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 10 #undef pr_fmt 11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt 12 13 #include <linux/module.h> 14 #include <linux/rculist.h> 15 #include <linux/random.h> 16 17 #include "rtrs-clt.h" 18 #include "rtrs-log.h" 19 #include "rtrs-clt-trace.h" 20 21 #define RTRS_CONNECT_TIMEOUT_MS 30000 22 /* 23 * Wait a bit before trying to reconnect after a failure 24 * in order to give server time to finish clean up which 25 * leads to "false positives" failed reconnect attempts 26 */ 27 #define RTRS_RECONNECT_BACKOFF 1000 28 /* 29 * Wait for additional random time between 0 and 8 seconds 30 * before starting to reconnect to avoid clients reconnecting 31 * all at once in case of a major network outage 32 */ 33 #define RTRS_RECONNECT_SEED 8 34 35 #define FIRST_CONN 0x01 36 /* limit to 128 * 4k = 512k max IO */ 37 #define RTRS_MAX_SEGMENTS 128 38 39 MODULE_DESCRIPTION("RDMA Transport Client"); 40 MODULE_LICENSE("GPL"); 41 42 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops; 43 static struct rtrs_rdma_dev_pd dev_pd = { 44 .ops = &dev_pd_ops 45 }; 46 47 static struct workqueue_struct *rtrs_wq; 48 static const struct class rtrs_clt_dev_class = { 49 .name = "rtrs-client", 50 }; 51 52 static inline bool rtrs_clt_is_connected(const struct rtrs_clt_sess *clt) 53 { 54 struct rtrs_clt_path *clt_path; 55 bool connected = false; 56 57 rcu_read_lock(); 58 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) 59 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED) { 60 connected = true; 61 break; 62 } 63 rcu_read_unlock(); 64 65 return connected; 66 } 67 68 static struct rtrs_permit * 69 __rtrs_get_permit(struct rtrs_clt_sess *clt, enum rtrs_clt_con_type con_type) 70 { 71 size_t max_depth = clt->queue_depth; 72 struct rtrs_permit *permit; 73 int bit; 74 75 /* 76 * Adapted from null_blk get_tag(). Callers from different cpus may 77 * grab the same bit, since find_first_zero_bit is not atomic. 78 * But then the test_and_set_bit_lock will fail for all the 79 * callers but one, so that they will loop again. 80 * This way an explicit spinlock is not required. 81 */ 82 do { 83 bit = find_first_zero_bit(clt->permits_map, max_depth); 84 if (bit >= max_depth) 85 return NULL; 86 } while (test_and_set_bit_lock(bit, clt->permits_map)); 87 88 permit = get_permit(clt, bit); 89 WARN_ON(permit->mem_id != bit); 90 permit->cpu_id = raw_smp_processor_id(); 91 permit->con_type = con_type; 92 93 return permit; 94 } 95 96 static inline void __rtrs_put_permit(struct rtrs_clt_sess *clt, 97 struct rtrs_permit *permit) 98 { 99 clear_bit_unlock(permit->mem_id, clt->permits_map); 100 } 101 102 /** 103 * rtrs_clt_get_permit() - allocates permit for future RDMA operation 104 * @clt: Current session 105 * @con_type: Type of connection to use with the permit 106 * @can_wait: Wait type 107 * 108 * Description: 109 * Allocates permit for the following RDMA operation. Permit is used 110 * to preallocate all resources and to propagate memory pressure 111 * up earlier. 112 * 113 * Context: 114 * Can sleep if @wait == RTRS_PERMIT_WAIT 115 */ 116 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt_sess *clt, 117 enum rtrs_clt_con_type con_type, 118 enum wait_type can_wait) 119 { 120 struct rtrs_permit *permit; 121 DEFINE_WAIT(wait); 122 123 permit = __rtrs_get_permit(clt, con_type); 124 if (permit || !can_wait) 125 return permit; 126 127 do { 128 prepare_to_wait(&clt->permits_wait, &wait, 129 TASK_UNINTERRUPTIBLE); 130 permit = __rtrs_get_permit(clt, con_type); 131 if (permit) 132 break; 133 134 io_schedule(); 135 } while (1); 136 137 finish_wait(&clt->permits_wait, &wait); 138 139 return permit; 140 } 141 EXPORT_SYMBOL(rtrs_clt_get_permit); 142 143 /** 144 * rtrs_clt_put_permit() - puts allocated permit 145 * @clt: Current session 146 * @permit: Permit to be freed 147 * 148 * Context: 149 * Does not matter 150 */ 151 void rtrs_clt_put_permit(struct rtrs_clt_sess *clt, 152 struct rtrs_permit *permit) 153 { 154 if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map))) 155 return; 156 157 __rtrs_put_permit(clt, permit); 158 159 /* 160 * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list 161 * before calling schedule(). So if rtrs_clt_get_permit() is sleeping 162 * it must have added itself to &clt->permits_wait before 163 * __rtrs_put_permit() finished. 164 * Hence it is safe to guard wake_up() with a waitqueue_active() test. 165 */ 166 if (waitqueue_active(&clt->permits_wait)) 167 wake_up(&clt->permits_wait); 168 } 169 EXPORT_SYMBOL(rtrs_clt_put_permit); 170 171 /** 172 * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit 173 * @clt_path: client path pointer 174 * @permit: permit for the allocation of the RDMA buffer 175 * Note: 176 * IO connection starts from 1. 177 * 0 connection is for user messages. 178 */ 179 static 180 struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_path *clt_path, 181 struct rtrs_permit *permit) 182 { 183 int id = 0; 184 185 if (permit->con_type == RTRS_IO_CON) 186 id = (permit->cpu_id % (clt_path->s.irq_con_num - 1)) + 1; 187 188 return to_clt_con(clt_path->s.con[id]); 189 } 190 191 /** 192 * rtrs_clt_change_state() - change the session state through session state 193 * machine. 194 * 195 * @clt_path: client path to change the state of. 196 * @new_state: state to change to. 197 * 198 * returns true if sess's state is changed to new state, otherwise return false. 199 * 200 * Locks: 201 * state_wq lock must be hold. 202 */ 203 static bool rtrs_clt_change_state(struct rtrs_clt_path *clt_path, 204 enum rtrs_clt_state new_state) 205 { 206 enum rtrs_clt_state old_state; 207 bool changed = false; 208 209 lockdep_assert_held(&clt_path->state_wq.lock); 210 211 old_state = clt_path->state; 212 switch (new_state) { 213 case RTRS_CLT_CONNECTING: 214 switch (old_state) { 215 case RTRS_CLT_RECONNECTING: 216 changed = true; 217 fallthrough; 218 default: 219 break; 220 } 221 break; 222 case RTRS_CLT_RECONNECTING: 223 switch (old_state) { 224 case RTRS_CLT_CONNECTED: 225 case RTRS_CLT_CONNECTING_ERR: 226 case RTRS_CLT_CLOSED: 227 changed = true; 228 fallthrough; 229 default: 230 break; 231 } 232 break; 233 case RTRS_CLT_CONNECTED: 234 switch (old_state) { 235 case RTRS_CLT_CONNECTING: 236 changed = true; 237 fallthrough; 238 default: 239 break; 240 } 241 break; 242 case RTRS_CLT_CONNECTING_ERR: 243 switch (old_state) { 244 case RTRS_CLT_CONNECTING: 245 changed = true; 246 fallthrough; 247 default: 248 break; 249 } 250 break; 251 case RTRS_CLT_CLOSING: 252 switch (old_state) { 253 case RTRS_CLT_CONNECTING: 254 case RTRS_CLT_CONNECTING_ERR: 255 case RTRS_CLT_RECONNECTING: 256 case RTRS_CLT_CONNECTED: 257 changed = true; 258 fallthrough; 259 default: 260 break; 261 } 262 break; 263 case RTRS_CLT_CLOSED: 264 switch (old_state) { 265 case RTRS_CLT_CLOSING: 266 changed = true; 267 fallthrough; 268 default: 269 break; 270 } 271 break; 272 case RTRS_CLT_DEAD: 273 switch (old_state) { 274 case RTRS_CLT_CLOSED: 275 changed = true; 276 fallthrough; 277 default: 278 break; 279 } 280 break; 281 default: 282 break; 283 } 284 if (changed) { 285 clt_path->state = new_state; 286 wake_up_locked(&clt_path->state_wq); 287 } 288 289 return changed; 290 } 291 292 static bool rtrs_clt_change_state_from_to(struct rtrs_clt_path *clt_path, 293 enum rtrs_clt_state old_state, 294 enum rtrs_clt_state new_state) 295 { 296 bool changed = false; 297 298 spin_lock_irq(&clt_path->state_wq.lock); 299 if (clt_path->state == old_state) 300 changed = rtrs_clt_change_state(clt_path, new_state); 301 spin_unlock_irq(&clt_path->state_wq.lock); 302 303 return changed; 304 } 305 306 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path); 307 static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con) 308 { 309 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 310 311 trace_rtrs_rdma_error_recovery(clt_path); 312 313 if (rtrs_clt_change_state_from_to(clt_path, 314 RTRS_CLT_CONNECTED, 315 RTRS_CLT_RECONNECTING)) { 316 queue_work(rtrs_wq, &clt_path->err_recovery_work); 317 } else { 318 /* 319 * Error can happen just on establishing new connection, 320 * so notify waiter with error state, waiter is responsible 321 * for cleaning the rest and reconnect if needed. 322 */ 323 rtrs_clt_change_state_from_to(clt_path, 324 RTRS_CLT_CONNECTING, 325 RTRS_CLT_CONNECTING_ERR); 326 } 327 } 328 329 static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc) 330 { 331 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context); 332 333 if (wc->status != IB_WC_SUCCESS) { 334 rtrs_err_rl(con->c.path, "Failed IB_WR_REG_MR: %s\n", 335 ib_wc_status_msg(wc->status)); 336 rtrs_rdma_error_recovery(con); 337 } 338 } 339 340 static struct ib_cqe fast_reg_cqe = { 341 .done = rtrs_clt_fast_reg_done 342 }; 343 344 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno, 345 bool notify, bool can_wait); 346 347 static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc) 348 { 349 struct rtrs_clt_io_req *req = 350 container_of(wc->wr_cqe, typeof(*req), inv_cqe); 351 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context); 352 353 if (wc->status != IB_WC_SUCCESS) { 354 rtrs_err_rl(con->c.path, "Failed IB_WR_LOCAL_INV: %s\n", 355 ib_wc_status_msg(wc->status)); 356 rtrs_rdma_error_recovery(con); 357 } 358 req->mr->need_inval = false; 359 if (req->need_inv_comp) 360 complete(&req->inv_comp); 361 else 362 /* Complete request from INV callback */ 363 complete_rdma_req(req, req->inv_errno, true, false); 364 } 365 366 static int rtrs_inv_rkey(struct rtrs_clt_io_req *req) 367 { 368 struct rtrs_clt_con *con = req->con; 369 struct ib_send_wr wr = { 370 .opcode = IB_WR_LOCAL_INV, 371 .wr_cqe = &req->inv_cqe, 372 .send_flags = IB_SEND_SIGNALED, 373 .ex.invalidate_rkey = req->mr->rkey, 374 }; 375 req->inv_cqe.done = rtrs_clt_inv_rkey_done; 376 377 return ib_post_send(con->c.qp, &wr, NULL); 378 } 379 380 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno, 381 bool notify, bool can_wait) 382 { 383 struct rtrs_clt_con *con = req->con; 384 struct rtrs_clt_path *clt_path; 385 int err; 386 387 if (!req->in_use) 388 return; 389 if (WARN_ON(!req->con)) 390 return; 391 clt_path = to_clt_path(con->c.path); 392 393 if (req->sg_cnt) { 394 if (req->mr->need_inval) { 395 /* 396 * We are here to invalidate read/write requests 397 * ourselves. In normal scenario server should 398 * send INV for all read requests, we do local 399 * invalidate for write requests ourselves, but 400 * we are here, thus three things could happen: 401 * 402 * 1. this is failover, when errno != 0 403 * and can_wait == 1, 404 * 405 * 2. something totally bad happened and 406 * server forgot to send INV, so we 407 * should do that ourselves. 408 * 409 * 3. write request finishes, we need to do local 410 * invalidate 411 */ 412 413 if (can_wait) { 414 req->need_inv_comp = true; 415 } else { 416 /* This should be IO path, so always notify */ 417 WARN_ON(!notify); 418 /* Save errno for INV callback */ 419 req->inv_errno = errno; 420 } 421 422 refcount_inc(&req->ref); 423 err = rtrs_inv_rkey(req); 424 if (err) { 425 rtrs_err_rl(con->c.path, "Send INV WR key=%#x: %pe\n", 426 req->mr->rkey, ERR_PTR(err)); 427 } else if (can_wait) { 428 wait_for_completion(&req->inv_comp); 429 } 430 if (!refcount_dec_and_test(&req->ref)) 431 return; 432 } 433 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist, 434 req->sg_cnt, req->dir); 435 } 436 if (!refcount_dec_and_test(&req->ref)) 437 return; 438 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT) 439 atomic_dec(&clt_path->stats->inflight); 440 441 req->in_use = false; 442 req->con = NULL; 443 444 if (errno) { 445 rtrs_err_rl(con->c.path, 446 "IO %s request failed: error=%pe path=%s [%s:%u] notify=%d\n", 447 req->dir == DMA_TO_DEVICE ? "write" : "read", ERR_PTR(errno), 448 kobject_name(&clt_path->kobj), clt_path->hca_name, 449 clt_path->hca_port, notify); 450 } 451 452 if (notify) 453 req->conf(req->priv, errno); 454 } 455 456 static int rtrs_post_send_rdma(struct rtrs_clt_con *con, 457 struct rtrs_clt_io_req *req, 458 struct rtrs_rbuf *rbuf, u32 off, 459 u32 imm, struct ib_send_wr *wr) 460 { 461 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 462 enum ib_send_flags flags; 463 struct ib_sge sge; 464 465 if (!req->sg_size) { 466 rtrs_wrn(con->c.path, 467 "Doing RDMA Write failed, no data supplied\n"); 468 return -EINVAL; 469 } 470 471 /* user data and user message in the first list element */ 472 sge.addr = req->iu->dma_addr; 473 sge.length = req->sg_size; 474 sge.lkey = clt_path->s.dev->ib_pd->local_dma_lkey; 475 476 /* 477 * From time to time we have to post signalled sends, 478 * or send queue will fill up and only QP reset can help. 479 */ 480 flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ? 