drivers/infiniband/ulp/rtrs/rtrs-clt.c

   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 struct class *rtrs_clt_dev_class;
  49
  50 static inline bool rtrs_clt_is_connected(const struct rtrs_clt_sess *clt)
  51 {
  52         struct rtrs_clt_path *clt_path;
  53         bool connected = false;
  54
  55         rcu_read_lock();
  56         list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry)
  57                 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED) {
  58                         connected = true;
  59                         break;
  60                 }
  61         rcu_read_unlock();
  62
  63         return connected;
  64 }
  65
  66 static struct rtrs_permit *
  67 __rtrs_get_permit(struct rtrs_clt_sess *clt, enum rtrs_clt_con_type con_type)
  68 {
  69         size_t max_depth = clt->queue_depth;
  70         struct rtrs_permit *permit;
  71         int bit;
  72
  73         /*
  74          * Adapted from null_blk get_tag(). Callers from different cpus may
  75          * grab the same bit, since find_first_zero_bit is not atomic.
  76          * But then the test_and_set_bit_lock will fail for all the
  77          * callers but one, so that they will loop again.
  78          * This way an explicit spinlock is not required.
  79          */
  80         do {
  81                 bit = find_first_zero_bit(clt->permits_map, max_depth);
  82                 if (bit >= max_depth)
  83                         return NULL;
  84         } while (test_and_set_bit_lock(bit, clt->permits_map));
  85
  86         permit = get_permit(clt, bit);
  87         WARN_ON(permit->mem_id != bit);
  88         permit->cpu_id = raw_smp_processor_id();
  89         permit->con_type = con_type;
  90
  91         return permit;
  92 }
  93
  94 static inline void __rtrs_put_permit(struct rtrs_clt_sess *clt,
  95                                       struct rtrs_permit *permit)
  96 {
  97         clear_bit_unlock(permit->mem_id, clt->permits_map);
  98 }
  99
 100 /**
 101  * rtrs_clt_get_permit() - allocates permit for future RDMA operation
 102  * @clt:        Current session
 103  * @con_type:   Type of connection to use with the permit
 104  * @can_wait:   Wait type
 105  *
 106  * Description:
 107  *    Allocates permit for the following RDMA operation.  Permit is used
 108  *    to preallocate all resources and to propagate memory pressure
 109  *    up earlier.
 110  *
 111  * Context:
 112  *    Can sleep if @wait == RTRS_PERMIT_WAIT
 113  */
 114 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt_sess *clt,
 115                                           enum rtrs_clt_con_type con_type,
 116                                           enum wait_type can_wait)
 117 {
 118         struct rtrs_permit *permit;
 119         DEFINE_WAIT(wait);
 120
 121         permit = __rtrs_get_permit(clt, con_type);
 122         if (permit || !can_wait)
 123                 return permit;
 124
 125         do {
 126                 prepare_to_wait(&clt->permits_wait, &wait,
 127                                 TASK_UNINTERRUPTIBLE);
 128                 permit = __rtrs_get_permit(clt, con_type);
 129                 if (permit)
 130                         break;
 131
 132                 io_schedule();
 133         } while (1);
 134
 135         finish_wait(&clt->permits_wait, &wait);
 136
 137         return permit;
 138 }
 139 EXPORT_SYMBOL(rtrs_clt_get_permit);
 140
 141 /**
 142  * rtrs_clt_put_permit() - puts allocated permit
 143  * @clt:        Current session
 144  * @permit:     Permit to be freed
 145  *
 146  * Context:
 147  *    Does not matter
 148  */
 149 void rtrs_clt_put_permit(struct rtrs_clt_sess *clt,
 150                          struct rtrs_permit *permit)
 151 {
 152         if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
 153                 return;
 154
 155         __rtrs_put_permit(clt, permit);
 156
 157         /*
 158          * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
 159          * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
 160          * it must have added itself to &clt->permits_wait before
 161          * __rtrs_put_permit() finished.
 162          * Hence it is safe to guard wake_up() with a waitqueue_active() test.
 163          */
 164         if (waitqueue_active(&clt->permits_wait))
 165                 wake_up(&clt->permits_wait);
 166 }
 167 EXPORT_SYMBOL(rtrs_clt_put_permit);
 168
 169 /**
 170  * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
 171  * @clt_path: client path pointer
 172  * @permit: permit for the allocation of the RDMA buffer
 173  * Note:
 174  *     IO connection starts from 1.
 175  *     0 connection is for user messages.
 176  */
 177 static
 178 struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_path *clt_path,
 179                                             struct rtrs_permit *permit)
 180 {
 181         int id = 0;
 182
 183         if (permit->con_type == RTRS_IO_CON)
 184                 id = (permit->cpu_id % (clt_path->s.irq_con_num - 1)) + 1;
 185
 186         return to_clt_con(clt_path->s.con[id]);
 187 }
 188
 189 /**
 190  * rtrs_clt_change_state() - change the session state through session state
 191  * machine.
 192  *
 193  * @clt_path: client path to change the state of.
 194  * @new_state: state to change to.
 195  *
 196  * returns true if sess's state is changed to new state, otherwise return false.
 197  *
 198  * Locks:
 199  * state_wq lock must be hold.
