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 static int rtrs_post_rdma_write_imm_empty(struct rtrs_con *con, 186 struct ib_cqe *cqe, 187 u32 imm_data, 188 struct ib_send_wr *head) 189 { 190 struct ib_rdma_wr wr; 191 struct rtrs_sess *sess = con->sess; 192 enum ib_send_flags sflags; 193 194 atomic_dec_if_positive(&con->sq_wr_avail); 195 sflags = (atomic_inc_return(&con->wr_cnt) % sess->signal_interval) ? 196 0 : IB_SEND_SIGNALED; 197 198 wr = (struct ib_rdma_wr) { 199 .wr.wr_cqe = cqe, 200 .wr.send_flags = sflags, 201 .wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM, 202 .wr.ex.imm_data = cpu_to_be32(imm_data), 203 }; 204 205 return rtrs_post_send(con->qp, head, &wr.wr, NULL); 206 } 207 208 static void qp_event_handler(struct ib_event *ev, void *ctx) 209 { 210 struct rtrs_con *con = ctx; 211 212 switch (ev->event) { 213 case IB_EVENT_COMM_EST: 214 rtrs_info(con->sess, "QP event %s (%d) received\n", 215 ib_event_msg(ev->event), ev->event); 216 rdma_notify(con->cm_id, IB_EVENT_COMM_EST); 217 break; 218 default: 219 rtrs_info(con->sess, "Unhandled QP event %s (%d) received\n", 220 ib_event_msg(ev->event), ev->event); 221 break; 222 } 223 } 224 225 static int create_cq(struct rtrs_con *con, int cq_vector, int nr_cqe, 226 enum ib_poll_context poll_ctx) 227 { 228 struct rdma_cm_id *cm_id = con->cm_id; 229 struct ib_cq *cq; 230 231 cq = ib_cq_pool_get(cm_id->device, nr_cqe, cq_vector, poll_ctx); 232 if (IS_ERR(cq)) { 233 rtrs_err(con->sess, "Creating completion queue failed, errno: %ld\n", 234 PTR_ERR(cq)); 235 return PTR_ERR(cq); 236 } 237 con->cq = cq; 238 con->nr_cqe = nr_cqe; 239 240 return 0; 241 } 242 243 static int create_qp(struct rtrs_con *con, struct ib_pd *pd, 244 u32 max_send_wr, u32 max_recv_wr, u32 max_sge) 245 { 246 struct ib_qp_init_attr init_attr = {NULL}; 247 struct rdma_cm_id *cm_id = con->cm_id; 248 int ret; 249 250 init_attr.cap.max_send_wr = max_send_wr; 251 init_attr.cap.max_recv_wr = max_recv_wr; 252 init_attr.cap.max_recv_sge = 1; 253 init_attr.event_handler = qp_event_handler; 254 init_attr.qp_context = con; 255 init_attr.cap.max_send_sge = max_sge; 256 257 init_attr.qp_type = IB_QPT_RC; 258 init_attr.send_cq = con->cq; 259 init_attr.recv_cq = con->cq; 260 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 261 262 ret = rdma_create_qp(cm_id, pd, &init_attr); 263 if (ret) { 264 rtrs_err(con->sess, "Creating QP failed, err: %d\n", ret); 265 return ret; 266 } 267 con->qp = cm_id->qp; 268 269 return ret; 270 } 271 272 int rtrs_cq_qp_create(struct rtrs_sess *sess, struct rtrs_con *con, 273 u32 max_send_sge, int cq_vector, int nr_cqe, 274 u32 max_send_wr, u32 max_recv_wr, 275 enum ib_poll_context poll_ctx) 276 { 277 int err; 278 279 err = create_cq(con, cq_vector, nr_cqe, poll_ctx); 280 if (err) 281 return err; 282 283 err = create_qp(con, sess->dev->ib_pd, max_send_wr, max_recv_wr, 284 max_send_sge); 285 if (err) { 286 ib_cq_pool_put(con->cq, con->nr_cqe); 287 con->cq = NULL; 288 return err; 289 } 290 con->sess = sess; 291 292 return 0; 293 } 294 EXPORT_SYMBOL_GPL(rtrs_cq_qp_create); 295 296 void rtrs_cq_qp_destroy(struct rtrs_con *con) 297 { 298 if (con->qp) { 299 