481 0 : IB_SEND_SIGNALED; 482 483 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, 484 req->iu->dma_addr, 485 req->sg_size, DMA_TO_DEVICE); 486 487 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1, 488 rbuf->rkey, rbuf->addr + off, 489 imm, flags, wr, NULL); 490 } 491 492 static void process_io_rsp(struct rtrs_clt_path *clt_path, u32 msg_id, 493 s16 errno, bool w_inval) 494 { 495 struct rtrs_clt_io_req *req; 496 497 if (WARN_ON(msg_id >= clt_path->queue_depth)) 498 return; 499 500 req = &clt_path->reqs[msg_id]; 501 /* Drop need_inv if server responded with send with invalidation */ 502 req->mr->need_inval &= !w_inval; 503 complete_rdma_req(req, errno, true, false); 504 } 505 506 static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc) 507 { 508 struct rtrs_iu *iu; 509 int err; 510 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 511 512 WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0); 513 iu = container_of(wc->wr_cqe, struct rtrs_iu, 514 cqe); 515 err = rtrs_iu_post_recv(&con->c, iu); 516 if (err) { 517 rtrs_err(con->c.path, "post iu failed %pe\n", ERR_PTR(err)); 518 rtrs_rdma_error_recovery(con); 519 } 520 } 521 522 static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc) 523 { 524 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 525 struct rtrs_msg_rkey_rsp *msg; 526 u32 imm_type, imm_payload; 527 bool w_inval = false; 528 struct rtrs_iu *iu; 529 u32 buf_id; 530 int err; 531 532 WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0); 533 534 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe); 535 536 if (wc->byte_len < sizeof(*msg)) { 537 rtrs_err(con->c.path, "rkey response is malformed: size %d\n", 538 wc->byte_len); 539 goto out; 540 } 541 ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr, 542 iu->size, DMA_FROM_DEVICE); 543 msg = iu->buf; 544 if (le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP) { 545 rtrs_err(clt_path->clt, 546 "rkey response is malformed: type %d\n", 547 le16_to_cpu(msg->type)); 548 goto out; 549 } 550 buf_id = le16_to_cpu(msg->buf_id); 551 if (WARN_ON(buf_id >= clt_path->queue_depth)) 552 goto out; 553 554 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload); 555 if (imm_type == RTRS_IO_RSP_IMM || 556 imm_type == RTRS_IO_RSP_W_INV_IMM) { 557 u32 msg_id; 558 559 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM); 560 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err); 561 562 if (WARN_ON(buf_id != msg_id)) 563 goto out; 564 clt_path->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey); 565 process_io_rsp(clt_path, msg_id, err, w_inval); 566 } 567 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, iu->dma_addr, 568 iu->size, DMA_FROM_DEVICE); 569 return rtrs_clt_recv_done(con, wc); 570 out: 571 rtrs_rdma_error_recovery(con); 572 } 573 574 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc); 575 576 static struct ib_cqe io_comp_cqe = { 577 .done = rtrs_clt_rdma_done 578 }; 579 580 /* 581 * Post x2 empty WRs: first is for this RDMA with IMM, 582 * second is for RECV with INV, which happened earlier. 583 */ 584 static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe) 585 { 586 struct ib_recv_wr wr_arr[2], *wr; 587 int i; 588 589 memset(wr_arr, 0, sizeof(wr_arr)); 590 for (i = 0; i < ARRAY_SIZE(wr_arr); i++) { 591 wr = &wr_arr[i]; 592 wr->wr_cqe = cqe; 593 if (i) 594 /* Chain backwards */ 595 wr->next = &wr_arr[i - 1]; 596 } 597 598 return ib_post_recv(con->qp, wr, NULL); 599 } 600 601 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc) 602 { 603 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context); 604 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 605 u32 imm_type, imm_payload; 606 bool w_inval = false; 607 int err; 608 609 if (wc->status != IB_WC_SUCCESS) { 610 if (wc->status != IB_WC_WR_FLUSH_ERR) { 611 rtrs_err(clt_path->clt, "RDMA failed: %s\n", 612 ib_wc_status_msg(wc->status)); 613 rtrs_rdma_error_recovery(con); 614 } 615 return; 616 } 617 rtrs_clt_update_wc_stats(con); 618 619 switch (wc->opcode) { 620 case IB_WC_RECV_RDMA_WITH_IMM: 621 /* 622 * post_recv() RDMA write completions of IO reqs (read/write) 623 * and hb 624 */ 625 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done)) 626 return; 627 clt_path->s.hb_missed_cnt = 0; 628 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), 629 &imm_type, &imm_payload); 630 if (imm_type == RTRS_IO_RSP_IMM || 631 imm_type == RTRS_IO_RSP_W_INV_IMM) { 632 u32 msg_id; 633 634 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM); 635 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err); 636 637 process_io_rsp(clt_path, msg_id, err, w_inval); 638 } else if (imm_type == RTRS_HB_MSG_IMM) { 639 WARN_ON(con->c.cid); 640 rtrs_send_hb_ack(&clt_path->s); 641 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) 642 return rtrs_clt_recv_done(con, wc); 643 } else if (imm_type == RTRS_HB_ACK_IMM) { 644 WARN_ON(con->c.cid); 645 clt_path->s.hb_cur_latency = 646 ktime_sub(ktime_get(), clt_path->s.hb_last_sent); 647 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) 648 return rtrs_clt_recv_done(con, wc); 649 } else { 650 rtrs_wrn(con->c.path, "Unknown IMM type %u\n", 651 imm_type); 652 } 653 if (w_inval) 654 /* 655 * Post x2 empty WRs: first is for this RDMA with IMM, 656 * second is for RECV with INV, which happened earlier. 657 */ 658 err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe); 659 else 660 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe); 661 if (err) { 662 rtrs_err(con->c.path, "rtrs_post_recv_empty(): %pe\n", 663 ERR_PTR(err)); 664 rtrs_rdma_error_recovery(con); 665 } 666 break; 667 case IB_WC_RECV: 668 /* 669 * Key invalidations from server side 670 */ 671 clt_path->s.hb_missed_cnt = 0; 672 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE || 673 wc->wc_flags & IB_WC_WITH_IMM)); 674 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done); 675 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) { 676 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) 677 return rtrs_clt_recv_done(con, wc); 678 679 return rtrs_clt_rkey_rsp_done(con, wc); 680 } 681 break; 682 case IB_WC_RDMA_WRITE: 683 /* 684 * post_send() RDMA write completions of IO reqs (read/write) 685 * and hb. 686 */ 687 break; 688 689 default: 690 rtrs_wrn(clt_path->clt, "Unexpected WC type: %d\n", wc->opcode); 691 return; 692 } 693 } 694 695 static int post_recv_io(struct rtrs_clt_con *con, size_t q_size) 696 { 697 int err, i; 698 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 699 700 for (i = 0; i < q_size; i++) { 701 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) { 702 struct rtrs_iu *iu = &con->rsp_ius[i]; 703 704 err = rtrs_iu_post_recv(&con->c, iu); 705 } else { 706 err = rtrs_post_recv_empty(&con->c, &io_comp_cqe); 707 } 708 if (err) 709 return err; 710 } 711 712 return 0; 713 } 714 715 static int post_recv_path(struct rtrs_clt_path *clt_path) 716 { 717 size_t q_size = 0; 718 int err, cid; 719 720 for (cid = 0; cid < clt_path->s.con_num; cid++) { 721 if (cid == 0) 722 q_size = SERVICE_CON_QUEUE_DEPTH; 723 else 724 q_size = clt_path->queue_depth; 725 726 /* 727 * x2 for RDMA read responses + FR key invalidations, 728 * RDMA writes do not require any FR registrations. 729 */ 730 q_size *= 2; 731 732 err = post_recv_io(to_clt_con(clt_path->s.con[cid]), q_size); 733 if (err) { 734 rtrs_err(clt_path->clt, "post_recv_io(), err: %pe\n", 735 ERR_PTR(err)); 736 return err; 737 } 738 } 739 740 return 0; 741 } 742 743 struct path_it { 744 int i; 745 struct list_head skip_list; 746 struct rtrs_clt_sess *clt; 747 struct rtrs_clt_path *(*next_path)(struct path_it *it); 748 }; 749 750 /* 751 * rtrs_clt_get_next_path_or_null - get clt path from the list or return NULL 752 * @head: the head for the list. 753 * @clt_path: The element to take the next clt_path from. 754 * 755 * Next clt path returned in round-robin fashion, i.e. head will be skipped, 756 * but if list is observed as empty, NULL will be returned. 757 * 758 * This function may safely run concurrently with the _rcu list-mutation 759 * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock(). 760 */ 761 static inline struct rtrs_clt_path * 762 rtrs_clt_get_next_path_or_null(struct list_head *head, struct rtrs_clt_path *clt_path) 763 { 764 return list_next_or_null_rcu(head, &clt_path->s.entry, typeof(*clt_path), s.entry) ?: 765 list_next_or_null_rcu(head, 766 READ_ONCE((&clt_path->s.entry)->next), 767 typeof(*clt_path), s.entry); 768 } 769 770 /** 771 * get_next_path_rr() - Returns path in round-robin fashion. 772 * @it: the path pointer 773 * 774 * Related to @MP_POLICY_RR 775 * 776 * Locks: 777 * rcu_read_lock() must be held. 778 */ 779 static struct rtrs_clt_path *get_next_path_rr(struct path_it *it) 780 { 781 struct rtrs_clt_path __rcu **ppcpu_path; 782 struct rtrs_clt_path *path; 783 struct rtrs_clt_sess *clt; 784 785 /* 786 * Assert that rcu lock must be held 787 */ 788 RCU_LOCKDEP_WARN(!rcu_read_lock_held(), "no rcu read lock held"); 789 790 clt = it->clt; 791 792 /* 793 * Here we use two RCU objects: @paths_list and @pcpu_path 794 * pointer. See rtrs_clt_remove_path_from_arr() for details 795 * how that is handled. 796 */ 797 798 ppcpu_path = this_cpu_ptr(clt->pcpu_path); 799 path = rcu_dereference(*ppcpu_path); 800 if (!path) 801 path = list_first_or_null_rcu(&clt->paths_list, 802 typeof(*path), s.entry); 803 else 804 path = rtrs_clt_get_next_path_or_null(&clt->paths_list, path); 805 806 rcu_assign_pointer(*ppcpu_path, path); 807 808 return path; 809 } 810 811 /** 812 * get_next_path_min_inflight() - Returns path with minimal inflight count. 813 * @it: the path pointer 814 * 815 * Related to @MP_POLICY_MIN_INFLIGHT 816 * 817 * Locks: 818 * rcu_read_lock() must be hold. 819 */ 820 static struct rtrs_clt_path *get_next_path_min_inflight(struct path_it *it) 821 { 822 struct rtrs_clt_path *min_path = NULL; 823 struct rtrs_clt_sess *clt = it->clt; 824 struct rtrs_clt_path *clt_path; 825 int min_inflight = INT_MAX; 826 int inflight; 827 828 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) { 829 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) 830 continue; 831 832 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry))) 833 continue; 834 835 inflight = atomic_read(&clt_path->stats->inflight); 836 837 if (inflight < min_inflight) { 838 min_inflight = inflight; 839 min_path = clt_path; 840 } 841 } 842 843 /* 844 * add the path to the skip list, so that next time we can get 845 * a different one 846 */ 847 if (min_path) 848 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list); 849 850 return min_path; 851 } 852 853 /** 854 * get_next_path_min_latency() - Returns path with minimal latency. 855 * @it: the path pointer 856 * 857 * Return: a path with the lowest latency or NULL if all paths are tried 858 * 859 * Locks: 860 * rcu_read_lock() must be hold. 861 * 862 * Related to @MP_POLICY_MIN_LATENCY 863 * 864 * This DOES skip an already-tried path. 865 * There is a skip-list to skip a path if the path has tried but failed. 866 * It will try the minimum latency path and then the second minimum latency 867 * path and so on. Finally it will return NULL if all paths are tried. 868 * Therefore the caller MUST check the returned 869 * path is NULL and trigger the IO error. 