 200  */
 201 static bool rtrs_clt_change_state(struct rtrs_clt_path *clt_path,
 202                                      enum rtrs_clt_state new_state)
 203 {
 204         enum rtrs_clt_state old_state;
 205         bool changed = false;
 206
 207         lockdep_assert_held(&clt_path->state_wq.lock);
 208
 209         old_state = clt_path->state;
 210         switch (new_state) {
 211         case RTRS_CLT_CONNECTING:
 212                 switch (old_state) {
 213                 case RTRS_CLT_RECONNECTING:
 214                         changed = true;
 215                         fallthrough;
 216                 default:
 217                         break;
 218                 }
 219                 break;
 220         case RTRS_CLT_RECONNECTING:
 221                 switch (old_state) {
 222                 case RTRS_CLT_CONNECTED:
 223                 case RTRS_CLT_CONNECTING_ERR:
 224                 case RTRS_CLT_CLOSED:
 225                         changed = true;
 226                         fallthrough;
 227                 default:
 228                         break;
 229                 }
 230                 break;
 231         case RTRS_CLT_CONNECTED:
 232                 switch (old_state) {
 233                 case RTRS_CLT_CONNECTING:
 234                         changed = true;
 235                         fallthrough;
 236                 default:
 237                         break;
 238                 }
 239                 break;
 240         case RTRS_CLT_CONNECTING_ERR:
 241                 switch (old_state) {
 242                 case RTRS_CLT_CONNECTING:
 243                         changed = true;
 244                         fallthrough;
 245                 default:
 246                         break;
 247                 }
 248                 break;
 249         case RTRS_CLT_CLOSING:
 250                 switch (old_state) {
 251                 case RTRS_CLT_CONNECTING:
 252                 case RTRS_CLT_CONNECTING_ERR:
 253                 case RTRS_CLT_RECONNECTING:
 254                 case RTRS_CLT_CONNECTED:
 255                         changed = true;
 256                         fallthrough;
 257                 default:
 258                         break;
 259                 }
 260                 break;
 261         case RTRS_CLT_CLOSED:
 262                 switch (old_state) {
 263                 case RTRS_CLT_CLOSING:
 264                         changed = true;
 265                         fallthrough;
 266                 default:
 267                         break;
 268                 }
 269                 break;
 270         case RTRS_CLT_DEAD:
 271                 switch (old_state) {
 272                 case RTRS_CLT_CLOSED:
 273                         changed = true;
 274                         fallthrough;
 275                 default:
 276                         break;
 277                 }
 278                 break;
 279         default:
 280                 break;
 281         }
 282         if (changed) {
 283                 clt_path->state = new_state;
 284                 wake_up_locked(&clt_path->state_wq);
 285         }
 286
 287         return changed;
 288 }
 289
 290 static bool rtrs_clt_change_state_from_to(struct rtrs_clt_path *clt_path,
 291                                            enum rtrs_clt_state old_state,
 292                                            enum rtrs_clt_state new_state)
 293 {
 294         bool changed = false;
 295
 296         spin_lock_irq(&clt_path->state_wq.lock);
 297         if (clt_path->state == old_state)
 298                 changed = rtrs_clt_change_state(clt_path, new_state);
 299         spin_unlock_irq(&clt_path->state_wq.lock);
 300
 301         return changed;
 302 }
 303
 304 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path);
 305 static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
 306 {
 307         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 308
 309         trace_rtrs_rdma_error_recovery(clt_path);
 310
 311         if (rtrs_clt_change_state_from_to(clt_path,
 312                                            RTRS_CLT_CONNECTED,
 313                                            RTRS_CLT_RECONNECTING)) {
 314                 queue_work(rtrs_wq, &clt_path->err_recovery_work);
 315         } else {
 316                 /*
 317                  * Error can happen just on establishing new connection,
 318                  * so notify waiter with error state, waiter is responsible
 319                  * for cleaning the rest and reconnect if needed.
 320                  */
 321                 rtrs_clt_change_state_from_to(clt_path,
 322                                                RTRS_CLT_CONNECTING,
 323                                                RTRS_CLT_CONNECTING_ERR);
 324         }
 325 }
 326
 327 static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
 328 {
 329         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 330
 331         if (wc->status != IB_WC_SUCCESS) {
 332                 rtrs_err(con->c.path, "Failed IB_WR_REG_MR: %s\n",
 333                           ib_wc_status_msg(wc->status));
 334                 rtrs_rdma_error_recovery(con);
 335         }
 336 }
 337
 338 static struct ib_cqe fast_reg_cqe = {
 339         .done = rtrs_clt_fast_reg_done
 340 };
 341
 342 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 343                               bool notify, bool can_wait);
 344
 345 static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 346 {
 347         struct rtrs_clt_io_req *req =
 348                 container_of(wc->wr_cqe, typeof(*req), inv_cqe);
 349         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 350
 351         if (wc->status != IB_WC_SUCCESS) {
 352                 rtrs_err(con->c.path, "Failed IB_WR_LOCAL_INV: %s\n",
 353                           ib_wc_status_msg(wc->status));
 354                 rtrs_rdma_error_recovery(con);
 355         }
 356         req->need_inv = false;
 357         if (req->need_inv_comp)
 358                 complete(&req->inv_comp);
 359         else
 360                 /* Complete request from INV callback */
 361                 complete_rdma_req(req, req->inv_errno, true, false);
 362 }
 363
 364 static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
 365 {
 366         struct rtrs_clt_con *con = req->con;
 367         struct ib_send_wr wr = {
 368                 .opcode             = IB_WR_LOCAL_INV,
 369                 .wr_cqe             = &req->inv_cqe,
 370                 .send_flags         = IB_SEND_SIGNALED,
 371                 .ex.invalidate_rkey = req->mr->rkey,
 372         };
 373         req->inv_cqe.done = rtrs_clt_inv_rkey_done;
 374
 375         return ib_post_send(con->c.qp, &wr, NULL);
 376 }
 377
 378 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 379                               bool notify, bool can_wait)
 380 {
 381         struct rtrs_clt_con *con = req->con;
 382         struct rtrs_clt_path *clt_path;
 383         int err;
 384
 385         if (WARN_ON(!req->in_use))
 386                 return;
 387         if (WARN_ON(!req->con))
 388                 return;
 389         clt_path = to_clt_path(con->c.path);
 390
 391         if (req->sg_cnt) {
 392                 if (req->dir == DMA_FROM_DEVICE && req->need_inv) {
 393                         /*
 394                          * We are here to invalidate read requests
 395                          * ourselves.  In normal scenario server should
 396                          * send INV for all read requests, but
 397                          * we are here, thus two things could happen:
 398                          *
 399                          *    1.  this is failover, when errno != 0
 400                          *        and can_wait == 1,
 401                          *
 402                          *    2.  something totally bad happened and
 403                          *        server forgot to send INV, so we
 404                          *        should do that ourselves.