rdma_destroy_qp(con->cm_id); 300 con->qp = NULL; 301 } 302 if (con->cq) { 303 ib_cq_pool_put(con->cq, con->nr_cqe); 304 con->cq = NULL; 305 } 306 } 307 EXPORT_SYMBOL_GPL(rtrs_cq_qp_destroy); 308 309 static void schedule_hb(struct rtrs_sess *sess) 310 { 311 queue_delayed_work(sess->hb_wq, &sess->hb_dwork, 312 msecs_to_jiffies(sess->hb_interval_ms)); 313 } 314 315 void rtrs_send_hb_ack(struct rtrs_sess *sess) 316 { 317 struct rtrs_con *usr_con = sess->con[0]; 318 u32 imm; 319 int err; 320 321 imm = rtrs_to_imm(RTRS_HB_ACK_IMM, 0); 322 err = rtrs_post_rdma_write_imm_empty(usr_con, sess->hb_cqe, imm, 323 NULL); 324 if (err) { 325 rtrs_err(sess, "send HB ACK failed, errno: %d\n", err); 326 sess->hb_err_handler(usr_con); 327 return; 328 } 329 } 330 EXPORT_SYMBOL_GPL(rtrs_send_hb_ack); 331 332 static void hb_work(struct work_struct *work) 333 { 334 struct rtrs_con *usr_con; 335 struct rtrs_sess *sess; 336 u32 imm; 337 int err; 338 339 sess = container_of(to_delayed_work(work), typeof(*sess), hb_dwork); 340 usr_con = sess->con[0]; 341 342 if (sess->hb_missed_cnt > sess->hb_missed_max) { 343 rtrs_err(sess, "HB missed max reached.\n"); 344 sess->hb_err_handler(usr_con); 345 return; 346 } 347 if (sess->hb_missed_cnt++) { 348 /* Reschedule work without sending hb */ 349 schedule_hb(sess); 350 return; 351 } 352 353 sess->hb_last_sent = ktime_get(); 354 355 imm = rtrs_to_imm(RTRS_HB_MSG_IMM, 0); 356 err = rtrs_post_rdma_write_imm_empty(usr_con, sess->hb_cqe, imm, 357 NULL); 358 if (err) { 359 rtrs_err(sess, "HB send failed, errno: %d\n", err); 360 sess->hb_err_handler(usr_con); 361 return; 362 } 363 364 schedule_hb(sess); 365 } 366 367 void rtrs_init_hb(struct rtrs_sess *sess, struct ib_cqe *cqe, 368 unsigned int interval_ms, unsigned int missed_max, 369 void (*err_handler)(struct rtrs_con *con), 370 struct workqueue_struct *wq) 371 { 372 sess->hb_cqe = cqe; 373 sess->hb_interval_ms = interval_ms; 374 sess->hb_err_handler = err_handler; 375 sess->hb_wq = wq; 376 sess->hb_missed_max = missed_max; 377 sess->hb_missed_cnt = 0; 378 INIT_DELAYED_WORK(&sess->hb_dwork, hb_work); 379 } 380 EXPORT_SYMBOL_GPL(rtrs_init_hb); 381 382 void rtrs_start_hb(struct rtrs_sess *sess) 383 { 384 schedule_hb(sess); 385 } 386 EXPORT_SYMBOL_GPL(rtrs_start_hb); 387 388 void rtrs_stop_hb(struct rtrs_sess *sess) 389 { 390 cancel_delayed_work_sync(&sess->hb_dwork); 391 sess->hb_missed_cnt = 0; 392 } 393 EXPORT_SYMBOL_GPL(rtrs_stop_hb); 394 395 static int rtrs_str_gid_to_sockaddr(const char *addr, size_t len, 396 short port, struct sockaddr_storage *dst) 397 { 398 struct sockaddr_ib *dst_ib = (struct sockaddr_ib *)dst; 399 int ret; 400 401 /* 402 * We can use some of the IPv6 functions since GID is a valid 403 * IPv6 address format 404 */ 405 ret = in6_pton(addr, len, dst_ib->sib_addr.sib_raw, '\0', NULL); 406 if (ret == 0) 407 return -EINVAL; 408 409 dst_ib->sib_family = AF_IB; 410 /* 411 * Use the same TCP server port number as the IB service ID 412 * on the IB port space range 413 */ 414 dst_ib->sib_sid = cpu_to_be64(RDMA_IB_IP_PS_IB | port); 415 dst_ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL); 416 dst_ib->sib_pkey = cpu_to_be16(0xffff); 417 418 return 0; 419 } 420 421 /** 422 * rtrs_str_to_sockaddr() - Convert rtrs address string to sockaddr 423 * @addr: String representation of an addr (IPv4, IPv6 or IB GID): 424 * - "ip:192.168.1.1" 425 * - "ip:fe80::200:5aee:feaa:20a2" 426 * - "gid:fe80::200:5aee:feaa:20a2" 427 * @len: String address length 428 * @port: Destination port 429 * @dst: Destination sockaddr structure 430 * 431 * Returns 0 if conversion successful. Non-zero on error. 