870 */ 871 static struct rtrs_clt_path *get_next_path_min_latency(struct path_it *it) 872 { 873 struct rtrs_clt_path *min_path = NULL; 874 struct rtrs_clt_sess *clt = it->clt; 875 struct rtrs_clt_path *clt_path; 876 ktime_t min_latency = KTIME_MAX; 877 ktime_t latency; 878 879 list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) { 880 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) 881 continue; 882 883 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry))) 884 continue; 885 886 latency = clt_path->s.hb_cur_latency; 887 888 if (latency < min_latency) { 889 min_latency = latency; 890 min_path = clt_path; 891 } 892 } 893 894 /* 895 * add the path to the skip list, so that next time we can get 896 * a different one 897 */ 898 if (min_path) 899 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list); 900 901 return min_path; 902 } 903 904 static inline void path_it_init(struct path_it *it, struct rtrs_clt_sess *clt) 905 { 906 INIT_LIST_HEAD(&it->skip_list); 907 it->clt = clt; 908 it->i = 0; 909 910 if (clt->mp_policy == MP_POLICY_RR) 911 it->next_path = get_next_path_rr; 912 else if (clt->mp_policy == MP_POLICY_MIN_INFLIGHT) 913 it->next_path = get_next_path_min_inflight; 914 else 915 it->next_path = get_next_path_min_latency; 916 } 917 918 static inline void path_it_deinit(struct path_it *it) 919 { 920 struct list_head *skip, *tmp; 921 /* 922 * The skip_list is used only for the MIN_INFLIGHT and MIN_LATENCY policies. 923 * We need to remove paths from it, so that next IO can insert 924 * paths (->mp_skip_entry) into a skip_list again. 925 */ 926 list_for_each_safe(skip, tmp, &it->skip_list) 927 list_del_init(skip); 928 } 929 930 /** 931 * rtrs_clt_init_req() - Initialize an rtrs_clt_io_req holding information 932 * about an inflight IO. 933 * The user buffer holding user control message (not data) is copied into 934 * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will 935 * also hold the control message of rtrs. 936 * @req: an io request holding information about IO. 937 * @clt_path: client path 938 * @conf: conformation callback function to notify upper layer. 939 * @permit: permit for allocation of RDMA remote buffer 940 * @priv: private pointer 941 * @vec: kernel vector containing control message 942 * @usr_len: length of the user message 943 * @sg: scater list for IO data 944 * @sg_cnt: number of scater list entries 945 * @data_len: length of the IO data 946 * @dir: direction of the IO. 947 */ 948 static void rtrs_clt_init_req(struct rtrs_clt_io_req *req, 949 struct rtrs_clt_path *clt_path, 950 void (*conf)(void *priv, int errno), 951 struct rtrs_permit *permit, void *priv, 952 const struct kvec *vec, size_t usr_len, 953 struct scatterlist *sg, size_t sg_cnt, 954 size_t data_len, int dir) 955 { 956 struct iov_iter iter; 957 size_t len; 958 959 req->permit = permit; 960 req->in_use = true; 961 req->usr_len = usr_len; 962 req->data_len = data_len; 963 req->sglist = sg; 964 req->sg_cnt = sg_cnt; 965 req->priv = priv; 966 req->dir = dir; 967 req->con = rtrs_permit_to_clt_con(clt_path, permit); 968 req->conf = conf; 969 req->mr->need_inval = false; 970 req->need_inv_comp = false; 971 req->inv_errno = 0; 972 refcount_set(&req->ref, 1); 973 req->mp_policy = clt_path->clt->mp_policy; 974 975 iov_iter_kvec(&iter, ITER_SOURCE, vec, 1, usr_len); 976 len = _copy_from_iter(req->iu->buf, usr_len, &iter); 977 WARN_ON(len != usr_len); 978 979 reinit_completion(&req->inv_comp); 980 } 981 982 static struct rtrs_clt_io_req * 983 rtrs_clt_get_req(struct rtrs_clt_path *clt_path, 984 void (*conf)(void *priv, int errno), 985 struct rtrs_permit *permit, void *priv, 986 const struct kvec *vec, size_t usr_len, 987 struct scatterlist *sg, size_t sg_cnt, 988 size_t data_len, int dir) 989 { 990 struct rtrs_clt_io_req *req; 991 992 req = &clt_path->reqs[permit->mem_id]; 993 rtrs_clt_init_req(req, clt_path, conf, permit, priv, vec, usr_len, 994 sg, sg_cnt, data_len, dir); 995 return req; 996 } 997 998 static struct rtrs_clt_io_req * 999 rtrs_clt_get_copy_req(struct rtrs_clt_path *alive_path, 1000 struct rtrs_clt_io_req *fail_req) 1001 { 1002 struct rtrs_clt_io_req *req; 1003 struct kvec vec = { 1004 .iov_base = fail_req->iu->buf, 1005 .iov_len = fail_req->usr_len 1006 }; 1007 1008 req = &alive_path->reqs[fail_req->permit->mem_id]; 1009 rtrs_clt_init_req(req, alive_path, fail_req->conf, fail_req->permit, 1010 fail_req->priv, &vec, fail_req->usr_len, 1011 fail_req->sglist, fail_req->sg_cnt, 1012 fail_req->data_len, fail_req->dir); 1013 return req; 1014 } 1015 1016 static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con, 1017 struct rtrs_clt_io_req *req, 1018 struct rtrs_rbuf *rbuf, bool fr_en, 1019 u32 count, u32 size, u32 imm, 1020 struct ib_send_wr *wr, 1021 struct ib_send_wr *tail) 1022 { 1023 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 1024 struct ib_sge *sge = req->sge; 1025 enum ib_send_flags flags; 1026 struct scatterlist *sg; 1027 size_t num_sge; 1028 int i; 1029 struct ib_send_wr *ptail = NULL; 1030 1031 if (fr_en) { 1032 i = 0; 1033 sge[i].addr = req->mr->iova; 1034 sge[i].length = req->mr->length; 1035 sge[i].lkey = req->mr->lkey; 1036 i++; 1037 num_sge = 2; 1038 ptail = tail; 1039 } else { 1040 for_each_sg(req->sglist, sg, count, i) { 1041 sge[i].addr = sg_dma_address(sg); 1042 sge[i].length = sg_dma_len(sg); 1043 sge[i].lkey = clt_path->s.dev->ib_pd->local_dma_lkey; 1044 } 1045 num_sge = 1 + count; 1046 } 1047 sge[i].addr = req->iu->dma_addr; 1048 sge[i].length = size; 1049 sge[i].lkey = clt_path->s.dev->ib_pd->local_dma_lkey; 1050 1051 /* 1052 * From time to time we have to post signalled sends, 1053 * or send queue will fill up and only QP reset can help. 1054 */ 1055 flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ? 1056 0 : IB_SEND_SIGNALED; 1057 1058 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, 1059 req->iu->dma_addr, 1060 size, DMA_TO_DEVICE); 1061 1062 return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge, 1063 rbuf->rkey, rbuf->addr, imm, 1064 flags, wr, ptail); 1065 } 1066 1067 static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count) 1068 { 1069 int nr; 1070 1071 /* Align the MR to a 4K page size to match the block virt boundary */ 1072 nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K); 1073 if (nr != count) 1074 return nr < 0 ? nr : -EINVAL; 1075 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey)); 1076 1077 return nr; 1078 } 1079 1080 static int rtrs_clt_write_req(struct rtrs_clt_io_req *req) 1081 { 1082 struct rtrs_clt_con *con = req->con; 1083 struct rtrs_path *s = con->c.path; 1084 struct rtrs_clt_path *clt_path = to_clt_path(s); 1085 struct rtrs_msg_rdma_write *msg; 1086 1087 struct rtrs_rbuf *rbuf; 1088 int ret, count = 0; 1089 u32 imm, buf_id; 1090 struct ib_reg_wr rwr; 1091 struct ib_send_wr *wr = NULL; 1092 bool fr_en = false; 1093 1094 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len; 1095 1096 if (tsize > clt_path->chunk_size) { 1097 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n", 1098 tsize, clt_path->chunk_size); 1099 return -EMSGSIZE; 1100 } 1101 if (req->sg_cnt) { 1102 count = ib_dma_map_sg(clt_path->s.dev->ib_dev, req->sglist, 1103 req->sg_cnt, req->dir); 1104 if (!count) { 1105 rtrs_wrn(s, "Write request failed, map failed\n"); 1106 return -EINVAL; 1107 } 1108 } 1109 /* put rtrs msg after sg and user message */ 1110 msg = req->iu->buf + req->usr_len; 1111 msg->type = cpu_to_le16(RTRS_MSG_WRITE); 1112 msg->usr_len = cpu_to_le16(req->usr_len); 1113 1114 /* rtrs message on server side will be after user data and message */ 1115 imm = req->permit->mem_off + req->data_len + req->usr_len; 1116 imm = rtrs_to_io_req_imm(imm); 1117 buf_id = req->permit->mem_id; 1118 req->sg_size = tsize; 1119 rbuf = &clt_path->rbufs[buf_id]; 1120 1121 if (count) { 1122 ret = rtrs_map_sg_fr(req, count); 1123 if (ret < 0) { 1124 rtrs_err_rl(s, 1125 "Write request failed, failed to map fast reg. data, err: %pe\n", 1126 ERR_PTR(ret)); 1127 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist, 1128 req->sg_cnt, req->dir); 1129 return ret; 1130 } 1131 rwr = (struct ib_reg_wr) { 1132 .wr.opcode = IB_WR_REG_MR, 1133 .wr.wr_cqe = &fast_reg_cqe, 1134 .mr = req->mr, 1135 .key = req->mr->rkey, 1136 .access = (IB_ACCESS_LOCAL_WRITE), 1137 }; 1138 wr = &rwr.wr; 1139 fr_en = true; 1140 req->mr->need_inval = true; 1141 } 1142 /* 1143 * Update stats now, after request is successfully sent it is not 1144 * safe anymore to touch it. 1145 */ 1146 rtrs_clt_update_all_stats(req, WRITE); 1147 1148 ret = rtrs_post_rdma_write_sg(req->con, req, rbuf, fr_en, count, 1149 req->usr_len + sizeof(*msg), 1150 imm, wr, NULL); 1151 if (ret) { 1152 rtrs_err_rl(s, 1153 "Write request failed: error=%pe path=%s [%s:%u]\n", 1154 ERR_PTR(ret), kobject_name(&clt_path->kobj), 1155 clt_path->hca_name, clt_path->hca_port); 1156 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT) 1157 atomic_dec(&clt_path->stats->inflight); 1158 if (req->mr->need_inval) { 1159 req->mr->need_inval = false; 1160 refcount_dec(&req->ref); 1161 } 1162 if (req->sg_cnt) 1163 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist, 1164 req->sg_cnt, req->dir); 1165 } 1166 1167 return ret; 1168 } 1169 1170 static int rtrs_clt_read_req(struct rtrs_clt_io_req *req) 1171 { 1172 struct rtrs_clt_con *con = req->con; 1173 struct rtrs_path *s = con->c.path; 1174 struct rtrs_clt_path *clt_path = to_clt_path(s); 1175 struct rtrs_msg_rdma_read *msg; 1176 struct rtrs_ib_dev *dev = clt_path->s.dev; 1177 1178 struct ib_reg_wr rwr; 1179 struct ib_send_wr *wr = NULL; 1180 1181 int ret, count = 0; 1182 u32 imm, buf_id; 1183 1184 const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len; 1185 1186 if (tsize > clt_path->chunk_size) { 1187 rtrs_wrn(s, 1188 "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n", 1189 tsize, clt_path->chunk_size); 1190 return -EMSGSIZE; 1191 } 1192 1193 if (req->sg_cnt) { 1194 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt, 1195 req->dir); 1196 if (!count) { 1197 rtrs_wrn(s, 1198 "Read request failed, dma map failed\n"); 1199 return -EINVAL; 1200 } 1201 } 1202 /* put our message into req->buf after user message*/ 1203 msg = req->iu->buf + req->usr_len; 1204 msg->type = cpu_to_le16(RTRS_MSG_READ); 1205 msg->usr_len = cpu_to_le16(req->usr_len); 1206 1207 if (count) { 1208 ret = rtrs_map_sg_fr(req, count); 1209 if (ret < 0) { 1210 rtrs_err_rl(s, 1211 "Read request failed, failed to map fast reg. data, err: %pe\n", 1212 ERR_PTR(ret)); 1213 ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt, 1214 req->dir); 1215 return ret; 1216 } 1217 rwr = (struct ib_reg_wr) { 1218 .wr.opcode = IB_WR_REG_MR, 1219 .wr.wr_cqe = &fast_reg_cqe, 1220 .mr = req->mr, 1221 .key = req->mr->rkey, 1222 .access = (IB_ACCESS_LOCAL_WRITE | 1223 IB_ACCESS_REMOTE_WRITE), 1224 }; 1225 wr = &rwr.wr; 1226 1227 msg->sg_cnt = cpu_to_le16(1); 1228 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F); 1229 1230 msg->desc[0].addr = cpu_to_le64(req->mr->iova); 1231 msg->desc[0].key = cpu_to_le32(req->mr->rkey); 1232 msg->desc[0].len = cpu_to_le32(req->mr->length); 1233 1234 /* Further invalidation is required */ 1235 req->mr->need_inval = !!RTRS_MSG_NEED_INVAL_F; 1236 1237 } else { 1238 msg->sg_cnt = 0; 1239 msg->flags = 0; 1240 } 1241 /* 1242 * rtrs message will be after the space reserved for disk data and 1243 * user message 1244 */ 1245 imm = req->permit->mem_off + req->data_len + req->usr_len; 1246 imm = rtrs_to_io_req_imm(imm); 1247 buf_id = req->permit->mem_id; 1248 1249 req->sg_size = sizeof(*msg); 1250 req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc); 1251 req->sg_size += req->usr_len; 1252 1253 /* 1254 * Update stats now, after request is successfully sent it is not 1255 * safe anymore to touch it. 1256 */ 1257 rtrs_clt_update_all_stats(req, READ); 1258 1259 ret = rtrs_post_send_rdma(req->con, req, &clt_path->rbufs[buf_id], 1260 req->data_len, imm, wr); 1261 if (ret) { 1262 rtrs_err_rl(s, 1263 "Read request failed: error=%pe path=%s [%s:%u]\n", 1264 ERR_PTR(ret), kobject_name(&clt_path->kobj), 1265 clt_path->hca_name, clt_path->hca_port); 1266 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT) 1267 atomic_dec(&clt_path->stats->inflight); 1268 req->mr->need_inval = false; 1269 if (req->sg_cnt) 1270 ib_dma_unmap_sg(dev->ib_dev, req->sglist, 1271 req->sg_cnt, req->dir); 1272 } 1273 1274 return ret; 1275 } 1276 1277 /** 1278 * rtrs_clt_failover_req() - Try to find an active path for a failed request 1279 * @clt: clt context 1280 * @fail_req: a failed io request. 