 405                          */
 406
 407                         if (can_wait) {
 408                                 req->need_inv_comp = true;
 409                         } else {
 410                                 /* This should be IO path, so always notify */
 411                                 WARN_ON(!notify);
 412                                 /* Save errno for INV callback */
 413                                 req->inv_errno = errno;
 414                         }
 415
 416                         refcount_inc(&req->ref);
 417                         err = rtrs_inv_rkey(req);
 418                         if (err) {
 419                                 rtrs_err(con->c.path, "Send INV WR key=%#x: %d\n",
 420                                           req->mr->rkey, err);
 421                         } else if (can_wait) {
 422                                 wait_for_completion(&req->inv_comp);
 423                         } else {
 424                                 /*
 425                                  * Something went wrong, so request will be
 426                                  * completed from INV callback.
 427                                  */
 428                                 WARN_ON_ONCE(1);
 429
 430                                 return;
 431                         }
 432                         if (!refcount_dec_and_test(&req->ref))
 433                                 return;
 434                 }
 435                 ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
 436                                 req->sg_cnt, req->dir);
 437         }
 438         if (!refcount_dec_and_test(&req->ref))
 439                 return;
 440         if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
 441                 atomic_dec(&clt_path->stats->inflight);
 442
 443         req->in_use = false;
 444         req->con = NULL;
 445
 446         if (errno) {
 447                 rtrs_err_rl(con->c.path, "IO request failed: error=%d path=%s [%s:%u] notify=%d\n",
 448                             errno, kobject_name(&clt_path->kobj), clt_path->hca_name,
 449                             clt_path->hca_port, notify);
 450         }
 451
 452         if (notify)
 453                 req->conf(req->priv, errno);
 454 }
 455
 456 static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
 457                                 struct rtrs_clt_io_req *req,
 458                                 struct rtrs_rbuf *rbuf, u32 off,
 459                                 u32 imm, struct ib_send_wr *wr)
 460 {
 461         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 462         enum ib_send_flags flags;
 463         struct ib_sge sge;
 464
 465         if (!req->sg_size) {
 466                 rtrs_wrn(con->c.path,
 467                          "Doing RDMA Write failed, no data supplied\n");
 468                 return -EINVAL;
 469         }
 470
 471         /* user data and user message in the first list element */
 472         sge.addr   = req->iu->dma_addr;
 473         sge.length = req->sg_size;
 474         sge.lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
 475
 476         /*
 477          * From time to time we have to post signalled sends,
 478          * or send queue will fill up and only QP reset can help.
 479          */
 480         flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
 481                         0 : IB_SEND_SIGNALED;
 482
 483         ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
 484                                       req->iu->dma_addr,
 485                                       req->sg_size, DMA_TO_DEVICE);
 486
 487         return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
 488                                             rbuf->rkey, rbuf->addr + off,
 489                                             imm, flags, wr, NULL);
 490 }
 491
 492 static void process_io_rsp(struct rtrs_clt_path *clt_path, u32 msg_id,
 493                            s16 errno, bool w_inval)
 494 {
 495         struct rtrs_clt_io_req *req;
 496
 497         if (WARN_ON(msg_id >= clt_path->queue_depth))
 498                 return;
 499
 500         req = &clt_path->reqs[msg_id];
 501         /* Drop need_inv if server responded with send with invalidation */
 502         req->need_inv &= !w_inval;
 503         complete_rdma_req(req, errno, true, false);
 504 }
 505
 506 static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
 507 {
 508         struct rtrs_iu *iu;
 509         int err;
 510         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 511
 512         WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
 513         iu = container_of(wc->wr_cqe, struct rtrs_iu,
 514                           cqe);
 515         err = rtrs_iu_post_recv(&con->c, iu);
 516         if (err) {
 517                 rtrs_err(con->c.path, "post iu failed %d\n", err);
 518                 rtrs_rdma_error_recovery(con);
 519         }
 520 }
 521
 522 static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
 523 {
 524         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 525         struct rtrs_msg_rkey_rsp *msg;
 526         u32 imm_type, imm_payload;
 527         bool w_inval = false;
 528         struct rtrs_iu *iu;
 529         u32 buf_id;
 530         int err;
 531
 532         WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
 533
 534         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
 535
 536         if (wc->byte_len < sizeof(*msg)) {
 537                 rtrs_err(con->c.path, "rkey response is malformed: size %d\n",
 538                           wc->byte_len);
 539                 goto out;
 540         }
 541         ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
 542                                    iu->size, DMA_FROM_DEVICE);
 543         msg = iu->buf;
 544         if (le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP) {
 545                 rtrs_err(clt_path->clt,
 546                           "rkey response is malformed: type %d\n",
 547                           le16_to_cpu(msg->type));
 548                 goto out;
 549         }
 550         buf_id = le16_to_cpu(msg->buf_id);
 551         if (WARN_ON(buf_id >= clt_path->queue_depth))
 552                 goto out;
 553
 554         rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
 555         if (imm_type == RTRS_IO_RSP_IMM ||
 556             imm_type == RTRS_IO_RSP_W_INV_IMM) {
 557                 u32 msg_id;
 558
 559                 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 560                 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 561
 562                 if (WARN_ON(buf_id != msg_id))
 563                         goto out;
 564                 clt_path->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
 565                 process_io_rsp(clt_path, msg_id, err, w_inval);
 566         }
 567         ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, iu->dma_addr,
 568                                       iu->size, DMA_FROM_DEVICE);
 569         return rtrs_clt_recv_done(con, wc);
 570 out:
 571         rtrs_rdma_error_recovery(con);
 572 }
 573
 574 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
 575
 576 static struct ib_cqe io_comp_cqe = {
 577         .done = rtrs_clt_rdma_done
 578 };
 579
 580 /*
 581  * Post x2 empty WRs: first is for this RDMA with IMM,
 582  * second is for RECV with INV, which happened earlier.
