_unmap_single(ibdev, iu->dma_addr, iu->size, iu->direction); 64 kfree(iu->buf); 65 } 66 kfree(ius); 67 } 68 EXPORT_SYMBOL_GPL(rtrs_iu_free); 69 70 int rtrs_iu_post_recv(struct rtrs_con *con, struct rtrs_iu *iu) 71 { 72 struct rtrs_sess *sess = con->sess; 73 struct ib_recv_wr wr; 74 struct ib_sge list; 75 76 list.addr = iu->dma_addr; 77 list.length = iu->size; 78 list.lkey = sess->dev->ib_pd->local_dma_lkey; 79 80 if (list.length == 0) { 81 rtrs_wrn(con->sess, 82 "Posting receive work request failed, sg list is empty\n"); 83 return -EINVAL; 84 } 85 wr = (struct ib_recv_wr) { 86 .wr_cqe = &iu->cqe, 87 .sg_list = &list, 88 .num_sge = 1, 89 }; 90 91 return ib_post_recv(con->qp, &wr, NULL); 92 } 93 EXPORT_SYMBOL_GPL(rtrs_iu_post_recv); 94 95 int rtrs_post_recv_empty(struct rtrs_con *con, struct ib_cqe *cqe) 96 { 97 struct ib_recv_wr wr; 98 99 wr = (struct ib_recv_wr) { 100 .wr_cqe = cqe, 101 }; 102 103 return ib_post_recv(con->qp, &wr, NULL); 104 } 105 EXPORT_SYMBOL_GPL(rtrs_post_recv_empty); 106 107 static int rtrs_post_send(struct ib_qp *qp, struct ib_send_wr *head, 108 struct ib_send_wr *wr, struct ib_send_wr *tail) 109 { 110 if (head) { 111 struct ib_send_wr *next = head; 112 113 while (next->next) 114 next = next->next; 115 next->next = wr; 116 } else { 117 head = wr; 118 } 119 120 if (tail) 121 wr->next = tail; 122 123 return ib_post_send(qp, head, NULL); 124 } 125 126 int rtrs_iu_post_send(struct rtrs_con *con, struct rtrs_iu *iu, size_t size, 127 struct ib_send_wr *head) 128 { 129 struct rtrs_sess *sess = con->sess; 130 struct ib_send_wr wr; 131 struct ib_sge list; 132 133 if (WARN_ON(size == 0)) 134 return -EINVAL; 135 136 list.addr = iu->dma_addr; 137 list.length = size; 138 list.lkey = sess->dev->ib_pd->local_dma_lkey; 139 140 wr = (struct ib_send_wr) { 141 .wr_cqe = &iu->cqe, 142 .sg_list = &list, 143 .num_sge = 1, 144 .opcode = IB_WR_SEND, 145 .send_flags = IB_SEND_SIGNALED, 146 }; 147 148 return rtrs_post_send(con->qp, head, &wr, NULL); 149 } 150 EXPORT_SYMBOL_GPL(rtrs_iu_post_send); 151 152 int rtrs_iu_post_rdma_write_imm(struct rtrs_con *con, struct rtrs_iu *iu, 153 struct ib_sge *sge, unsigned int num_sge, 154 u32 rkey, u64 rdma_addr, u32 imm_data, 155 enum ib_send_flags flags, 156 struct ib_send_wr *head, 157 struct ib_send_wr *tail) 158 { 159 struct ib_rdma_wr wr; 160 int i; 161 162 wr = (struct ib_rdma_wr) { 163 .wr.wr_cqe = &iu->cqe, 164 .wr.sg_list = sge, 165 .wr.num_sge = num_sge, 166 .rkey = rkey, 167 .remote_addr = rdma_addr, 168 .wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM, 169 .wr.ex.imm_data = cpu_to_be32(imm_data), 170 .wr.send_flags = flags, 171 }; 172 173 /* 174 * If one of the sges has 0 size, the operation will fail with a 175 * length error 176 */ 177 for (i = 0; i < num_sge; i++) 178 if (WARN_ON(sge[i].length == 0)) 179 return -EINVAL; 180 181 return rtrs_post_send(con->qp, head, &wr.wr, tail); 182 } 183 EXPORT_SYMBOL_GPL(rtrs_iu_post_rdma_write_imm); 184 185 int rtrs_post_rdma_write_imm_empty(struct rtrs_con *con, struct ib_cqe *cqe, 186 u32 