1281 */ 1282 static int rtrs_clt_failover_req(struct rtrs_clt_sess *clt, 1283 struct rtrs_clt_io_req *fail_req) 1284 { 1285 struct rtrs_clt_path *alive_path; 1286 struct rtrs_clt_io_req *req; 1287 int err = -ECONNABORTED; 1288 struct path_it it; 1289 1290 rcu_read_lock(); 1291 for (path_it_init(&it, clt); 1292 (alive_path = it.next_path(&it)) && it.i < it.clt->paths_num; 1293 it.i++) { 1294 if (READ_ONCE(alive_path->state) != RTRS_CLT_CONNECTED) 1295 continue; 1296 req = rtrs_clt_get_copy_req(alive_path, fail_req); 1297 if (req->dir == DMA_TO_DEVICE) 1298 err = rtrs_clt_write_req(req); 1299 else 1300 err = rtrs_clt_read_req(req); 1301 if (err) { 1302 req->in_use = false; 1303 continue; 1304 } 1305 /* Success path */ 1306 rtrs_clt_inc_failover_cnt(alive_path->stats); 1307 break; 1308 } 1309 path_it_deinit(&it); 1310 rcu_read_unlock(); 1311 1312 return err; 1313 } 1314 1315 static void fail_all_outstanding_reqs(struct rtrs_clt_path *clt_path) 1316 { 1317 struct rtrs_clt_sess *clt = clt_path->clt; 1318 struct rtrs_clt_io_req *req; 1319 int i, err; 1320 1321 if (!clt_path->reqs) 1322 return; 1323 for (i = 0; i < clt_path->queue_depth; ++i) { 1324 req = &clt_path->reqs[i]; 1325 if (!req->in_use) 1326 continue; 1327 1328 /* 1329 * Safely (without notification) complete failed request. 1330 * After completion this request is still useble and can 1331 * be failovered to another path. 1332 */ 1333 complete_rdma_req(req, -ECONNABORTED, false, true); 1334 1335 err = rtrs_clt_failover_req(clt, req); 1336 if (err) 1337 /* Failover failed, notify anyway */ 1338 req->conf(req->priv, err); 1339 } 1340 } 1341 1342 static void free_path_reqs(struct rtrs_clt_path *clt_path) 1343 { 1344 struct rtrs_clt_io_req *req; 1345 int i; 1346 1347 if (!clt_path->reqs) 1348 return; 1349 for (i = 0; i < clt_path->queue_depth; ++i) { 1350 req = &clt_path->reqs[i]; 1351 if (req->mr) 1352 ib_dereg_mr(req->mr); 1353 kfree(req->sge); 1354 rtrs_iu_free(req->iu, clt_path->s.dev->ib_dev, 1); 1355 } 1356 kfree(clt_path->reqs); 1357 clt_path->reqs = NULL; 1358 } 1359 1360 static int alloc_path_reqs(struct rtrs_clt_path *clt_path) 1361 { 1362 struct ib_device *ib_dev = clt_path->s.dev->ib_dev; 1363 struct rtrs_clt_io_req *req; 1364 enum ib_mr_type mr_type; 1365 int i, err = -ENOMEM; 1366 1367 clt_path->reqs = kcalloc(clt_path->queue_depth, 1368 sizeof(*clt_path->reqs), 1369 GFP_KERNEL); 1370 if (!clt_path->reqs) 1371 return -ENOMEM; 1372 1373 if (ib_dev->attrs.kernel_cap_flags & IBK_SG_GAPS_REG) 1374 mr_type = IB_MR_TYPE_SG_GAPS; 1375 else 1376 mr_type = IB_MR_TYPE_MEM_REG; 1377 1378 for (i = 0; i < clt_path->queue_depth; ++i) { 1379 req = &clt_path->reqs[i]; 1380 req->iu = rtrs_iu_alloc(1, clt_path->max_hdr_size, GFP_KERNEL, 1381 clt_path->s.dev->ib_dev, 1382 DMA_TO_DEVICE, 1383 rtrs_clt_rdma_done); 1384 if (!req->iu) 1385 goto out; 1386 1387 req->sge = kcalloc(2, sizeof(*req->sge), GFP_KERNEL); 1388 if (!req->sge) 1389 goto out; 1390 1391 req->mr = ib_alloc_mr(clt_path->s.dev->ib_pd, mr_type, 1392 clt_path->max_pages_per_mr); 1393 if (IS_ERR(req->mr)) { 1394 err = PTR_ERR(req->mr); 1395 pr_err("Failed to alloc clt_path->max_pages_per_mr %d: %pe\n", 1396 clt_path->max_pages_per_mr, req->mr); 1397 req->mr = NULL; 1398 goto out; 1399 } 1400 1401 init_completion(&req->inv_comp); 1402 } 1403 1404 return 0; 1405 1406 out: 1407 free_path_reqs(clt_path); 1408 1409 return err; 1410 } 1411 1412 static int alloc_permits(struct rtrs_clt_sess *clt) 1413 { 1414 unsigned int chunk_bits; 1415 int err, i; 1416 1417 clt->permits_map = bitmap_zalloc(clt->queue_depth, GFP_KERNEL); 1418 if (!clt->permits_map) { 1419 err = -ENOMEM; 1420 goto out_err; 1421 } 1422 clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL); 1423 if (!clt->permits) { 1424 err = -ENOMEM; 1425 goto err_map; 1426 } 1427 chunk_bits = ilog2(clt->queue_depth - 1) + 1; 1428 for (i = 0; i < clt->queue_depth; i++) { 1429 struct rtrs_permit *permit; 1430 1431 permit = get_permit(clt, i); 1432 permit->mem_id = i; 1433 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits); 1434 } 1435 1436 return 0; 1437 1438 err_map: 1439 bitmap_free(clt->permits_map); 1440 clt->permits_map = NULL; 1441 out_err: 1442 return err; 1443 } 1444 1445 static void free_permits(struct rtrs_clt_sess *clt) 1446 { 1447 if (clt->permits_map) 1448 wait_event(clt->permits_wait, 1449 bitmap_empty(clt->permits_map, clt->queue_depth)); 1450 1451 bitmap_free(clt->permits_map); 1452 clt->permits_map = NULL; 1453 kfree(clt->permits); 1454 clt->permits = NULL; 1455 } 1456 1457 static void query_fast_reg_mode(struct rtrs_clt_path *clt_path) 1458 { 1459 struct ib_device *ib_dev; 1460 u64 max_pages_per_mr; 1461 int mr_page_shift; 1462 1463 ib_dev = clt_path->s.dev->ib_dev; 1464 1465 /* 1466 * Use the smallest page size supported by the HCA, down to a 1467 * minimum of 4096 bytes. We're unlikely to build large sglists 1468 * out of smaller entries. 1469 */ 1470 mr_page_shift = max(12, ffs(ib_dev->attrs.page_size_cap) - 1); 1471 max_pages_per_mr = ib_dev->attrs.max_mr_size; 1472 do_div(max_pages_per_mr, (1ull << mr_page_shift)); 1473 max_pages_per_mr = min_not_zero((u32)max_pages_per_mr, U32_MAX); 1474 clt_path->max_pages_per_mr = 1475 min3(clt_path->max_pages_per_mr, (u32)max_pages_per_mr, 1476 ib_dev->attrs.max_fast_reg_page_list_len); 1477 clt_path->clt->max_segments = 1478 min(clt_path->max_pages_per_mr, clt_path->clt->max_segments); 1479 } 1480 1481 static bool rtrs_clt_change_state_get_old(struct rtrs_clt_path *clt_path, 1482 enum rtrs_clt_state new_state, 1483 enum rtrs_clt_state *old_state) 1484 { 1485 bool changed; 1486 1487 spin_lock_irq(&clt_path->state_wq.lock); 1488 if (old_state) 1489 *old_state = clt_path->state; 1490 changed = rtrs_clt_change_state(clt_path, new_state); 1491 spin_unlock_irq(&clt_path->state_wq.lock); 1492 1493 return changed; 1494 } 1495 1496 static void rtrs_clt_hb_err_handler(struct rtrs_con *c) 1497 { 1498 struct rtrs_clt_con *con = container_of(c, typeof(*con), c); 1499 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 1500 1501 rtrs_err(con->c.path, "HB err handler for path=%s\n", kobject_name(&clt_path->kobj)); 1502 rtrs_rdma_error_recovery(con); 1503 } 1504 1505 static void rtrs_clt_init_hb(struct rtrs_clt_path *clt_path) 1506 { 1507 rtrs_init_hb(&clt_path->s, &io_comp_cqe, 1508 RTRS_HB_INTERVAL_MS, 1509 RTRS_HB_MISSED_MAX, 1510 rtrs_clt_hb_err_handler, 1511 rtrs_wq); 1512 } 1513 1514 static void rtrs_clt_reconnect_work(struct work_struct *work); 1515 static void rtrs_clt_close_work(struct work_struct *work); 1516 1517 static void rtrs_clt_err_recovery_work(struct work_struct *work) 1518 { 1519 struct rtrs_clt_path *clt_path; 1520 struct rtrs_clt_sess *clt; 1521 int delay_ms; 1522 1523 clt_path = container_of(work, struct rtrs_clt_path, err_recovery_work); 1524 clt = clt_path->clt; 1525 delay_ms = clt->reconnect_delay_sec * 1000; 1526 rtrs_clt_stop_and_destroy_conns(clt_path); 1527 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork, 1528 msecs_to_jiffies(delay_ms + 1529 get_random_u32_below(RTRS_RECONNECT_SEED))); 1530 } 1531 1532 static struct rtrs_clt_path *alloc_path(struct rtrs_clt_sess *clt, 1533 const struct rtrs_addr *path, 1534 size_t con_num, u32 nr_poll_queues) 1535 { 1536 struct rtrs_clt_path *clt_path; 1537 int err = -ENOMEM; 1538 int cpu; 1539 size_t total_con; 1540 1541 clt_path = kzalloc(sizeof(*clt_path), GFP_KERNEL); 1542 if (!clt_path) 1543 goto err; 1544 1545 /* 1546 * irqmode and poll 1547 * +1: Extra connection for user messages 1548 */ 1549 total_con = con_num + nr_poll_queues + 1; 1550 clt_path->s.con = kcalloc(total_con, sizeof(*clt_path->s.con), 1551 GFP_KERNEL); 1552 if (!clt_path->s.con) 1553 goto err_free_path; 1554 1555 clt_path->s.con_num = total_con; 1556 clt_path->s.irq_con_num = con_num + 1; 1557 1558 clt_path->stats = kzalloc(sizeof(*clt_path->stats), GFP_KERNEL); 1559 if (!clt_path->stats) 1560 goto err_free_con; 1561 1562 mutex_init(&clt_path->init_mutex); 1563 uuid_gen(&clt_path->s.uuid); 1564 memcpy(&clt_path->s.dst_addr, path->dst, 1565 rdma_addr_size((struct sockaddr *)path->dst)); 1566 1567 /* 1568 * rdma_resolve_addr() passes src_addr to cma_bind_addr, which 1569 * checks the sa_family to be non-zero. If user passed src_addr=NULL 1570 * the sess->src_addr will contain only zeros, which is then fine. 1571 */ 1572 if (path->src) 1573 memcpy(&clt_path->s.src_addr, path->src, 1574 rdma_addr_size((struct sockaddr *)path->src)); 1575 strscpy(clt_path->s.sessname, clt->sessname, 1576 sizeof(clt_path->s.sessname)); 1577 clt_path->clt = clt; 1578 clt_path->max_pages_per_mr = RTRS_MAX_SEGMENTS; 1579 init_waitqueue_head(&clt_path->state_wq); 1580 clt_path->state = RTRS_CLT_CONNECTING; 1581 atomic_set(&clt_path->connected_cnt, 0); 1582 INIT_WORK(&clt_path->close_work, rtrs_clt_close_work); 1583 INIT_WORK(&clt_path->err_recovery_work, rtrs_clt_err_recovery_work); 1584 INIT_DELAYED_WORK(&clt_path->reconnect_dwork, rtrs_clt_reconnect_work); 1585 rtrs_clt_init_hb(clt_path); 1586 1587 clt_path->mp_skip_entry = alloc_percpu(typeof(*clt_path->mp_skip_entry)); 1588 if (!clt_path->mp_skip_entry) 1589 goto err_free_stats; 1590 1591 for_each_possible_cpu(cpu) 1592 INIT_LIST_HEAD(per_cpu_ptr(clt_path->mp_skip_entry, cpu)); 1593 1594 err = rtrs_clt_init_stats(clt_path->stats); 1595 if (err) 1596 goto err_free_percpu; 1597 1598 return clt_path; 1599 1600 err_free_percpu: 1601 free_percpu(clt_path->mp_skip_entry); 1602 err_free_stats: 1603 kfree(clt_path->stats); 1604 err_free_con: 1605 kfree(clt_path->s.con); 1606 err_free_path: 1607 kfree(clt_path); 1608 err: 1609 return ERR_PTR(err); 1610 } 1611 1612 void free_path(struct rtrs_clt_path *clt_path) 1613 { 1614 free_percpu(clt_path->mp_skip_entry); 1615 mutex_destroy(&clt_path->init_mutex); 1616 kfree(clt_path->s.con); 1617 kfree(clt_path->rbufs); 1618 kfree(clt_path); 1619 } 1620 1621 static int create_con(struct rtrs_clt_path *clt_path, unsigned int cid) 1622 { 1623 struct rtrs_clt_con *con; 1624 1625 con = kzalloc(sizeof(*con), GFP_KERNEL); 1626 if (!con) 1627 return -ENOMEM; 1628 1629 /* Map first two connections to the first CPU */ 1630 con->cpu = (cid ? cid - 1 : 0) % nr_cpu_ids; 1631 con->c.cid = cid; 1632 con->c.path = &clt_path->s; 1633 /* Align with srv, init as 1 */ 1634 atomic_set(&con->c.wr_cnt, 1); 1635 mutex_init(&con->con_mutex); 1636 1637 clt_path->s.con[cid] = &con->c; 1638 1639 return 0; 1640 } 1641 1642 static void destroy_con(struct rtrs_clt_con *con) 1643 { 1644 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 1645 1646 clt_path->s.con[con->c.cid] = NULL; 1647 mutex_destroy(&con->con_mutex); 1648 kfree(con); 1649 } 1650 1651 static int create_con_cq_qp(struct rtrs_clt_con *con) 1652 { 1653 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 1654 u32 max_send_wr, max_recv_wr, cq_num, max_send_sge, wr_limit; 1655 int err, cq_vector; 1656 struct rtrs_msg_rkey_rsp *rsp; 1657 1658 lockdep_assert_held(&con->con_mutex); 1659 if (con->c.cid == 0) { 1660 max_send_sge = 1; 1661 /* We must be the first here */ 1662 if (WARN_ON(clt_path->s.dev)) 1663 return -EINVAL; 1664 1665 /* 1666 * The whole session uses device from user connection. 