 583  */
 584 static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
 585 {
 586         struct ib_recv_wr wr_arr[2], *wr;
 587         int i;
 588
 589         memset(wr_arr, 0, sizeof(wr_arr));
 590         for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
 591                 wr = &wr_arr[i];
 592                 wr->wr_cqe  = cqe;
 593                 if (i)
 594                         /* Chain backwards */
 595                         wr->next = &wr_arr[i - 1];
 596         }
 597
 598         return ib_post_recv(con->qp, wr, NULL);
 599 }
 600
 601 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
 602 {
 603         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 604         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 605         u32 imm_type, imm_payload;
 606         bool w_inval = false;
 607         int err;
 608
 609         if (wc->status != IB_WC_SUCCESS) {
 610                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 611                         rtrs_err(clt_path->clt, "RDMA failed: %s\n",
 612                                   ib_wc_status_msg(wc->status));
 613                         rtrs_rdma_error_recovery(con);
 614                 }
 615                 return;
 616         }
 617         rtrs_clt_update_wc_stats(con);
 618
 619         switch (wc->opcode) {
 620         case IB_WC_RECV_RDMA_WITH_IMM:
 621                 /*
 622                  * post_recv() RDMA write completions of IO reqs (read/write)
 623                  * and hb
 624                  */
 625                 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
 626                         return;
 627                 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
 628                                &imm_type, &imm_payload);
 629                 if (imm_type == RTRS_IO_RSP_IMM ||
 630                     imm_type == RTRS_IO_RSP_W_INV_IMM) {
 631                         u32 msg_id;
 632
 633                         w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 634                         rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 635
 636                         process_io_rsp(clt_path, msg_id, err, w_inval);
 637                 } else if (imm_type == RTRS_HB_MSG_IMM) {
 638                         WARN_ON(con->c.cid);
 639                         rtrs_send_hb_ack(&clt_path->s);
 640                         if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
 641                                 return  rtrs_clt_recv_done(con, wc);
 642                 } else if (imm_type == RTRS_HB_ACK_IMM) {
 643                         WARN_ON(con->c.cid);
 644                         clt_path->s.hb_missed_cnt = 0;
 645                         clt_path->s.hb_cur_latency =
 646                                 ktime_sub(ktime_get(), clt_path->s.hb_last_sent);
 647                         if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
 648                                 return  rtrs_clt_recv_done(con, wc);
 649                 } else {
 650                         rtrs_wrn(con->c.path, "Unknown IMM type %u\n",
 651                                   imm_type);
 652                 }
 653                 if (w_inval)
 654                         /*
 655                          * Post x2 empty WRs: first is for this RDMA with IMM,
 656                          * second is for RECV with INV, which happened earlier.
 657                          */
 658                         err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
 659                 else
 660                         err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 661                 if (err) {
 662                         rtrs_err(con->c.path, "rtrs_post_recv_empty(): %d\n",
 663                                   err);
 664                         rtrs_rdma_error_recovery(con);
 665                 }
 666                 break;
 667         case IB_WC_RECV:
 668                 /*
 669                  * Key invalidations from server side
 670                  */
 671                 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
 672                           wc->wc_flags & IB_WC_WITH_IMM));
 673                 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
 674                 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
 675                         if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
 676                                 return  rtrs_clt_recv_done(con, wc);
 677
 678                         return  rtrs_clt_rkey_rsp_done(con, wc);
 679                 }
 680                 break;
 681         case IB_WC_RDMA_WRITE:
 682                 /*
 683                  * post_send() RDMA write completions of IO reqs (read/write)
 684                  * and hb.
 685                  */
 686                 break;
 687
 688         default:
 689                 rtrs_wrn(clt_path->clt, "Unexpected WC type: %d\n", wc->opcode);
 690                 return;
 691         }
 692 }
 693
 694 static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
 695 {
 696         int err, i;
 697         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 698
 699         for (i = 0; i < q_size; i++) {
 700                 if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
 701                         struct rtrs_iu *iu = &con->rsp_ius[i];
 702
 703                         err = rtrs_iu_post_recv(&con->c, iu);
 704                 } else {
 705                         err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 706                 }
 707                 if (err)
 708                         return err;
 709         }
 710
 711         return 0;
 712 }
 713
 714 static int post_recv_path(struct rtrs_clt_path *clt_path)
 715 {
 716         size_t q_size = 0;
 717         int err, cid;
 718
 719         for (cid = 0; cid < clt_path->s.con_num; cid++) {
 720                 if (cid == 0)
 721                         q_size = SERVICE_CON_QUEUE_DEPTH;
 722                 else
 723                         q_size = clt_path->queue_depth;
 724
 725                 /*
 726                  * x2 for RDMA read responses + FR key invalidations,
 727                  * RDMA writes do not require any FR registrations.