432 */ 433 static int rtrs_str_to_sockaddr(const char *addr, size_t len, 434 u16 port, struct sockaddr_storage *dst) 435 { 436 if (strncmp(addr, "gid:", 4) == 0) { 437 return rtrs_str_gid_to_sockaddr(addr + 4, len - 4, port, dst); 438 } else if (strncmp(addr, "ip:", 3) == 0) { 439 char port_str[8]; 440 char *cpy; 441 int err; 442 443 snprintf(port_str, sizeof(port_str), "%u", port); 444 cpy = kstrndup(addr + 3, len - 3, GFP_KERNEL); 445 err = cpy ? inet_pton_with_scope(&init_net, AF_UNSPEC, 446 cpy, port_str, dst) : -ENOMEM; 447 kfree(cpy); 448 449 return err; 450 } 451 return -EPROTONOSUPPORT; 452 } 453 454 /** 455 * sockaddr_to_str() - convert sockaddr to a string. 456 * @addr: the sockadddr structure to be converted. 457 * @buf: string containing socket addr. 458 * @len: string length. 459 * 460 * The return value is the number of characters written into buf not 461 * including the trailing '\0'. If len is == 0 the function returns 0.. 462 */ 463 int sockaddr_to_str(const struct sockaddr *addr, char *buf, size_t len) 464 { 465 466 switch (addr->sa_family) { 467 case AF_IB: 468 return scnprintf(buf, len, "gid:%pI6", 469 &((struct sockaddr_ib *)addr)->sib_addr.sib_raw); 470 case AF_INET: 471 return scnprintf(buf, len, "ip:%pI4", 472 &((struct sockaddr_in *)addr)->sin_addr); 473 case AF_INET6: 474 return scnprintf(buf, len, "ip:%pI6c", 475 &((struct sockaddr_in6 *)addr)->sin6_addr); 476 } 477 return scnprintf(buf, len, "<invalid address family>"); 478 } 479 EXPORT_SYMBOL(sockaddr_to_str); 480 481 /** 482 * rtrs_addr_to_str() - convert rtrs_addr to a string "src@dst" 483 * @addr: the rtrs_addr structure to be converted 484 * @buf: string containing source and destination addr of a path 485 * separated by '@' I.e. "ip:1.1.1.1@ip:1.1.1.2" 486 * "ip:1.1.1.1@ip:1.1.1.2". 487 * @len: string length 488 * 489 * The return value is the number of characters written into buf not 490 * including the trailing '\0'. 491 */ 492 int rtrs_addr_to_str(const struct rtrs_addr *addr, char *buf, size_t len) 493 { 494 int cnt; 495 496 cnt = sockaddr_to_str((struct sockaddr *)addr->src, 497 buf, len); 498 cnt += scnprintf(buf + cnt, len - cnt, "@"); 499 sockaddr_to_str((struct sockaddr *)addr->dst, 500 buf + cnt, len - cnt); 501 return cnt; 502 } 503 EXPORT_SYMBOL(rtrs_addr_to_str); 504 505 /** 506 * rtrs_addr_to_sockaddr() - convert path string "src,dst" or "src@dst" 507 * to sockaddreses 508 * @str: string containing source and destination addr of a path 509 * separated by ',' or '@' I.e. "ip:1.1.1.1,ip:1.1.1.2" or 510 * "ip:1.1.1.1@ip:1.1.1.2". If str contains only one address it's 511 * considered to be destination. 512 * @len: string length 513 * @port: Destination port number. 