1667 * Be careful not to close user connection before ib dev 1668 * is gracefully put. 1669 */ 1670 clt_path->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device, 1671 &dev_pd); 1672 if (!clt_path->s.dev) { 1673 rtrs_wrn(clt_path->clt, 1674 "rtrs_ib_dev_find_get_or_add(): no memory\n"); 1675 return -ENOMEM; 1676 } 1677 clt_path->s.dev_ref = 1; 1678 query_fast_reg_mode(clt_path); 1679 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr; 1680 /* 1681 * Two (request + registration) completion for send 1682 * Two for recv if always_invalidate is set on server 1683 * or one for recv. 1684 * + 2 for drain and heartbeat 1685 * in case qp gets into error state. 1686 */ 1687 max_send_wr = 1688 min_t(int, wr_limit, SERVICE_CON_QUEUE_DEPTH * 2 + 2); 1689 max_recv_wr = max_send_wr; 1690 } else { 1691 /* 1692 * Here we assume that session members are correctly set. 1693 * This is always true if user connection (cid == 0) is 1694 * established first. 1695 */ 1696 if (WARN_ON(!clt_path->s.dev)) 1697 return -EINVAL; 1698 if (WARN_ON(!clt_path->queue_depth)) 1699 return -EINVAL; 1700 1701 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr; 1702 /* Shared between connections */ 1703 clt_path->s.dev_ref++; 1704 max_send_wr = min_t(int, wr_limit, 1705 /* QD * (REQ + RSP + FR REGS or INVS) + drain */ 1706 clt_path->queue_depth * 4 + 1); 1707 max_recv_wr = min_t(int, wr_limit, 1708 clt_path->queue_depth * 3 + 1); 1709 max_send_sge = 2; 1710 } 1711 atomic_set(&con->c.sq_wr_avail, max_send_wr); 1712 cq_num = max_send_wr + max_recv_wr; 1713 /* alloc iu to recv new rkey reply when server reports flags set */ 1714 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) { 1715 con->rsp_ius = rtrs_iu_alloc(cq_num, sizeof(*rsp), 1716 GFP_KERNEL, 1717 clt_path->s.dev->ib_dev, 1718 DMA_FROM_DEVICE, 1719 rtrs_clt_rdma_done); 1720 if (!con->rsp_ius) 1721 return -ENOMEM; 1722 con->queue_num = cq_num; 1723 } 1724 cq_vector = con->cpu % clt_path->s.dev->ib_dev->num_comp_vectors; 1725 if (con->c.cid >= clt_path->s.irq_con_num) 1726 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge, 1727 cq_vector, cq_num, max_send_wr, 1728 max_recv_wr, IB_POLL_DIRECT); 1729 else 1730 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge, 1731 cq_vector, cq_num, max_send_wr, 1732 max_recv_wr, IB_POLL_SOFTIRQ); 1733 /* 1734 * In case of error we do not bother to clean previous allocations, 1735 * since destroy_con_cq_qp() must be called. 1736 */ 1737 return err; 1738 } 1739 1740 static void destroy_con_cq_qp(struct rtrs_clt_con *con) 1741 { 1742 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 1743 1744 /* 1745 * Be careful here: destroy_con_cq_qp() can be called even 1746 * create_con_cq_qp() failed, see comments there. 1747 */ 1748 lockdep_assert_held(&con->con_mutex); 1749 rtrs_cq_qp_destroy(&con->c); 1750 if (con->rsp_ius) { 1751 rtrs_iu_free(con->rsp_ius, clt_path->s.dev->ib_dev, 1752 con->queue_num); 1753 con->rsp_ius = NULL; 1754 con->queue_num = 0; 1755 } 1756 if (clt_path->s.dev_ref && !--clt_path->s.dev_ref) { 1757 rtrs_ib_dev_put(clt_path->s.dev); 1758 clt_path->s.dev = NULL; 1759 } 1760 } 1761 1762 static void stop_cm(struct rtrs_clt_con *con) 1763 { 1764 rdma_disconnect(con->c.cm_id); 1765 if (con->c.qp) 1766 ib_drain_qp(con->c.qp); 1767 } 1768 1769 static void destroy_cm(struct rtrs_clt_con *con) 1770 { 1771 rdma_destroy_id(con->c.cm_id); 1772 con->c.cm_id = NULL; 1773 } 1774 1775 static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con) 1776 { 1777 struct rtrs_path *s = con->c.path; 1778 int err; 1779 1780 mutex_lock(&con->con_mutex); 1781 err = create_con_cq_qp(con); 1782 mutex_unlock(&con->con_mutex); 1783 if (err) { 1784 rtrs_err(s, "create_con_cq_qp(), err: %pe\n", ERR_PTR(err)); 1785 return err; 1786 } 1787 err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS); 1788 if (err) 1789 rtrs_err(s, "Resolving route failed, err: %pe\n", ERR_PTR(err)); 1790 1791 return err; 1792 } 1793 1794 static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con) 1795 { 1796 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 1797 struct rtrs_clt_sess *clt = clt_path->clt; 1798 struct rtrs_msg_conn_req msg; 1799 struct rdma_conn_param param; 1800 1801 int err; 1802 1803 param = (struct rdma_conn_param) { 1804 .retry_count = 7, 1805 .rnr_retry_count = 7, 1806 .private_data = &msg, 1807 .private_data_len = sizeof(msg), 1808 }; 1809 1810 msg = (struct rtrs_msg_conn_req) { 1811 .magic = cpu_to_le16(RTRS_MAGIC), 1812 .version = cpu_to_le16(RTRS_PROTO_VER), 1813 .cid = cpu_to_le16(con->c.cid), 1814 .cid_num = cpu_to_le16(clt_path->s.con_num), 1815 .recon_cnt = cpu_to_le16(clt_path->s.recon_cnt), 1816 }; 1817 msg.first_conn = clt_path->for_new_clt ? FIRST_CONN : 0; 1818 uuid_copy(&msg.sess_uuid, &clt_path->s.uuid); 1819 uuid_copy(&msg.paths_uuid, &clt->paths_uuid); 1820 1821 err = rdma_connect_locked(con->c.cm_id, ¶m); 1822 if (err) 1823 rtrs_err(clt, "rdma_connect_locked(): %pe\n", ERR_PTR(err)); 1824 1825 return err; 1826 } 1827 1828 static int rtrs_rdma_conn_established(struct rtrs_clt_con *con, 1829 struct rdma_cm_event *ev) 1830 { 1831 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 1832 struct rtrs_clt_sess *clt = clt_path->clt; 1833 const struct rtrs_msg_conn_rsp *msg; 1834 u16 version, queue_depth; 1835 int errno; 1836 u8 len; 1837 1838 msg = ev->param.conn.private_data; 1839 len = ev->param.conn.private_data_len; 1840 if (len < sizeof(*msg)) { 1841 rtrs_err(clt, "Invalid RTRS connection response\n"); 1842 return -ECONNRESET; 1843 } 1844 if (le16_to_cpu(msg->magic) != RTRS_MAGIC) { 1845 rtrs_err(clt, "Invalid RTRS magic\n"); 1846 return -ECONNRESET; 1847 } 1848 version = le16_to_cpu(msg->version); 1849 if (version >> 8 != RTRS_PROTO_VER_MAJOR) { 1850 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n", 1851 version >> 8, RTRS_PROTO_VER_MAJOR); 1852 return -ECONNRESET; 1853 } 1854 errno = le16_to_cpu(msg->errno); 1855 if (errno) { 1856 rtrs_err(clt, "Invalid RTRS message: errno %pe\n", 1857 ERR_PTR(errno)); 1858 return -ECONNRESET; 1859 } 1860 if (con->c.cid == 0) { 1861 queue_depth = le16_to_cpu(msg->queue_depth); 1862 1863 if (clt_path->queue_depth > 0 && queue_depth != clt_path->queue_depth) { 1864 rtrs_err(clt, "Error: queue depth changed\n"); 1865 1866 /* 1867 * Stop any more reconnection attempts 1868 */ 1869 clt_path->reconnect_attempts = -1; 1870 rtrs_err(clt, 1871 "Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n"); 1872 return -ECONNRESET; 1873 } 1874 1875 if (!clt_path->rbufs) { 1876 clt_path->rbufs = kcalloc(queue_depth, 1877 sizeof(*clt_path->rbufs), 1878 GFP_KERNEL); 1879 if (!clt_path->rbufs) 1880 return -ENOMEM; 1881 } 1882 clt_path->queue_depth = queue_depth; 1883 clt_path->s.signal_interval = min_not_zero(queue_depth, 1884 (unsigned short) SERVICE_CON_QUEUE_DEPTH); 1885 clt_path->max_hdr_size = le32_to_cpu(msg->max_hdr_size); 1886 clt_path->max_io_size = le32_to_cpu(msg->max_io_size); 1887 clt_path->flags = le32_to_cpu(msg->flags); 1888 clt_path->chunk_size = clt_path->max_io_size + clt_path->max_hdr_size; 1889 1890 /* 1891 * Global IO size is always a minimum. 1892 * If while a reconnection server sends us a value a bit 1893 * higher - client does not care and uses cached minimum. 1894 * 1895 * Since we can have several sessions (paths) restablishing 1896 * connections in parallel, use lock. 1897 */ 1898 mutex_lock(&clt->paths_mutex); 1899 clt->queue_depth = clt_path->queue_depth; 1900 clt->max_io_size = min_not_zero(clt_path->max_io_size, 1901 clt->max_io_size); 1902 mutex_unlock(&clt->paths_mutex); 1903 1904 /* 1905 * Cache the hca_port and hca_name for sysfs 1906 */ 1907 clt_path->hca_port = con->c.cm_id->port_num; 1908 scnprintf(clt_path->hca_name, sizeof(clt_path->hca_name), 1909 clt_path->s.dev->ib_dev->name); 1910 clt_path->s.src_addr = con->c.cm_id->route.addr.src_addr; 1911 /* set for_new_clt, to allow future reconnect on any path */ 1912 clt_path->for_new_clt = 1; 1913 } 1914 1915 return 0; 1916 } 1917 1918 static inline void flag_success_on_conn(struct rtrs_clt_con *con) 1919 { 1920 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 1921 1922 atomic_inc(&clt_path->connected_cnt); 1923 con->cm_err = 1; 1924 } 1925 1926 static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con, 1927 struct rdma_cm_event *ev) 1928 { 1929 struct rtrs_path *s = con->c.path; 1930 const struct rtrs_msg_conn_rsp *msg; 1931 const char *rej_msg; 1932 int status, errno = -ECONNRESET; 1933 u8 data_len; 1934 1935 status = ev->status; 1936 rej_msg = rdma_reject_msg(con->c.cm_id, status); 1937 msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len); 1938 1939 if (msg && data_len >= sizeof(*msg)) { 1940 errno = (int16_t)le16_to_cpu(msg->errno); 1941 if (errno == -EBUSY) 1942 rtrs_err(s, 1943 "Previous session is still exists on the server, please reconnect later\n"); 1944 else 1945 rtrs_err(s, 1946 "Connect rejected: status %d (%s), rtrs errno %pe\n", 1947 status, rej_msg, ERR_PTR(errno)); 1948 } else { 1949 rtrs_err(s, 1950 "Connect rejected but with malformed message: status %d (%s)\n", 1951 status, rej_msg); 1952 } 1953 1954 return errno; 1955 } 1956 1957 void rtrs_clt_close_conns(struct rtrs_clt_path *clt_path, bool wait) 1958 { 1959 trace_rtrs_clt_close_conns(clt_path); 1960 1961 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSING, NULL)) 1962 queue_work(rtrs_wq, &clt_path->close_work); 1963 if (wait) 1964 flush_work(&clt_path->close_work); 1965 } 1966 1967 static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err) 1968 { 1969 if (con->cm_err == 1) { 1970 struct rtrs_clt_path *clt_path; 1971 1972 clt_path = to_clt_path(con->c.path); 1973 if (atomic_dec_and_test(&clt_path->connected_cnt)) 1974 1975 wake_up(&clt_path->state_wq); 1976 } 1977 con->cm_err = cm_err; 1978 } 1979 1980 static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id, 1981 struct rdma_cm_event *ev) 1982 { 1983 struct rtrs_clt_con *con = cm_id->context; 1984 struct rtrs_path *s = con->c.path; 1985 struct rtrs_clt_path *clt_path = to_clt_path(s); 1986 int cm_err = 0; 1987 1988 switch (ev->event) { 1989 case RDMA_CM_EVENT_ADDR_RESOLVED: 1990 cm_err = rtrs_rdma_addr_resolved(con); 1991 break; 1992 case RDMA_CM_EVENT_ROUTE_RESOLVED: 1993 cm_err = rtrs_rdma_route_resolved(con); 1994 break; 1995 case RDMA_CM_EVENT_ESTABLISHED: 1996 cm_err = rtrs_rdma_conn_established(con, ev); 1997 if (!cm_err) { 1998 /* 1999 * Report success and wake up. Here we abuse state_wq, 2000 * i.e. wake up without state change, but we set cm_err. 2001 */ 2002 flag_success_on_conn(con); 2003 wake_up(&clt_path->state_wq); 2004 return 0; 2005 } 2006 break; 2007 case RDMA_CM_EVENT_REJECTED: 2008 cm_err = rtrs_rdma_conn_rejected(con, ev); 2009 break; 2010 case RDMA_CM_EVENT_DISCONNECTED: 2011 /* No message for disconnecting */ 2012 cm_err = -ECONNRESET; 2013 break; 2014 case RDMA_CM_EVENT_CONNECT_ERROR: 2015 case RDMA_CM_EVENT_UNREACHABLE: 2016 case RDMA_CM_EVENT_ADDR_CHANGE: 2017 case RDMA_CM_EVENT_TIMEWAIT_EXIT: 2018 if (ev->status < 0) { 2019 rtrs_wrn(s, "CM error (CM event: %s, err: %pe)\n", 2020 rdma_event_msg(ev->event), ERR_PTR(ev->status)); 2021 } else if (ev->status > 0) { 2022 rtrs_wrn(s, "CM error (CM event: %s, err: %s)\n", 2023 rdma_event_msg(ev->event), 2024 rdma_reject_msg(cm_id, ev->status)); 2025 } 2026 cm_err = -ECONNRESET; 2027 break; 2028 case RDMA_CM_EVENT_ADDR_ERROR: 2029 case RDMA_CM_EVENT_ROUTE_ERROR: 2030 if (ev->status < 0) { 2031 rtrs_wrn(s, "CM error (CM event: %s, err: %pe)\n", 2032 rdma_event_msg(ev->event), 2033 ERR_PTR(ev->status)); 2034 } else if (ev->status > 0) { 2035 rtrs_wrn(s, "CM error (CM event: %s, err: %s)\n", 2036 rdma_event_msg(ev->event), 2037 rdma_reject_msg(cm_id, ev->status)); 2038 } 2039 cm_err = -EHOSTUNREACH; 2040 break; 2041 case RDMA_CM_EVENT_DEVICE_REMOVAL: 2042 /* 2043 * Device removal is a special case. Queue close and return 0. 