 728                  */
 729                 q_size *= 2;
 730
 731                 err = post_recv_io(to_clt_con(clt_path->s.con[cid]), q_size);
 732                 if (err) {
 733                         rtrs_err(clt_path->clt, "post_recv_io(), err: %d\n",
 734                                  err);
 735                         return err;
 736                 }
 737         }
 738
 739         return 0;
 740 }
 741
 742 struct path_it {
 743         int i;
 744         struct list_head skip_list;
 745         struct rtrs_clt_sess *clt;
 746         struct rtrs_clt_path *(*next_path)(struct path_it *it);
 747 };
 748
 749 /*
 750  * rtrs_clt_get_next_path_or_null - get clt path from the list or return NULL
 751  * @head:       the head for the list.
 752  * @clt_path:   The element to take the next clt_path from.
 753  *
 754  * Next clt path returned in round-robin fashion, i.e. head will be skipped,
 755  * but if list is observed as empty, NULL will be returned.
 756  *
 757  * This function may safely run concurrently with the _rcu list-mutation
 758  * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
 759  */
 760 static inline struct rtrs_clt_path *
 761 rtrs_clt_get_next_path_or_null(struct list_head *head, struct rtrs_clt_path *clt_path)
 762 {
 763         return list_next_or_null_rcu(head, &clt_path->s.entry, typeof(*clt_path), s.entry) ?:
 764                                      list_next_or_null_rcu(head,
 765                                                            READ_ONCE((&clt_path->s.entry)->next),
 766                                                            typeof(*clt_path), s.entry);
 767 }
 768
 769 /**
 770  * get_next_path_rr() - Returns path in round-robin fashion.
 771  * @it: the path pointer
 772  *
 773  * Related to @MP_POLICY_RR
 774  *
 775  * Locks:
 776  *    rcu_read_lock() must be hold.
 777  */
 778 static struct rtrs_clt_path *get_next_path_rr(struct path_it *it)
 779 {
 780         struct rtrs_clt_path __rcu **ppcpu_path;
 781         struct rtrs_clt_path *path;
 782         struct rtrs_clt_sess *clt;
 783
 784         clt = it->clt;
 785
 786         /*
 787          * Here we use two RCU objects: @paths_list and @pcpu_path
 788          * pointer.  See rtrs_clt_remove_path_from_arr() for details
 789          * how that is handled.
 790          */
 791
 792         ppcpu_path = this_cpu_ptr(clt->pcpu_path);
 793         path = rcu_dereference(*ppcpu_path);
 794         if (!path)
 795                 path = list_first_or_null_rcu(&clt->paths_list,
 796                                               typeof(*path), s.entry);
 797         else
 798                 path = rtrs_clt_get_next_path_or_null(&clt->paths_list, path);
 799
 800         rcu_assign_pointer(*ppcpu_path, path);
 801
 802         return path;
 803 }
 804
 805 /**
 806  * get_next_path_min_inflight() - Returns path with minimal inflight count.
 807  * @it: the path pointer
 808  *
 809  * Related to @MP_POLICY_MIN_INFLIGHT
 810  *
 811  * Locks:
 812  *    rcu_read_lock() must be hold.
 813  */
 814 static struct rtrs_clt_path *get_next_path_min_inflight(struct path_it *it)
 815 {
 816         struct rtrs_clt_path *min_path = NULL;
 817         struct rtrs_clt_sess *clt = it->clt;
 818         struct rtrs_clt_path *clt_path;
 819         int min_inflight = INT_MAX;
 820         int inflight;
 821
 822         list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
 823                 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
 824                         continue;
 825
 826                 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
 827                         continue;
 828
 829                 inflight = atomic_read(&clt_path->stats->inflight);
 830
 831                 if (inflight < min_inflight) {
 832                         min_inflight = inflight;
 833                         min_path = clt_path;
 834                 }
 835         }
 836
 837         /*
 838          * add the path to the skip list, so that next time we can get
 839          * a different one
 840          */
 841         if (min_path)
 842                 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
 843
 844         return min_path;
 845 }
 846
 847 /**
 848  * get_next_path_min_latency() - Returns path with minimal latency.
 849  * @it: the path pointer
 850  *
 851  * Return: a path with the lowest latency or NULL if all paths are tried
 852  *
 853  * Locks:
 854  *    rcu_read_lock() must be hold.
 855  *
 856  * Related to @MP_POLICY_MIN_LATENCY
 857  *
 858  * This DOES skip an already-tried path.
 859  * There is a skip-list to skip a path if the path has tried but failed.
 860  * It will try the minimum latency path and then the second minimum latency
 861  * path and so on. Finally it will return NULL if all paths are tried.
 862  * Therefore the caller MUST check the returned
 863  * path is NULL and trigger the IO error.