imm_data, enum ib_send_flags flags, 187 struct ib_send_wr *head) 188 { 189 struct ib_rdma_wr wr; 190 191 wr = (struct ib_rdma_wr) { 192 .wr.wr_cqe = cqe, 193 .wr.send_flags = flags, 194 .wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM, 195 .wr.ex.imm_data = cpu_to_be32(imm_data), 196 }; 197 198 return rtrs_post_send(con->qp, head, &wr.wr, NULL); 199 } 200 EXPORT_SYMBOL_GPL(rtrs_post_rdma_write_imm_empty); 201 202 static void qp_event_handler(struct ib_event *ev, void *ctx) 203 { 204 struct rtrs_con *con = ctx; 205 206 switch (ev->event) { 207 case IB_EVENT_COMM_EST: 208 rtrs_info(con->sess, "QP event %s (%d) received\n", 209 ib_event_msg(ev->event), ev->event); 210 rdma_notify(con->cm_id, IB_EVENT_COMM_EST); 211 break; 212 default: 213 rtrs_info(con->sess, "Unhandled QP event %s (%d) received\n", 214 ib_event_msg(ev->event), ev->event); 215 break; 216 } 217 } 218 219 static int create_cq(struct rtrs_con *con, int cq_vector, int nr_cqe, 220 enum ib_poll_context poll_ctx) 221 { 222 struct rdma_cm_id *cm_id = con->cm_id; 223 struct ib_cq *cq; 224 225 cq = ib_cq_pool_get(cm_id->device, nr_cqe, cq_vector, poll_ctx); 226 if (IS_ERR(cq)) { 227 rtrs_err(con->sess, "Creating completion queue failed, errno: %ld\n", 228 PTR_ERR(cq)); 229 return PTR_ERR(cq); 230 } 231 con->cq = cq; 232 con->nr_cqe = nr_cqe; 233 234 return 0; 235 } 236 237 static int create_qp(struct rtrs_con *con, struct ib_pd *pd, 238 u32 max_send_wr, u32 max_recv_wr, u32 max_sge) 239 { 240 struct ib_qp_init_attr init_attr = {NULL}; 241 struct rdma_cm_id *cm_id = con->cm_id; 242 int ret; 243 244 init_attr.cap.max_send_wr = max_send_wr; 245 init_attr.cap.max_recv_wr = max_recv_wr; 246 init_attr.cap.max_recv_sge = 1; 247 init_attr.event_handler = qp_event_handler; 248 init_attr.qp_context = con; 249 init_attr.cap.max_send_sge = max_sge; 250 251 init_attr.qp_type = IB_QPT_RC; 252 init_attr.send_cq = con->cq; 253 init_attr.recv_cq = con->cq; 254 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 255 256 ret = rdma_create_qp(cm_id, pd, &init_attr); 257 if (ret) { 258 rtrs_err(con->sess, "Creating QP failed, err: %d\n", ret); 259 return ret; 260 } 261 con->qp = cm_id->qp; 262 263 return ret; 264 } 265 266 int rtrs_cq_qp_create(struct rtrs_sess *sess, struct rtrs_con *con, 267 u32 max_send_sge, int cq_vector, int nr_cqe, 268 u32 max_send_wr, u32 max_recv_wr, 269 enum ib_poll_context poll_ctx) 270 { 271 int err; 272 273 err = create_cq(con, cq_vector, nr_cqe, poll_ctx); 274 if (err) 275 return err; 276 277 err = create_qp(con, sess->dev->ib_pd, max_send_wr, max_recv_wr, 278 max_send_sge); 279 if (err) { 280 ib_cq_pool_put(con->cq, con->nr_cqe); 281 con->cq = NULL; 282 return err; 283 } 284 con->sess = sess; 285 286 return 0; 287 } 288 EXPORT_SYMBOL_GPL(rtrs_cq_qp_create); 289 290 void rtrs_cq_qp_destroy(struct rtrs_con *con) 291 { 292 if (con->qp) { 293 rdma_destroy_qp(con->cm_id); 294 con->qp = NULL; 295 } 296 if (con->cq) { 297 ib_cq_pool_put(con->cq, con->nr_cqe); 298 con->cq = NULL; 299 } 300 } 301 EXPORT_SYMBOL_GPL(rtrs_cq_qp_destroy); 