514 * @addr: will be set to the source/destination address or to NULL 515 * if str doesn't contain any source address. 516 * 517 * Returns zero if conversion successful. Non-zero otherwise. 518 */ 519 int rtrs_addr_to_sockaddr(const char *str, size_t len, u16 port, 520 struct rtrs_addr *addr) 521 { 522 const char *d; 523 524 d = strchr(str, ','); 525 if (!d) 526 d = strchr(str, '@'); 527 if (d) { 528 if (rtrs_str_to_sockaddr(str, d - str, 0, addr->src)) 529 return -EINVAL; 530 d += 1; 531 len -= d - str; 532 str = d; 533 534 } else { 535 addr->src = NULL; 536 } 537 return rtrs_str_to_sockaddr(str, len, port, addr->dst); 538 } 539 EXPORT_SYMBOL(rtrs_addr_to_sockaddr); 540 541 void rtrs_rdma_dev_pd_init(enum ib_pd_flags pd_flags, 542 struct rtrs_rdma_dev_pd *pool) 543 { 544 WARN_ON(pool->ops && (!pool->ops->alloc ^ !pool->ops->free)); 545 INIT_LIST_HEAD(&pool->list); 546 mutex_init(&pool->mutex); 547 pool->pd_flags = pd_flags; 548 } 549 EXPORT_SYMBOL(rtrs_rdma_dev_pd_init); 550 551 void rtrs_rdma_dev_pd_deinit(struct rtrs_rdma_dev_pd *pool) 552 { 553 mutex_destroy(&pool->mutex); 554 WARN_ON(!list_empty(&pool->list)); 555 } 556 EXPORT_SYMBOL(rtrs_rdma_dev_pd_deinit); 557 558 static void dev_free(struct kref *ref) 559 { 560 struct rtrs_rdma_dev_pd *pool; 561 struct rtrs_ib_dev *dev; 562 563 dev = container_of(ref, typeof(*dev), ref); 564 pool = dev->pool; 565 566 mutex_lock(&pool->mutex); 567 list_del(&dev->entry); 568 mutex_unlock(&pool->mutex); 569 570 if (pool->ops && pool->ops->deinit) 571 pool->ops->deinit(dev); 572 573 ib_dealloc_pd(dev->ib_pd); 574 575 if (pool->ops && pool->ops->free) 576 pool->ops->free(dev); 577 else 578 kfree(dev); 579 } 580 581 int rtrs_ib_dev_put(struct rtrs_ib_dev *dev) 582 { 583 return kref_put(&dev->ref, dev_free); 584 } 585 EXPORT_SYMBOL(rtrs_ib_dev_put); 586 587 static int rtrs_ib_dev_get(struct rtrs_ib_dev *dev) 588 { 589 return kref_get_unless_zero(&dev->ref); 590 } 591 592 struct rtrs_ib_dev * 593 rtrs_ib_dev_find_or_add(struct ib_device *ib_dev, 594 struct rtrs_rdma_dev_pd *pool) 595 { 596 struct rtrs_ib_dev *dev; 597 598 mutex_lock(&pool->mutex); 599 list_for_each_entry(dev, &pool->list, entry) { 600 if (dev->ib_dev->node_guid == ib_dev->node_guid && 601 rtrs_ib_dev_get(dev)) 602 goto out_unlock; 603 } 604 mutex_unlock(&pool->mutex); 605 if (pool->ops && pool->ops->alloc) 606 dev = pool->ops->alloc(); 607 else 608 dev = kzalloc(sizeof(*dev), GFP_KERNEL); 609 if (IS_ERR_OR_NULL(dev)) 610 goto out_err; 611 612 kref_init(&dev->ref); 613 dev->pool = pool; 614 dev->ib_dev = ib_dev; 615 dev->ib_pd = ib_alloc_pd(ib_dev, pool->pd_flags); 616 if (IS_ERR(dev->ib_pd)) 617 goto out_free_dev; 618 619 if (pool->ops && pool->ops->init && pool->ops->init(dev)) 620 goto out_free_pd; 621 622 mutex_lock(&pool->mutex); 623 list_add(&dev->entry, &pool->list); 624 out_unlock: 625 mutex_unlock(&pool->mutex); 626 return dev; 627 628 out_free_pd: 629 ib_dealloc_pd(dev->ib_pd); 630 out_free_dev: 631 if (pool->ops && pool->ops->free) 632 pool->ops->free(dev); 633 else 634 kfree(dev); 635 out_err: 636 return NULL; 637 } 638 EXPORT_SYMBOL(rtrs_ib_dev_find_or_add); 639