 864  */
 865 static struct rtrs_clt_path *get_next_path_min_latency(struct path_it *it)
 866 {
 867         struct rtrs_clt_path *min_path = NULL;
 868         struct rtrs_clt_sess *clt = it->clt;
 869         struct rtrs_clt_path *clt_path;
 870         ktime_t min_latency = KTIME_MAX;
 871         ktime_t latency;
 872
 873         list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
 874                 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
 875                         continue;
 876
 877                 if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
 878                         continue;
 879
 880                 latency = clt_path->s.hb_cur_latency;
 881
 882                 if (latency < min_latency) {
 883                         min_latency = latency;
 884                         min_path = clt_path;
 885                 }
 886         }
 887
 888         /*
 889          * add the path to the skip list, so that next time we can get
 890          * a different one
 891          */
 892         if (min_path)
 893                 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
 894
 895         return min_path;
 896 }
 897
 898 static inline void path_it_init(struct path_it *it, struct rtrs_clt_sess *clt)
 899 {
 900         INIT_LIST_HEAD(&it->skip_list);
 901         it->clt = clt;
 902         it->i = 0;
 903
 904         if (clt->mp_policy == MP_POLICY_RR)
 905                 it->next_path = get_next_path_rr;
 906         else if (clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
 907                 it->next_path = get_next_path_min_inflight;
 908         else
 909                 it->next_path = get_next_path_min_latency;
 910 }
 911
 912 static inline void path_it_deinit(struct path_it *it)
 913 {
 914         struct list_head *skip, *tmp;
 915         /*
 916          * The skip_list is used only for the MIN_INFLIGHT and MIN_LATENCY policies.
 917          * We need to remove paths from it, so that next IO can insert
 918          * paths (->mp_skip_entry) into a skip_list again.
 919          */
 920         list_for_each_safe(skip, tmp, &it->skip_list)
 921                 list_del_init(skip);
 922 }
 923
 924 /**
 925  * rtrs_clt_init_req() - Initialize an rtrs_clt_io_req holding information
 926  * about an inflight IO.
 927  * The user buffer holding user control message (not data) is copied into
 928  * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
 929  * also hold the control message of rtrs.
 930  * @req: an io request holding information about IO.
 931  * @clt_path: client path
 932  * @conf: conformation callback function to notify upper layer.
 933  * @permit: permit for allocation of RDMA remote buffer
 934  * @priv: private pointer
 935  * @vec: kernel vector containing control message
 936  * @usr_len: length of the user message
 937  * @sg: scater list for IO data
 938  * @sg_cnt: number of scater list entries
 939  * @data_len: length of the IO data
 940  * @dir: direction of the IO.
 941  */
 942 static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
 943                               struct rtrs_clt_path *clt_path,
 944                               void (*conf)(void *priv, int errno),
 945                               struct rtrs_permit *permit, void *priv,
 946                               const struct kvec *vec, size_t usr_len,
 947                               struct scatterlist *sg, size_t sg_cnt,
 948                               size_t data_len, int dir)
 949 {
 950         struct iov_iter iter;
 951         size_t len;
 952
 953         req->permit = permit;
 954         req->in_use = true;
 955         req->usr_len = usr_len;
 956         req->data_len = data_len;
 957         req->sglist = sg;
 958         req->sg_cnt = sg_cnt;
 959         req->priv = priv;
 960         req->dir = dir;
 961         req->con = rtrs_permit_to_clt_con(clt_path, permit);
 962         req->conf = conf;
 963         req->need_inv = false;
 964         req->need_inv_comp = false;
 965         req->inv_errno = 0;
 966         refcount_set(&req->ref, 1);
 967         req->mp_policy = clt_path->clt->mp_policy;
 968
 969         iov_iter_kvec(&iter, READ, vec, 1, usr_len);
 970         len = _copy_from_iter(req->iu->buf, usr_len, &iter);
 971         WARN_ON(len != usr_len);
 972
 973         reinit_completion(&req->inv_comp);
 974 }
 975
 976 static struct rtrs_clt_io_req *
 977 rtrs_clt_get_req(struct rtrs_clt_path *clt_path,
 978                  void (*conf)(void *priv, int errno),
 979                  struct rtrs_permit *permit, void *priv,
 980                  const struct kvec *vec, size_t usr_len,
 981                  struct scatterlist *sg, size_t sg_cnt,
 982                  size_t data_len, int dir)
 983 {
 984         struct rtrs_clt_io_req *req;
 985
 986         req = &clt_path->reqs[permit->mem_id];
 987         rtrs_clt_init_req(req, clt_path, conf, permit, priv, vec, usr_len,
 988                            sg, sg_cnt, data_len, dir);
 989         return req;
 990 }
 991
 992 static struct rtrs_clt_io_req *
 993 rtrs_clt_get_copy_req(struct rtrs_clt_path *alive_path,
 994                        struct rtrs_clt_io_req *fail_req)
 995 {
 996         struct rtrs_clt_io_req *req;
 997         struct kvec vec = {
 998                 .iov_base = fail_req->iu->buf,
 999                 .iov_len  = fail_req->usr_len
1000         };
1001
1002         req = &alive_path->reqs[fail_req->permit->mem_id];
1003         rtrs_clt_init_req(req, alive_path, fail_req->conf, fail_req->permit,
1004                            fail_req->priv, &vec, fail_req->usr_len,
1005                            fail_req->sglist, fail_req->sg_cnt,
1006                            fail_req->data_len, fail_req->dir);
1007         return req;
1008 }
1009
1010 static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
1011                                    struct rtrs_clt_io_req *req,
1012                                    struct rtrs_rbuf *rbuf, bool fr_en,
1013                                    u32 count, u32 size, u32 imm,
1014                                    struct ib_send_wr *wr,
1015                                    struct ib_send_wr *tail)
1016 {
1017         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1018         struct ib_sge *sge = req->sge;
1019         enum ib_send_flags flags;
1020         struct scatterlist *sg;
1021         size_t num_sge;
1022         int i;
1023         struct ib_send_wr *ptail = NULL;
1024
1025         if (fr_en) {
1026                 i = 0;
1027                 sge[i].addr   = req->mr->iova;
1028                 sge[i].length = req->mr->length;
1029                 sge[i].lkey   = req->mr->lkey;
1030                 i++;
1031                 num_sge = 2;
1032                 ptail = tail;
1033         } else {
1034                 for_each_sg(req->sglist, sg, count, i) {
1035                         sge[i].addr   = sg_dma_address(sg);
1036                         sge[i].length = sg_dma_len(sg);
1037                         sge[i].lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
1038                 }
1039                 num_sge = 1 + count;
1040         }
1041         sge[i].addr   = req->iu->dma_addr;
1042         sge[i].length = size;
1043         sge[i].lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
1044
1045         /*
1046          * From time to time we have to post signalled sends,
1047          * or send queue will fill up and only QP reset can help.