2044 */ 2045 if (ev->status < 0) { 2046 rtrs_wrn_rl(s, "CM event: %s, status: %pe\n", 2047 rdma_event_msg(ev->event), 2048 ERR_PTR(ev->status)); 2049 } else if (ev->status > 0) { 2050 rtrs_wrn_rl(s, "CM event: %s, status: %s\n", 2051 rdma_event_msg(ev->event), 2052 rdma_reject_msg(cm_id, ev->status)); 2053 } 2054 rtrs_clt_close_conns(clt_path, false); 2055 return 0; 2056 default: 2057 if (ev->status < 0) { 2058 rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %pe)\n", 2059 rdma_event_msg(ev->event), ERR_PTR(ev->status)); 2060 } else if (ev->status > 0) { 2061 rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %s)\n", 2062 rdma_event_msg(ev->event), 2063 rdma_reject_msg(cm_id, ev->status)); 2064 } 2065 cm_err = -ECONNRESET; 2066 break; 2067 } 2068 2069 if (cm_err) { 2070 /* 2071 * cm error makes sense only on connection establishing, 2072 * in other cases we rely on normal procedure of reconnecting. 2073 */ 2074 flag_error_on_conn(con, cm_err); 2075 rtrs_rdma_error_recovery(con); 2076 } 2077 2078 return 0; 2079 } 2080 2081 /* The caller should do the cleanup in case of error */ 2082 static int create_cm(struct rtrs_clt_con *con) 2083 { 2084 struct rtrs_path *s = con->c.path; 2085 struct rtrs_clt_path *clt_path = to_clt_path(s); 2086 struct rdma_cm_id *cm_id; 2087 int err; 2088 2089 cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con, 2090 clt_path->s.dst_addr.ss_family == AF_IB ? 2091 RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC); 2092 if (IS_ERR(cm_id)) { 2093 rtrs_err(s, "Failed to create CM ID, err: %pe\n", cm_id); 2094 return PTR_ERR(cm_id); 2095 } 2096 con->c.cm_id = cm_id; 2097 con->cm_err = 0; 2098 /* allow the port to be reused */ 2099 err = rdma_set_reuseaddr(cm_id, 1); 2100 if (err != 0) { 2101 rtrs_err(s, "Set address reuse failed, err: %pe\n", ERR_PTR(err)); 2102 return err; 2103 } 2104 err = rdma_resolve_addr(cm_id, (struct sockaddr *)&clt_path->s.src_addr, 2105 (struct sockaddr *)&clt_path->s.dst_addr, 2106 RTRS_CONNECT_TIMEOUT_MS); 2107 if (err) { 2108 rtrs_err(s, "Failed to resolve address, err: %pe\n", ERR_PTR(err)); 2109 return err; 2110 } 2111 /* 2112 * Combine connection status and session events. This is needed 2113 * for waiting two possible cases: cm_err has something meaningful 2114 * or session state was really changed to error by device removal. 2115 */ 2116 err = wait_event_interruptible_timeout( 2117 clt_path->state_wq, 2118 con->cm_err || clt_path->state != RTRS_CLT_CONNECTING, 2119 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS)); 2120 if (err == 0 || err == -ERESTARTSYS) { 2121 if (err == 0) 2122 err = -ETIMEDOUT; 2123 /* Timedout or interrupted */ 2124 return err; 2125 } 2126 if (con->cm_err < 0) 2127 return con->cm_err; 2128 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTING) 2129 /* Device removal */ 2130 return -ECONNABORTED; 2131 2132 return 0; 2133 } 2134 2135 static void rtrs_clt_path_up(struct rtrs_clt_path *clt_path) 2136 { 2137 struct rtrs_clt_sess *clt = clt_path->clt; 2138 int up; 2139 2140 /* 2141 * We can fire RECONNECTED event only when all paths were 2142 * connected on rtrs_clt_open(), then each was disconnected 2143 * and the first one connected again. That's why this nasty 2144 * game with counter value. 2145 */ 2146 2147 mutex_lock(&clt->paths_ev_mutex); 2148 up = ++clt->paths_up; 2149 /* 2150 * Here it is safe to access paths num directly since up counter 2151 * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is 2152 * in progress, thus paths removals are impossible. 2153 */ 2154 if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num) 2155 clt->paths_up = clt->paths_num; 2156 else if (up == 1) 2157 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED); 2158 mutex_unlock(&clt->paths_ev_mutex); 2159 2160 /* Mark session as established */ 2161 clt_path->established = true; 2162 clt_path->reconnect_attempts = 0; 2163 clt_path->stats->reconnects.successful_cnt++; 2164 } 2165 2166 static void rtrs_clt_path_down(struct rtrs_clt_path *clt_path) 2167 { 2168 struct rtrs_clt_sess *clt = clt_path->clt; 2169 2170 if (!clt_path->established) 2171 return; 2172 2173 clt_path->established = false; 2174 mutex_lock(&clt->paths_ev_mutex); 2175 WARN_ON(!clt->paths_up); 2176 if (--clt->paths_up == 0) 2177 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED); 2178 mutex_unlock(&clt->paths_ev_mutex); 2179 } 2180 2181 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path) 2182 { 2183 struct rtrs_clt_con *con; 2184 unsigned int cid; 2185 2186 WARN_ON(READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED); 2187 2188 /* 2189 * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes 2190 * exactly in between. Start destroying after it finishes. 2191 */ 2192 mutex_lock(&clt_path->init_mutex); 2193 mutex_unlock(&clt_path->init_mutex); 2194 2195 /* 2196 * All IO paths must observe !CONNECTED state before we 2197 * free everything. 2198 */ 2199 synchronize_rcu(); 2200 2201 rtrs_stop_hb(&clt_path->s); 2202 2203 /* 2204 * The order it utterly crucial: firstly disconnect and complete all 2205 * rdma requests with error (thus set in_use=false for requests), 2206 * then fail outstanding requests checking in_use for each, and 2207 * eventually notify upper layer about session disconnection. 2208 */ 2209 2210 for (cid = 0; cid < clt_path->s.con_num; cid++) { 2211 if (!clt_path->s.con[cid]) 2212 break; 2213 con = to_clt_con(clt_path->s.con[cid]); 2214 stop_cm(con); 2215 } 2216 fail_all_outstanding_reqs(clt_path); 2217 free_path_reqs(clt_path); 2218 rtrs_clt_path_down(clt_path); 2219 2220 /* 2221 * Wait for graceful shutdown, namely when peer side invokes 2222 * rdma_disconnect(). 'connected_cnt' is decremented only on 2223 * CM events, thus if other side had crashed and hb has detected 2224 * something is wrong, here we will stuck for exactly timeout ms, 2225 * since CM does not fire anything. That is fine, we are not in 2226 * hurry. 2227 */ 2228 wait_event_timeout(clt_path->state_wq, 2229 !atomic_read(&clt_path->connected_cnt), 2230 msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS)); 2231 2232 for (cid = 0; cid < clt_path->s.con_num; cid++) { 2233 if (!clt_path->s.con[cid]) 2234 break; 2235 con = to_clt_con(clt_path->s.con[cid]); 2236 mutex_lock(&con->con_mutex); 2237 destroy_con_cq_qp(con); 2238 mutex_unlock(&con->con_mutex); 2239 destroy_cm(con); 2240 destroy_con(con); 2241 } 2242 } 2243 2244 static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_path *clt_path) 2245 { 2246 struct rtrs_clt_sess *clt = clt_path->clt; 2247 struct rtrs_clt_path *next; 2248 bool wait_for_grace = false; 2249 int cpu; 2250 2251 mutex_lock(&clt->paths_mutex); 2252 list_del_rcu(&clt_path->s.entry); 2253 2254 /* Make sure everybody observes path removal. */ 2255 synchronize_rcu(); 2256 2257 /* 2258 * At this point nobody sees @sess in the list, but still we have 2259 * dangling pointer @pcpu_path which _can_ point to @sess. Since 2260 * nobody can observe @sess in the list, we guarantee that IO path 2261 * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal 2262 * to @sess, but can never again become @sess. 2263 */ 2264 2265 /* 2266 * Decrement paths number only after grace period, because 2267 * caller of do_each_path() must firstly observe list without 2268 * path and only then decremented paths number. 2269 * 2270 * Otherwise there can be the following situation: 2271 * o Two paths exist and IO is coming. 2272 * o One path is removed: 2273 * CPU#0 CPU#1 2274 * do_each_path(): rtrs_clt_remove_path_from_arr(): 2275 * path = get_next_path() 2276 * ^^^ list_del_rcu(path) 2277 * [!CONNECTED path] clt->paths_num-- 2278 * ^^^^^^^^^ 2279 * load clt->paths_num from 2 to 1 2280 * ^^^^^^^^^ 2281 * sees 1 2282 * 2283 * path is observed as !CONNECTED, but do_each_path() loop 2284 * ends, because expression i < clt->paths_num is false. 2285 */ 2286 clt->paths_num--; 2287 2288 /* 2289 * Get @next connection from current @sess which is going to be 2290 * removed. If @sess is the last element, then @next is NULL. 2291 */ 2292 rcu_read_lock(); 2293 next = rtrs_clt_get_next_path_or_null(&clt->paths_list, clt_path); 2294 rcu_read_unlock(); 2295 2296 /* 2297 * @pcpu paths can still point to the path which is going to be 2298 * removed, so change the pointer manually. 2299 */ 2300 for_each_possible_cpu(cpu) { 2301 struct rtrs_clt_path __rcu **ppcpu_path; 2302 2303 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu); 2304 if (rcu_dereference_protected(*ppcpu_path, 2305 lockdep_is_held(&clt->paths_mutex)) != clt_path) 2306 /* 2307 * synchronize_rcu() was called just after deleting 2308 * entry from the list, thus IO code path cannot 2309 * change pointer back to the pointer which is going 2310 * to be removed, we are safe here. 2311 */ 2312 continue; 2313 2314 /* 2315 * We race with IO code path, which also changes pointer, 2316 * thus we have to be careful not to overwrite it. 2317 */ 2318 if (try_cmpxchg((struct rtrs_clt_path **)ppcpu_path, &clt_path, 2319 next)) 2320 /* 2321 * @ppcpu_path was successfully replaced with @next, 2322 * that means that someone could also pick up the 2323 * @sess and dereferencing it right now, so wait for 2324 * a grace period is required. 2325 */ 2326 wait_for_grace = true; 2327 } 2328 if (wait_for_grace) 2329 synchronize_rcu(); 2330 2331 mutex_unlock(&clt->paths_mutex); 2332 } 2333 2334 static void rtrs_clt_add_path_to_arr(struct rtrs_clt_path *clt_path) 2335 { 2336 struct rtrs_clt_sess *clt = clt_path->clt; 2337 2338 mutex_lock(&clt->paths_mutex); 2339 clt->paths_num++; 2340 2341 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list); 2342 mutex_unlock(&clt->paths_mutex); 2343 } 2344 2345 static void rtrs_clt_close_work(struct work_struct *work) 2346 { 2347 struct rtrs_clt_path *clt_path; 2348 2349 clt_path = container_of(work, struct rtrs_clt_path, close_work); 2350 2351 cancel_work_sync(&clt_path->err_recovery_work); 2352 cancel_delayed_work_sync(&clt_path->reconnect_dwork); 2353 rtrs_clt_stop_and_destroy_conns(clt_path); 2354 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSED, NULL); 2355 } 2356 2357 static int init_conns(struct rtrs_clt_path *clt_path) 2358 { 2359 unsigned int cid; 2360 int err, i; 2361 2362 /* 2363 * On every new session connections increase reconnect counter 2364 * to avoid clashes with previous sessions not yet closed 2365 * sessions on a server side. 2366 */ 2367 clt_path->s.recon_cnt++; 2368 2369 /* Establish all RDMA connections */ 2370 for (cid = 0; cid < clt_path->s.con_num; cid++) { 2371 err = create_con(clt_path, cid); 2372 if (err) 2373 goto destroy; 2374 2375 err = create_cm(to_clt_con(clt_path->s.con[cid])); 2376 if (err) 2377 goto destroy; 2378 } 2379 2380 /* 2381 * Set the cid to con_num - 1, since if we fail later, we want to stay in bounds. 2382 */ 2383 cid = clt_path->s.con_num - 1; 2384 2385 err = alloc_path_reqs(clt_path); 2386 if (err) 2387 goto destroy; 2388 2389 return 0; 2390 2391 destroy: 2392 /* Make sure we do the cleanup in the order they are created */ 2393 for (i = 0; i <= cid; i++) { 2394 struct rtrs_clt_con *con; 2395 2396 if (!clt_path->s.con[i]) 2397 break; 2398 2399 con = to_clt_con(clt_path->s.con[i]); 2400 if (con->c.cm_id) { 2401 stop_cm(con); 2402 mutex_lock(&con->con_mutex); 2403 destroy_con_cq_qp(con); 2404 mutex_unlock(&con->con_mutex); 2405 destroy_cm(con); 2406 } 2407 destroy_con(con); 2408 } 2409 /* 2410 * If we've never taken async path and got an error, say, 2411 * doing rdma_resolve_addr(), switch to CONNECTION_ERR state 2412 * manually to keep reconnecting. 2413 */ 2414 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL); 2415 2416 return err; 2417 } 2418 2419 static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc) 2420 { 2421 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context); 2422 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 2423 struct rtrs_iu *iu; 2424 2425 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe); 2426 rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1); 2427 2428 if (wc->status != IB_WC_SUCCESS) { 2429 rtrs_err(clt_path->clt, "Path info request send failed: %s\n", 2430 ib_wc_status_msg(wc->status)); 2431 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL); 2432 return; 2433 } 2434 2435 rtrs_clt_update_wc_stats(con); 2436 } 2437 2438 static int process_info_rsp(struct rtrs_clt_path *clt_path, 2439 const struct rtrs_msg_info_rsp *msg) 2440 { 2441 unsigned int sg_cnt, total_len; 2442 int i, sgi; 2443 2444 sg_cnt = le16_to_cpu(msg->sg_cnt); 2445 if (!sg_cnt || (clt_path->queue_depth % sg_cnt)) { 2446 rtrs_err(clt_path->clt, 2447 "Incorrect sg_cnt %d, is not multiple\n", 2448 sg_cnt); 2449 return -EINVAL; 2450 } 2451 2452 /* 2453 * Check if IB immediate data size is enough to hold the mem_id and 2454 * the offset inside the memory chunk. 