302 303 static void schedule_hb(struct rtrs_sess *sess) 304 { 305 queue_delayed_work(sess->hb_wq, &sess->hb_dwork, 306 msecs_to_jiffies(sess->hb_interval_ms)); 307 } 308 309 void rtrs_send_hb_ack(struct rtrs_sess *sess) 310 { 311 struct rtrs_con *usr_con = sess->con[0]; 312 u32 imm; 313 int err; 314 315 imm = rtrs_to_imm(RTRS_HB_ACK_IMM, 0); 316 err = rtrs_post_rdma_write_imm_empty(usr_con, sess->hb_cqe, imm, 317 0, NULL); 318 if (err) { 319 sess->hb_err_handler(usr_con); 320 return; 321 } 322 } 323 EXPORT_SYMBOL_GPL(rtrs_send_hb_ack); 324 325 static void hb_work(struct work_struct *work) 326 { 327 struct rtrs_con *usr_con; 328 struct rtrs_sess *sess; 329 u32 imm; 330 int err; 331 332 sess = container_of(to_delayed_work(work), typeof(*sess), hb_dwork); 333 usr_con = sess->con[0]; 334 335 if (sess->hb_missed_cnt > sess->hb_missed_max) { 336 sess->hb_err_handler(usr_con); 337 return; 338 } 339 if (sess->hb_missed_cnt++) { 340 /* Reschedule work without sending hb */ 341 schedule_hb(sess); 342 return; 343 } 344 345 sess->hb_last_sent = ktime_get(); 346 347 imm = rtrs_to_imm(RTRS_HB_MSG_IMM, 0); 348 err = rtrs_post_rdma_write_imm_empty(usr_con, sess->hb_cqe, imm, 349 0, NULL); 350 if (err) { 351 sess->hb_err_handler(usr_con); 352 return; 353 } 354 355 schedule_hb(sess); 356 } 357 358 void rtrs_init_hb(struct rtrs_sess *sess, struct ib_cqe *cqe, 359 unsigned int interval_ms, unsigned int missed_max, 360 void (*err_handler)(struct rtrs_con *con), 361 struct workqueue_struct *wq) 362 { 363 sess->hb_cqe = cqe; 364 sess->hb_interval_ms = interval_ms; 365 sess->hb_err_handler = err_handler; 366 sess->hb_wq = wq; 367 sess->hb_missed_max = missed_max; 368 sess->hb_missed_cnt = 0; 369 INIT_DELAYED_WORK(&sess->hb_dwork, hb_work); 370 } 371 EXPORT_SYMBOL_GPL(rtrs_init_hb); 372 373 void rtrs_start_hb(struct rtrs_sess *sess) 374 { 375 schedule_hb(sess); 376 } 377 EXPORT_SYMBOL_GPL(rtrs_start_hb); 378 379 void rtrs_stop_hb(struct rtrs_sess *sess) 380 { 381 cancel_delayed_work_sync(&sess->hb_dwork); 382 sess->hb_missed_cnt = 0; 383 } 384 EXPORT_SYMBOL_GPL(rtrs_stop_hb); 385 386 static int rtrs_str_gid_to_sockaddr(const char *addr, size_t len, 387 short port, struct sockaddr_storage *dst) 388 { 389 struct sockaddr_ib *dst_ib = (struct sockaddr_ib *)dst; 390 int ret; 391 392 /* 393 * We can use some of the IPv6 functions since GID is a valid 394 * IPv6 address format 395 */ 396 ret = in6_pton(addr, len, dst_ib->sib_addr.sib_raw, '\0', NULL); 397 if (ret == 0) 398 return -EINVAL; 399 400 dst_ib->sib_family = AF_IB; 401 /* 402 * Use the same TCP server port number as the IB service ID 403 * on the IB port space range 404 */ 405 dst_ib->sib_sid = cpu_to_be64(RDMA_IB_IP_PS_IB | port); 406 dst_ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL); 407 dst_ib->sib_pkey = cpu_to_be16(0xffff); 408 409 return 0; 410 } 411 412 /** 413 * rtrs_str_to_sockaddr() - Convert rtrs address string to sockaddr 414 * @addr: String representation of an addr (IPv4, IPv6 or IB GID): 415 * - "ip:192.168.1.1" 416 * - "ip:fe80::200:5aee:feaa:20a2" 417 * - "gid:fe80::200:5aee:feaa:20a2" 418 * @len: String address length 419 * @port: Destination port 420 * @dst: Destination sockaddr structure 421 * 422 * Returns 0 if conversion successful. Non-zero on error. 