1048          */
1049         flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
1050                         0 : IB_SEND_SIGNALED;
1051
1052         ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
1053                                       req->iu->dma_addr,
1054                                       size, DMA_TO_DEVICE);
1055
1056         return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
1057                                             rbuf->rkey, rbuf->addr, imm,
1058                                             flags, wr, ptail);
1059 }
1060
1061 static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1062 {
1063         int nr;
1064
1065         /* Align the MR to a 4K page size to match the block virt boundary */
1066         nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1067         if (nr < 0)
1068                 return nr;
1069         if (nr < req->sg_cnt)
1070                 return -EINVAL;
1071         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1072
1073         return nr;
1074 }
1075
1076 static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
1077 {
1078         struct rtrs_clt_con *con = req->con;
1079         struct rtrs_path *s = con->c.path;
1080         struct rtrs_clt_path *clt_path = to_clt_path(s);
1081         struct rtrs_msg_rdma_write *msg;
1082
1083         struct rtrs_rbuf *rbuf;
1084         int ret, count = 0;
1085         u32 imm, buf_id;
1086         struct ib_reg_wr rwr;
1087         struct ib_send_wr inv_wr;
1088         struct ib_send_wr *wr = NULL;
1089         bool fr_en = false;
1090
1091         const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1092
1093         if (tsize > clt_path->chunk_size) {
1094                 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
1095                           tsize, clt_path->chunk_size);
1096                 return -EMSGSIZE;
1097         }
1098         if (req->sg_cnt) {
1099                 count = ib_dma_map_sg(clt_path->s.dev->ib_dev, req->sglist,
1100                                       req->sg_cnt, req->dir);
1101                 if (!count) {
1102                         rtrs_wrn(s, "Write request failed, map failed\n");
1103                         return -EINVAL;
1104                 }
1105         }
1106         /* put rtrs msg after sg and user message */
1107         msg = req->iu->buf + req->usr_len;
1108         msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1109         msg->usr_len = cpu_to_le16(req->usr_len);
1110
1111         /* rtrs message on server side will be after user data and message */
1112         imm = req->permit->mem_off + req->data_len + req->usr_len;
1113         imm = rtrs_to_io_req_imm(imm);
1114         buf_id = req->permit->mem_id;
1115         req->sg_size = tsize;
1116         rbuf = &clt_path->rbufs[buf_id];
1117
1118         if (count) {
1119                 ret = rtrs_map_sg_fr(req, count);
1120                 if (ret < 0) {
1121                         rtrs_err_rl(s,
1122                                     "Write request failed, failed to map fast reg. data, err: %d\n",
1123                                     ret);
1124                         ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1125                                         req->sg_cnt, req->dir);
1126                         return ret;
1127                 }
1128                 inv_wr = (struct ib_send_wr) {
1129                         .opcode             = IB_WR_LOCAL_INV,
1130                         .wr_cqe             = &req->inv_cqe,
1131                         .send_flags         = IB_SEND_SIGNALED,
1132                         .ex.invalidate_rkey = req->mr->rkey,
1133                 };
1134                 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
1135                 rwr = (struct ib_reg_wr) {
1136                         .wr.opcode = IB_WR_REG_MR,
1137                         .wr.wr_cqe = &fast_reg_cqe,
1138                         .mr = req->mr,
1139                         .key = req->mr->rkey,
1140                         .access = (IB_ACCESS_LOCAL_WRITE),
1141                 };
1142                 wr = &rwr.wr;
1143                 fr_en = true;
1144                 refcount_inc(&req->ref);
1145         }
1146         /*
1147          * Update stats now, after request is successfully sent it is not
1148          * safe anymore to touch it.
