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