2455 */ 2456 if ((ilog2(sg_cnt - 1) + 1) + (ilog2(clt_path->chunk_size - 1) + 1) > 2457 MAX_IMM_PAYL_BITS) { 2458 rtrs_err(clt_path->clt, 2459 "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n", 2460 MAX_IMM_PAYL_BITS, sg_cnt, clt_path->chunk_size); 2461 return -EINVAL; 2462 } 2463 total_len = 0; 2464 for (sgi = 0, i = 0; sgi < sg_cnt && i < clt_path->queue_depth; sgi++) { 2465 const struct rtrs_sg_desc *desc = &msg->desc[sgi]; 2466 u32 len, rkey; 2467 u64 addr; 2468 2469 addr = le64_to_cpu(desc->addr); 2470 rkey = le32_to_cpu(desc->key); 2471 len = le32_to_cpu(desc->len); 2472 2473 total_len += len; 2474 2475 if (!len || (len % clt_path->chunk_size)) { 2476 rtrs_err(clt_path->clt, "Incorrect [%d].len %d\n", 2477 sgi, 2478 len); 2479 return -EINVAL; 2480 } 2481 for ( ; len && i < clt_path->queue_depth; i++) { 2482 clt_path->rbufs[i].addr = addr; 2483 clt_path->rbufs[i].rkey = rkey; 2484 2485 len -= clt_path->chunk_size; 2486 addr += clt_path->chunk_size; 2487 } 2488 } 2489 /* Sanity check */ 2490 if (sgi != sg_cnt || i != clt_path->queue_depth) { 2491 rtrs_err(clt_path->clt, 2492 "Incorrect sg vector, not fully mapped\n"); 2493 return -EINVAL; 2494 } 2495 if (total_len != clt_path->chunk_size * clt_path->queue_depth) { 2496 rtrs_err(clt_path->clt, "Incorrect total_len %d\n", total_len); 2497 return -EINVAL; 2498 } 2499 2500 return 0; 2501 } 2502 2503 static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc) 2504 { 2505 struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context); 2506 struct rtrs_clt_path *clt_path = to_clt_path(con->c.path); 2507 struct rtrs_msg_info_rsp *msg; 2508 enum rtrs_clt_state state; 2509 struct rtrs_iu *iu; 2510 size_t rx_sz; 2511 int err; 2512 2513 state = RTRS_CLT_CONNECTING_ERR; 2514 2515 WARN_ON(con->c.cid); 2516 iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe); 2517 if (wc->status != IB_WC_SUCCESS) { 2518 rtrs_err(clt_path->clt, "Path info response recv failed: %s\n", 2519 ib_wc_status_msg(wc->status)); 2520 goto out; 2521 } 2522 WARN_ON(wc->opcode != IB_WC_RECV); 2523 2524 if (wc->byte_len < sizeof(*msg)) { 2525 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n", 2526 wc->byte_len); 2527 goto out; 2528 } 2529 ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr, 2530 iu->size, DMA_FROM_DEVICE); 2531 msg = iu->buf; 2532 if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP) { 2533 rtrs_err(clt_path->clt, "Path info response is malformed: type %d\n", 2534 le16_to_cpu(msg->type)); 2535 goto out; 2536 } 2537 rx_sz = sizeof(*msg); 2538 rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt); 2539 if (wc->byte_len < rx_sz) { 2540 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n", 2541 wc->byte_len); 2542 goto out; 2543 } 2544 err = process_info_rsp(clt_path, msg); 2545 if (err) 2546 goto out; 2547 2548 err = post_recv_path(clt_path); 2549 if (err) 2550 goto out; 2551 2552 state = RTRS_CLT_CONNECTED; 2553 2554 out: 2555 rtrs_clt_update_wc_stats(con); 2556 rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1); 2557 rtrs_clt_change_state_get_old(clt_path, state, NULL); 2558 } 2559 2560 static int rtrs_send_path_info(struct rtrs_clt_path *clt_path) 2561 { 2562 struct rtrs_clt_con *usr_con = to_clt_con(clt_path->s.con[0]); 2563 struct rtrs_msg_info_req *msg; 2564 struct rtrs_iu *tx_iu, *rx_iu; 2565 size_t rx_sz; 2566 int err; 2567 2568 rx_sz = sizeof(struct rtrs_msg_info_rsp); 2569 rx_sz += sizeof(struct rtrs_sg_desc) * clt_path->queue_depth; 2570 2571 tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL, 2572 clt_path->s.dev->ib_dev, DMA_TO_DEVICE, 2573 rtrs_clt_info_req_done); 2574 rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, clt_path->s.dev->ib_dev, 2575 DMA_FROM_DEVICE, rtrs_clt_info_rsp_done); 2576 if (!tx_iu || !rx_iu) { 2577 err = -ENOMEM; 2578 goto out; 2579 } 2580 /* Prepare for getting info response */ 2581 err = rtrs_iu_post_recv(&usr_con->c, rx_iu); 2582 if (err) { 2583 rtrs_err(clt_path->clt, "rtrs_iu_post_recv(), err: %pe\n", ERR_PTR(err)); 2584 goto out; 2585 } 2586 rx_iu = NULL; 2587 2588 msg = tx_iu->buf; 2589 msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ); 2590 memcpy(msg->pathname, clt_path->s.sessname, sizeof(msg->pathname)); 2591 2592 ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, 2593 tx_iu->dma_addr, 2594 tx_iu->size, DMA_TO_DEVICE); 2595 2596 /* Send info request */ 2597 err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL); 2598 if (err) { 2599 rtrs_err(clt_path->clt, "rtrs_iu_post_send(), err: %pe\n", ERR_PTR(err)); 2600 goto out; 2601 } 2602 tx_iu = NULL; 2603 2604 /* Wait for state change */ 2605 wait_event_interruptible_timeout(clt_path->state_wq, 2606 clt_path->state != RTRS_CLT_CONNECTING, 2607 msecs_to_jiffies( 2608 RTRS_CONNECT_TIMEOUT_MS)); 2609 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) { 2610 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTING_ERR) 2611 err = -ECONNRESET; 2612 else 2613 err = -ETIMEDOUT; 2614 } 2615 2616 out: 2617 if (tx_iu) 2618 rtrs_iu_free(tx_iu, clt_path->s.dev->ib_dev, 1); 2619 if (rx_iu) 2620 rtrs_iu_free(rx_iu, clt_path->s.dev->ib_dev, 1); 2621 if (err) 2622 /* If we've never taken async path because of malloc problems */ 2623 rtrs_clt_change_state_get_old(clt_path, 2624 RTRS_CLT_CONNECTING_ERR, NULL); 2625 2626 return err; 2627 } 2628 2629 /** 2630 * init_path() - establishes all path connections and does handshake 2631 * @clt_path: client path. 2632 * In case of error full close or reconnect procedure should be taken, 2633 * because reconnect or close async works can be started. 2634 */ 2635 static int init_path(struct rtrs_clt_path *clt_path) 2636 { 2637 int err; 2638 char str[NAME_MAX]; 2639 struct rtrs_addr path = { 2640 .src = &clt_path->s.src_addr, 2641 .dst = &clt_path->s.dst_addr, 2642 }; 2643 2644 rtrs_addr_to_str(&path, str, sizeof(str)); 2645 2646 mutex_lock(&clt_path->init_mutex); 2647 err = init_conns(clt_path); 2648 if (err) { 2649 rtrs_err(clt_path->clt, 2650 "init_conns() failed: err=%pe path=%s [%s:%u]\n", 2651 ERR_PTR(err), str, clt_path->hca_name, clt_path->hca_port); 2652 goto out; 2653 } 2654 err = rtrs_send_path_info(clt_path); 2655 if (err) { 2656 rtrs_err(clt_path->clt, 2657 "rtrs_send_path_info() failed: err=%pe path=%s [%s:%u]\n", 2658 ERR_PTR(err), str, clt_path->hca_name, clt_path->hca_port); 2659 goto out; 2660 } 2661 rtrs_clt_path_up(clt_path); 2662 rtrs_start_hb(&clt_path->s); 2663 out: 2664 mutex_unlock(&clt_path->init_mutex); 2665 2666 return err; 2667 } 2668 2669 static void rtrs_clt_reconnect_work(struct work_struct *work) 2670 { 2671 struct rtrs_clt_path *clt_path; 2672 struct rtrs_clt_sess *clt; 2673 int err; 2674 2675 clt_path = container_of(to_delayed_work(work), struct rtrs_clt_path, 2676 reconnect_dwork); 2677 clt = clt_path->clt; 2678 2679 trace_rtrs_clt_reconnect_work(clt_path); 2680 2681 if (READ_ONCE(clt_path->state) != RTRS_CLT_RECONNECTING) 2682 return; 2683 2684 if (clt_path->reconnect_attempts >= clt->max_reconnect_attempts) { 2685 /* Close a path completely if max attempts is reached */ 2686 rtrs_clt_close_conns(clt_path, false); 2687 return; 2688 } 2689 clt_path->reconnect_attempts++; 2690 2691 msleep(RTRS_RECONNECT_BACKOFF); 2692 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING, NULL)) { 2693 err = init_path(clt_path); 2694 if (err) 2695 goto reconnect_again; 2696 } 2697 2698 return; 2699 2700 reconnect_again: 2701 if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_RECONNECTING, NULL)) { 2702 clt_path->stats->reconnects.fail_cnt++; 2703 queue_work(rtrs_wq, &clt_path->err_recovery_work); 2704 } 2705 } 2706 2707 static void rtrs_clt_dev_release(struct device *dev) 2708 { 2709 struct rtrs_clt_sess *clt = container_of(dev, struct rtrs_clt_sess, 2710 dev); 2711 2712 mutex_destroy(&clt->paths_ev_mutex); 2713 mutex_destroy(&clt->paths_mutex); 2714 kfree(clt); 2715 } 2716 2717 static struct rtrs_clt_sess *alloc_clt(const char *sessname, size_t paths_num, 2718 u16 port, size_t pdu_sz, void *priv, 2719 void (*link_ev)(void *priv, 2720 enum rtrs_clt_link_ev ev), 2721 unsigned int reconnect_delay_sec, 2722 unsigned int max_reconnect_attempts) 2723 { 2724 struct rtrs_clt_sess *clt; 2725 int err; 2726 2727 if (!paths_num || paths_num > MAX_PATHS_NUM) 2728 return ERR_PTR(-EINVAL); 2729 2730 if (strlen(sessname) >= sizeof(clt->sessname)) 2731 return ERR_PTR(-EINVAL); 2732 2733 clt = kzalloc(sizeof(*clt), GFP_KERNEL); 2734 if (!clt) 2735 return ERR_PTR(-ENOMEM); 2736 2737 clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path)); 2738 if (!clt->pcpu_path) { 2739 kfree(clt); 2740 return ERR_PTR(-ENOMEM); 2741 } 2742 2743 clt->dev.class = &rtrs_clt_dev_class; 2744 clt->dev.release = rtrs_clt_dev_release; 2745 uuid_gen(&clt->paths_uuid); 2746 INIT_LIST_HEAD_RCU(&clt->paths_list); 2747 clt->paths_num = paths_num; 2748 clt->paths_up = MAX_PATHS_NUM; 2749 clt->port = port; 2750 clt->pdu_sz = pdu_sz; 2751 clt->max_segments = RTRS_MAX_SEGMENTS; 2752 clt->reconnect_delay_sec = reconnect_delay_sec; 2753 clt->max_reconnect_attempts = max_reconnect_attempts; 2754 clt->priv = priv; 2755 clt->link_ev = link_ev; 2756 clt->mp_policy = MP_POLICY_MIN_INFLIGHT; 2757 strscpy(clt->sessname, sessname, sizeof(clt->sessname)); 2758 init_waitqueue_head(&clt->permits_wait); 2759 mutex_init(&clt->paths_ev_mutex); 2760 mutex_init(&clt->paths_mutex); 2761 device_initialize(&clt->dev); 2762 2763 err = dev_set_name(&clt->dev, "%s", sessname); 2764 if (err) 2765 goto err_put; 2766 2767 /* 2768 * Suppress user space notification until 2769 * sysfs files are created 2770 */ 2771 dev_set_uevent_suppress(&clt->dev, true); 2772 err = device_add(&clt->dev); 2773 if (err) 2774 goto err_put; 2775 2776 clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj); 2777 if (!clt->kobj_paths) { 2778 err = -ENOMEM; 2779 goto err_del; 2780 } 2781 err = rtrs_clt_create_sysfs_root_files(clt); 2782 if (err) { 2783 kobject_del(clt->kobj_paths); 2784 kobject_put(clt->kobj_paths); 2785 goto err_del; 2786 } 2787 dev_set_uevent_suppress(&clt->dev, false); 2788 kobject_uevent(&clt->dev.kobj, KOBJ_ADD); 2789 2790 return clt; 2791 err_del: 2792 device_del(&clt->dev); 2793 err_put: 2794 free_percpu(clt->pcpu_path); 2795 put_device(&clt->dev); 2796 return ERR_PTR(err); 2797 } 2798 2799 static void free_clt(struct rtrs_clt_sess *clt) 2800 { 2801 free_percpu(clt->pcpu_path); 2802 2803 /* 2804 * release callback will free clt and destroy mutexes in last put 2805 */ 2806 device_unregister(&clt->dev); 2807 } 2808 2809 /** 2810 * rtrs_clt_open() - Open a path to an RTRS server 2811 * @ops: holds the link event callback and the private pointer. 2812 * @pathname: name of the path to an RTRS server 2813 * @paths: Paths to be established defined by their src and dst addresses 2814 * @paths_num: Number of elements in the @paths array 2815 * @port: port to be used by the RTRS session 2816 * @pdu_sz: Size of extra payload which can be accessed after permit allocation. 2817 * @reconnect_delay_sec: time between reconnect tries 2818 * @max_reconnect_attempts: Number of times to reconnect on error before giving 2819 * up, 0 for * disabled, -1 for forever 2820 * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag 2821 * 2822 * Starts session establishment with the rtrs_server. The function can block 2823 * up to ~2000ms before it returns. 