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