423 */ 424 static int rtrs_str_to_sockaddr(const char *addr, size_t len, 425 u16 port, struct sockaddr_storage *dst) 426 { 427 if (strncmp(addr, "gid:", 4) == 0) { 428 return rtrs_str_gid_to_sockaddr(addr + 4, len - 4, port, dst); 429 } else if (strncmp(addr, "ip:", 3) == 0) { 430 char port_str[8]; 431 char *cpy; 432 int err; 433 434 snprintf(port_str, sizeof(port_str), "%u", port); 435 cpy = kstrndup(addr + 3, len - 3, GFP_KERNEL); 436 err = cpy ? inet_pton_with_scope(&init_net, AF_UNSPEC, 437 cpy, port_str, dst) : -ENOMEM; 438 kfree(cpy); 439 440 return err; 441 } 442 return -EPROTONOSUPPORT; 443 } 444 445 /** 446 * sockaddr_to_str() - convert sockaddr to a string. 447 * @addr: the sockadddr structure to be converted. 448 * @buf: string containing socket addr. 449 * @len: string length. 450 * 451 * The return value is the number of characters written into buf not 452 * including the trailing '\0'. If len is == 0 the function returns 0.. 453 */ 454 int sockaddr_to_str(const struct sockaddr *addr, char *buf, size_t len) 455 { 456 457 switch (addr->sa_family) { 458 case AF_IB: 459 return scnprintf(buf, len, "gid:%pI6", 460 &((struct sockaddr_ib *)addr)->sib_addr.sib_raw); 461 case AF_INET: 462 return scnprintf(buf, len, "ip:%pI4", 463 &((struct sockaddr_in *)addr)->sin_addr); 464 case AF_INET6: 465 return scnprintf(buf, len, "ip:%pI6c", 466 &((struct sockaddr_in6 *)addr)->sin6_addr); 467 } 468 return scnprintf(buf, len, "<invalid address family>"); 469 } 470 EXPORT_SYMBOL(sockaddr_to_str); 471 472 /** 473 * rtrs_addr_to_str() - convert rtrs_addr to a string "src@dst" 474 * @addr: the rtrs_addr structure to be converted 475 * @buf: string containing source and destination addr of a path 476 * separated by '@' I.e. "ip:1.1.1.1@ip:1.1.1.2" 477 * "ip:1.1.1.1@ip:1.1.1.2". 478 * @len: string length 479 * 480 * The return value is the number of characters written into buf not 481 * including the trailing '\0'. 482 */ 483 int rtrs_addr_to_str(const struct rtrs_addr *addr, char *buf, size_t len) 484 { 485 int cnt; 486 487 cnt = sockaddr_to_str((struct sockaddr *)addr->src, 488 buf, len); 489 cnt += scnprintf(buf + cnt, len - cnt, "@"); 490 sockaddr_to_str((struct sockaddr *)addr->dst, 491 buf + cnt, len - cnt); 492 return cnt; 493 } 494 EXPORT_SYMBOL(rtrs_addr_to_str); 495 496 /** 497 * rtrs_addr_to_sockaddr() - convert path string "src,dst" or "src@dst" 498 * to sockaddreses 499 * @str: string containing source and destination addr of a path 500 * separated by ',' or '@' I.e. "ip:1.1.1.1,ip:1.1.1.2" or 501 * "ip:1.1.1.1@ip:1.1.1.2". If str contains only one address it's 502 * considered to be destination. 503 * @len: string length 504 * @port: Destination port number. 505 * @addr: will be set to the source/destination address or to NULL 506 * if str doesn't contain any source address. 