2824 * 2825 * Return a valid pointer on success otherwise PTR_ERR. 2826 */ 2827 struct rtrs_clt_sess *rtrs_clt_open(struct rtrs_clt_ops *ops, 2828 const char *pathname, 2829 const struct rtrs_addr *paths, 2830 size_t paths_num, u16 port, 2831 size_t pdu_sz, u8 reconnect_delay_sec, 2832 s16 max_reconnect_attempts, u32 nr_poll_queues) 2833 { 2834 struct rtrs_clt_path *clt_path, *tmp; 2835 struct rtrs_clt_sess *clt; 2836 int err, i; 2837 2838 if (strchr(pathname, '/') || strchr(pathname, '.')) { 2839 pr_err("pathname cannot contain / and .\n"); 2840 err = -EINVAL; 2841 goto out; 2842 } 2843 2844 clt = alloc_clt(pathname, paths_num, port, pdu_sz, ops->priv, 2845 ops->link_ev, 2846 reconnect_delay_sec, 2847 max_reconnect_attempts); 2848 if (IS_ERR(clt)) { 2849 err = PTR_ERR(clt); 2850 goto out; 2851 } 2852 for (i = 0; i < paths_num; i++) { 2853 struct rtrs_clt_path *clt_path; 2854 2855 clt_path = alloc_path(clt, &paths[i], nr_cpu_ids, 2856 nr_poll_queues); 2857 if (IS_ERR(clt_path)) { 2858 err = PTR_ERR(clt_path); 2859 goto close_all_path; 2860 } 2861 if (!i) 2862 clt_path->for_new_clt = 1; 2863 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list); 2864 2865 err = init_path(clt_path); 2866 if (err) { 2867 list_del_rcu(&clt_path->s.entry); 2868 rtrs_clt_close_conns(clt_path, true); 2869 free_percpu(clt_path->stats->pcpu_stats); 2870 kfree(clt_path->stats); 2871 free_path(clt_path); 2872 goto close_all_path; 2873 } 2874 2875 err = rtrs_clt_create_path_files(clt_path); 2876 if (err) { 2877 list_del_rcu(&clt_path->s.entry); 2878 rtrs_clt_close_conns(clt_path, true); 2879 free_percpu(clt_path->stats->pcpu_stats); 2880 kfree(clt_path->stats); 2881 free_path(clt_path); 2882 goto close_all_path; 2883 } 2884 } 2885 err = alloc_permits(clt); 2886 if (err) 2887 goto close_all_path; 2888 2889 return clt; 2890 2891 close_all_path: 2892 list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) { 2893 rtrs_clt_destroy_path_files(clt_path, NULL); 2894 rtrs_clt_close_conns(clt_path, true); 2895 kobject_put(&clt_path->kobj); 2896 } 2897 rtrs_clt_destroy_sysfs_root(clt); 2898 free_clt(clt); 2899 2900 out: 2901 return ERR_PTR(err); 2902 } 2903 EXPORT_SYMBOL(rtrs_clt_open); 2904 2905 /** 2906 * rtrs_clt_close() - Close a path 2907 * @clt: Session handle. Session is freed upon return. 2908 */ 2909 void rtrs_clt_close(struct rtrs_clt_sess *clt) 2910 { 2911 struct rtrs_clt_path *clt_path, *tmp; 2912 2913 /* Firstly forbid sysfs access */ 2914 rtrs_clt_destroy_sysfs_root(clt); 2915 2916 /* Now it is safe to iterate over all paths without locks */ 2917 list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) { 2918 rtrs_clt_close_conns(clt_path, true); 2919 rtrs_clt_destroy_path_files(clt_path, NULL); 2920 kobject_put(&clt_path->kobj); 2921 } 2922 free_permits(clt); 2923 free_clt(clt); 2924 } 2925 EXPORT_SYMBOL(rtrs_clt_close); 2926 2927 int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_path *clt_path) 2928 { 2929 enum rtrs_clt_state old_state; 2930 int err = -EBUSY; 2931 bool changed; 2932 2933 changed = rtrs_clt_change_state_get_old(clt_path, 2934 RTRS_CLT_RECONNECTING, 2935 &old_state); 2936 if (changed) { 2937 clt_path->reconnect_attempts = 0; 2938 rtrs_clt_stop_and_destroy_conns(clt_path); 2939 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork, 0); 2940 } 2941 if (changed || old_state == RTRS_CLT_RECONNECTING) { 2942 /* 2943 * flush_delayed_work() queues pending work for immediate 2944 * execution, so do the flush if we have queued something 2945 * right now or work is pending. 2946 */ 2947 flush_delayed_work(&clt_path->reconnect_dwork); 2948 err = (READ_ONCE(clt_path->state) == 2949 RTRS_CLT_CONNECTED ? 0 : -ENOTCONN); 2950 } 2951 2952 return err; 2953 } 2954 2955 int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_path *clt_path, 2956 const struct attribute *sysfs_self) 2957 { 2958 enum rtrs_clt_state old_state; 2959 bool changed; 2960 2961 /* 2962 * Continue stopping path till state was changed to DEAD or 2963 * state was observed as DEAD: 2964 * 1. State was changed to DEAD - we were fast and nobody 2965 * invoked rtrs_clt_reconnect(), which can again start 2966 * reconnecting. 2967 * 2. State was observed as DEAD - we have someone in parallel 2968 * removing the path. 2969 */ 2970 do { 2971 rtrs_clt_close_conns(clt_path, true); 2972 changed = rtrs_clt_change_state_get_old(clt_path, 2973 RTRS_CLT_DEAD, 2974 &old_state); 2975 } while (!changed && old_state != RTRS_CLT_DEAD); 2976 2977 if (changed) { 2978 rtrs_clt_remove_path_from_arr(clt_path); 2979 rtrs_clt_destroy_path_files(clt_path, sysfs_self); 2980 kobject_put(&clt_path->kobj); 2981 } 2982 2983 return 0; 2984 } 2985 2986 void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt_sess *clt, int value) 2987 { 2988 clt->max_reconnect_attempts = (unsigned int)value; 2989 } 2990 2991 int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt_sess *clt) 2992 { 2993 return (int)clt->max_reconnect_attempts; 2994 } 2995 2996 /** 2997 * rtrs_clt_request() - Request data transfer to/from server via RDMA. 2998 * 2999 * @dir: READ/WRITE 3000 * @ops: callback function to be called as confirmation, and the pointer. 3001 * @clt: Session 3002 * @permit: Preallocated permit 3003 * @vec: Message that is sent to server together with the request. 3004 * Sum of len of all @vec elements limited to <= IO_MSG_SIZE. 3005 * Since the msg is copied internally it can be allocated on stack. 3006 * @nr: Number of elements in @vec. 3007 * @data_len: length of data sent to/from server 3008 * @sg: Pages to be sent/received to/from server. 3009 * @sg_cnt: Number of elements in the @sg 3010 * 3011 * Return: 3012 * 0: Success 3013 * <0: Error 3014 * 3015 * On dir=READ rtrs client will request a data transfer from Server to client. 3016 * The data that the server will respond with will be stored in @sg when 3017 * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event. 3018 * On dir=WRITE rtrs client will rdma write data in sg to server side. 3019 */ 3020 int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops, 3021 struct rtrs_clt_sess *clt, struct rtrs_permit *permit, 3022 const struct kvec *vec, size_t nr, size_t data_len, 3023 struct scatterlist *sg, unsigned int sg_cnt) 3024 { 3025 struct rtrs_clt_io_req *req; 3026 struct rtrs_clt_path *clt_path; 3027 3028 enum dma_data_direction dma_dir; 3029 int err = -ECONNABORTED, i; 3030 size_t usr_len, hdr_len; 3031 struct path_it it; 3032 3033 /* Get kvec length */ 3034 for (i = 0, usr_len = 0; i < nr; i++) 3035 usr_len += vec[i].iov_len; 3036 3037 if (dir == READ) { 3038 hdr_len = sizeof(struct rtrs_msg_rdma_read) + 3039 sg_cnt * sizeof(struct rtrs_sg_desc); 3040 dma_dir = DMA_FROM_DEVICE; 3041 } else { 3042 hdr_len = sizeof(struct rtrs_msg_rdma_write); 3043 dma_dir = DMA_TO_DEVICE; 3044 } 3045 3046 rcu_read_lock(); 3047 for (path_it_init(&it, clt); 3048 (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) { 3049 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) 3050 continue; 3051 3052 if (usr_len + hdr_len > clt_path->max_hdr_size) { 3053 rtrs_wrn_rl(clt_path->clt, 3054 "%s request failed, user message size is %zu and header length %zu, but max size is %u\n", 3055 dir == READ ? "Read" : "Write", 3056 usr_len, hdr_len, clt_path->max_hdr_size); 3057 err = -EMSGSIZE; 3058 break; 3059 } 3060 req = rtrs_clt_get_req(clt_path, ops->conf_fn, permit, ops->priv, 3061 vec, usr_len, sg, sg_cnt, data_len, 3062 dma_dir); 3063 if (dir == READ) 3064 err = rtrs_clt_read_req(req); 3065 else 3066 err = rtrs_clt_write_req(req); 3067 if (err) { 3068 req->in_use = false; 3069 continue; 3070 } 3071 /* Success path */ 3072 break; 3073 } 3074 path_it_deinit(&it); 3075 rcu_read_unlock(); 3076 3077 return err; 3078 } 3079 EXPORT_SYMBOL(rtrs_clt_request); 3080 3081 int rtrs_clt_rdma_cq_direct(struct rtrs_clt_sess *clt, unsigned int index) 3082 { 3083 /* If no path, return -1 for block layer not to try again */ 3084 int cnt = -1; 3085 struct rtrs_con *con; 3086 struct rtrs_clt_path *clt_path; 3087 struct path_it it; 3088 3089 rcu_read_lock(); 3090 for (path_it_init(&it, clt); 3091 (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) { 3092 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) 3093 continue; 3094 3095 con = clt_path->s.con[index + 1]; 3096 cnt = ib_process_cq_direct(con->cq, -1); 3097 if (cnt) 3098 break; 3099 } 3100 path_it_deinit(&it); 3101 rcu_read_unlock(); 3102 3103 return cnt; 3104 } 3105 EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct); 3106 3107 /** 3108 * rtrs_clt_query() - queries RTRS session attributes 3109 *@clt: session pointer 3110 *@attr: query results for session attributes. 3111 * Returns: 3112 * 0 on success 3113 * -ECOMM no connection to the server 3114 */ 3115 int rtrs_clt_query(struct rtrs_clt_sess *clt, struct rtrs_attrs *attr) 3116 { 3117 if (!rtrs_clt_is_connected(clt)) 3118 return -ECOMM; 3119 3120 attr->queue_depth = clt->queue_depth; 3121 attr->max_segments = clt->max_segments; 3122 /* Cap max_io_size to min of remote buffer size and the fr pages */ 3123 attr->max_io_size = min_t(int, clt->max_io_size, 3124 clt->max_segments * SZ_4K); 3125 3126 return 0; 3127 } 3128 EXPORT_SYMBOL(rtrs_clt_query); 3129 3130 int rtrs_clt_create_path_from_sysfs(struct rtrs_clt_sess *clt, 3131 struct rtrs_addr *addr) 3132 { 3133 struct rtrs_clt_path *clt_path; 3134 int err; 3135 3136 clt_path = alloc_path(clt, addr, nr_cpu_ids, 0); 3137 if (IS_ERR(clt_path)) 3138 return PTR_ERR(clt_path); 3139 3140 mutex_lock(&clt->paths_mutex); 3141 if (clt->paths_num == 0) { 3142 /* 3143 * When all the paths are removed for a session, 3144 * the addition of the first path is like a new session for 3145 * the storage server 3146 */ 3147 clt_path->for_new_clt = 1; 3148 } 3149 3150 mutex_unlock(&clt->paths_mutex); 3151 3152 /* 3153 * It is totally safe to add path in CONNECTING state: coming 3154 * IO will never grab it. Also it is very important to add 3155 * path before init, since init fires LINK_CONNECTED event. 3156 */ 3157 rtrs_clt_add_path_to_arr(clt_path); 3158 3159 err = init_path(clt_path); 3160 if (err) 3161 goto close_path; 3162 3163 err = rtrs_clt_create_path_files(clt_path); 3164 if (err) 3165 goto close_path; 3166 3167 return 0; 3168 3169 close_path: 3170 rtrs_clt_remove_path_from_arr(clt_path); 3171 rtrs_clt_close_conns(clt_path, true); 3172 free_percpu(clt_path->stats->pcpu_stats); 3173 kfree(clt_path->stats); 3174 free_path(clt_path); 3175 3176 return err; 3177 } 3178 3179 void rtrs_clt_ib_event_handler(struct ib_event_handler *handler, 3180 struct ib_event *ibevent) 3181 { 3182 struct ib_device *idev = ibevent->device; 3183 u32 port_num = ibevent->element.port_num; 3184 3185 pr_info("Handling event: %s (%d). HCA name: %s, port num: %u\n", 3186 ib_event_msg(ibevent->event), ibevent->event, idev->name, port_num); 3187 } 3188 3189 3190 static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev) 3191 { 3192 INIT_IB_EVENT_HANDLER(&dev->event_handler, dev->ib_dev, 3193 rtrs_clt_ib_event_handler); 3194 ib_register_event_handler(&dev->event_handler); 3195 3196 if (!(dev->ib_dev->attrs.device_cap_flags & 3197 IB_DEVICE_MEM_MGT_EXTENSIONS)) { 3198 pr_err("Memory registrations not supported.\n"); 3199 return -ENOTSUPP; 3200 } 3201 3202 return 0; 3203 } 3204 3205 static void rtrs_clt_ib_dev_deinit(struct rtrs_ib_dev *dev) 3206 { 3207 ib_unregister_event_handler(&dev->event_handler); 3208 } 3209 3210 3211 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = { 3212 .init = rtrs_clt_ib_dev_init, 3213 .deinit = rtrs_clt_ib_dev_deinit 3214 }; 3215 3216 static int __init rtrs_client_init(void) 3217 { 3218 int ret = 0; 3219 3220 rtrs_rdma_dev_pd_init(0, &dev_pd); 3221 ret = class_register(&rtrs_clt_dev_class); 3222 if (ret) { 3223 pr_err("Failed to create rtrs-client dev class\n"); 3224 return ret; 3225 } 3226 rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0); 3227 if (!rtrs_wq) { 3228 class_unregister(&rtrs_clt_dev_class); 3229 return -ENOMEM; 3230 } 3231 3232 return 0; 3233 } 3234 3235 static void __exit rtrs_client_exit(void) 3236 { 3237 destroy_workqueue(rtrs_wq); 3238 class_unregister(&rtrs_clt_dev_class); 3239 rtrs_rdma_dev_pd_deinit(&dev_pd); 3240 } 3241 3242 module_init(rtrs_client_init); 3243 module_exit(rtrs_client_exit); 3244