507 * 508 * Returns zero if conversion successful. Non-zero otherwise. 509 */ 510 int rtrs_addr_to_sockaddr(const char *str, size_t len, u16 port, 511 struct rtrs_addr *addr) 512 { 513 const char *d; 514 515 d = strchr(str, ','); 516 if (!d) 517 d = strchr(str, '@'); 518 if (d) { 519 if (rtrs_str_to_sockaddr(str, d - str, 0, addr->src)) 520 return -EINVAL; 521 d += 1; 522 len -= d - str; 523 str = d; 524 525 } else { 526 addr->src = NULL; 527 } 528 return rtrs_str_to_sockaddr(str, len, port, addr->dst); 529 } 530 EXPORT_SYMBOL(rtrs_addr_to_sockaddr); 531 532 void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags, 533 struct rtrs_rdma_dev_pd *pool) 534 { 535 WARN_ON(pool->ops && (!pool->ops->alloc ^ !pool->ops->free)); 536 INIT_LIST_HEAD(&pool->list); 537 mutex_init(&pool->mutex); 538 pool->pd_flags = pd_flags; 539 } 540 EXPORT_SYMBOL(rtrs_rdma_dev_pd_init); 541 542 void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool) 543 { 544 mutex_destroy(&pool->mutex); 545 WARN_ON(!list_empty(&pool->list)); 546 } 547 EXPORT_SYMBOL(rtrs_rdma_dev_pd_deinit); 548 549 static void dev_free(struct kref *ref) 550 { 551 struct rtrs_rdma_dev_pd *pool; 552 struct rtrs_ib_dev *dev; 553 554 dev = container_of(ref, typeof(*dev), ref); 555 pool = dev->pool; 556 557 mutex_lock(&pool->mutex); 558 list_del(&dev->entry); 559 mutex_unlock(&pool->mutex); 560 561 if (pool->ops && pool->ops->deinit) 562 pool->ops->deinit(dev); 563 564 ib_dealloc_pd(dev->ib_pd); 565 566 if (pool->ops && pool->ops->free) 567 pool->ops->free(dev); 568 else 569 kfree(dev); 570 } 571 572 int rtrs_ib_dev_put(struct rtrs_ib_dev *dev) 573 { 574 return kref_put(&dev->ref, dev_free); 575 } 576 EXPORT_SYMBOL(rtrs_ib_dev_put); 577 578 static int rtrs_ib_dev_get(struct rtrs_ib_dev *dev) 579 { 580 return kref_get_unless_zero(&dev->ref); 581 } 582 583 struct rtrs_ib_dev * 584 rtrs_ib_dev_find_or_add(struct ib_device *ib_dev, 585 struct rtrs_rdma_dev_pd *pool) 586 { 587 struct rtrs_ib_dev *dev; 588 589 mutex_lock(&pool->mutex); 590 list_for_each_entry(dev, &pool->list, entry) { 591 if (dev->ib_dev->node_guid == ib_dev->node_guid && 592 rtrs_ib_dev_get(dev)) 593 goto out_unlock; 594 } 595 mutex_unlock(&pool->mutex); 596 if (pool->ops && pool->ops->alloc) 597 dev = pool->ops->alloc(); 598 else 599 dev = kzalloc(sizeof(*dev), GFP_KERNEL); 600 if (IS_ERR_OR_NULL(dev)) 601 goto out_err; 602 603 kref_init(&dev->ref); 604 dev->pool = pool; 605 dev->ib_dev = ib_dev; 606 dev->ib_pd = ib_alloc_pd(ib_dev, pool->pd_flags); 607 if (IS_ERR(dev->ib_pd)) 608 goto out_free_dev; 609 610 if (pool->ops && pool->ops->init && pool->ops->init(dev)) 611 goto out_free_pd; 612 613 mutex_lock(&pool->mutex); 614 list_add(&dev->entry, &pool->list); 615 out_unlock: 616 mutex_unlock(&pool->mutex); 617 return dev; 618 619 out_free_pd: 620 ib_dealloc_pd(dev->ib_pd); 621 out_free_dev: 622 if (pool->ops && pool->ops->free) 623 pool->ops->free(dev); 624 else 625 kfree(dev); 626 out_err: 627 return NULL; 628 } 629 EXPORT_SYMBOL(rtrs_ib_dev_find_or_add); 630