1149          */
1150         rtrs_clt_update_all_stats(req, WRITE);
1151
1152         ret = rtrs_post_rdma_write_sg(req->con, req, rbuf, fr_en, count,
1153                                       req->usr_len + sizeof(*msg),
1154                                       imm, wr, &inv_wr);
1155         if (ret) {
1156                 rtrs_err_rl(s,
1157                             "Write request failed: error=%d path=%s [%s:%u]\n",
1158                             ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1159                             clt_path->hca_port);
1160                 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1161                         atomic_dec(&clt_path->stats->inflight);
1162                 if (req->sg_cnt)
1163                         ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1164                                         req->sg_cnt, req->dir);
1165         }
1166
1167         return ret;
1168 }
1169
1170 static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1171 {
1172         struct rtrs_clt_con *con = req->con;
1173         struct rtrs_path *s = con->c.path;
1174         struct rtrs_clt_path *clt_path = to_clt_path(s);
1175         struct rtrs_msg_rdma_read *msg;
1176         struct rtrs_ib_dev *dev = clt_path->s.dev;
1177
1178         struct ib_reg_wr rwr;
1179         struct ib_send_wr *wr = NULL;
1180
1181         int ret, count = 0;
1182         u32 imm, buf_id;
1183
1184         const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1185
1186         if (tsize > clt_path->chunk_size) {
1187                 rtrs_wrn(s,
1188                           "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1189                           tsize, clt_path->chunk_size);
1190                 return -EMSGSIZE;
1191         }
1192
1193         if (req->sg_cnt) {
1194                 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1195                                       req->dir);
1196                 if (!count) {
1197                         rtrs_wrn(s,
1198                                   "Read request failed, dma map failed\n");
1199                         return -EINVAL;
1200                 }
1201         }
1202         /* put our message into req->buf after user message*/
1203         msg = req->iu->buf + req->usr_len;
1204         msg->type = cpu_to_le16(RTRS_MSG_READ);
1205         msg->usr_len = cpu_to_le16(req->usr_len);
1206
1207         if (count) {
1208                 ret = rtrs_map_sg_fr(req, count);
1209                 if (ret < 0) {
1210                         rtrs_err_rl(s,
1211                                      "Read request failed, failed to map  fast reg. data, err: %d\n",
1212                                      ret);
1213                         ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1214                                         req->dir);
1215                         return ret;
1216                 }
1217                 rwr = (struct ib_reg_wr) {
1218                         .wr.opcode = IB_WR_REG_MR,
1219                         .wr.wr_cqe = &fast_reg_cqe,
1220                         .mr = req->mr,
1221                         .key = req->mr->rkey,
1222                         .access = (IB_ACCESS_LOCAL_WRITE |
1223                                    IB_ACCESS_REMOTE_WRITE),
1224                 };
1225                 wr = &rwr.wr;
1226
1227                 msg->sg_cnt = cpu_to_le16(1);
1228                 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1229
1230                 msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1231                 msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1232                 msg->desc[0].len = cpu_to_le32(req->mr->length);
1233
1234                 /* Further invalidation is required */
1235                 req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1236
1237         } else {
1238                 msg->sg_cnt = 0;
1239                 msg->flags = 0;
1240         }
1241         /*
1242          * rtrs message will be after the space reserved for disk data and
1243          * user message
1244          */
1245         imm = req->permit->mem_off + req->data_len + req->usr_len;
1246         imm = rtrs_to_io_req_imm(imm);
1247         buf_id = req->permit->mem_id;
1248
1249         req->sg_size  = sizeof(*msg);
1250         req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1251         req->sg_size += req->usr_len;
1252
1253         /*
1254          * Update stats now, after request is successfully sent it is not
1255          * safe anymore to touch it.
1256          */
1257         rtrs_clt_update_all_stats(req, READ);
1258
1259         ret = rtrs_post_send_rdma(req->con, req, &clt_path->rbufs[buf_id],
1260                                    req->data_len, imm, wr);
1261         if (ret) {
1262                 rtrs_err_rl(s,
1263                             "Read request failed: error=%d path=%s [%s:%u]\n",
1264                             ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1265                             clt_path->hca_port);
1266                 if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1267                         atomic_dec(&clt_path->stats->inflight);
1268                 req->need_inv = false;
1269                 if (req->sg_cnt)
1270                         ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1271                                         req->sg_cnt, req->dir);
1272         }
1273
1274         return ret;
1275 }
1276
1277 /**
1278  * rtrs_clt_failover_req() - Try to find an active path for a failed request
1279  * @clt: clt context
1280  * @fail_req: a failed io request.
1281  */
1282 static int rtrs_clt_failover_req(struct rtrs_clt_sess *clt,
1283                                  struct rtrs_clt_io_req *fail_req)
1284 {
1285         struct rtrs_clt_path *alive_path;
1286         struct rtrs_clt_io_req *req;
1287         int err = -ECONNABORTED;
1288         struct path_it it;
1289
1290         rcu_read_lock();
1291         for (path_it_init(&it, clt);
1292              (alive_path = it.next_path(&it)) && it.i < it.clt->paths_num;
1293              it.i++) {
1294                 if (READ_ONCE(alive_path->state) != RTRS_CLT_CONNECTED)
1295                         continue;
1296                 req = rtrs_clt_get_copy_req(alive_path, fail_req);
1297                 if (req->dir == DMA_TO_DEVICE)
1298                         err = rtrs_clt_write_req(req);
1299                 else
1300                         err = rtrs_clt_read_req(req);
1301                 if (err) {
1302                         req->in_use = false;
1303                         continue;
1304                 }
1305                 /* Success path */
1306                 rtrs_clt_inc_failover_cnt(alive_path->stats);
1307                 break;
1308         }
1309         path_it_deinit(&it);
1310         rcu_read_unlock();
1311
1312         return err;
1313 }
1314
1315 static void fail_all_outstanding_reqs(struct rtrs_clt_path *clt_path)
1316 {
1317         struct rtrs_clt_sess *clt = clt_path->clt;
1318         struct rtrs_clt_io_req *req;
1319         int i, err;
1320
1321         if (!clt_path->reqs)
1322                 return;
1323         for (i = 0; i < clt_path->queue_depth; ++i) {
1324                 req = &clt_path->reqs[i];
1325                 if (!req->in_use)
1326                         continue;
1327
1328                 /*
1329                  * Safely (without notification) complete failed request.
1330                  * After completion this request is still useble and can
1331                  * be failovered to another path.
1332                  */
1333                 complete_rdma_req(req, -ECONNABORTED, false, true);
1334
1335                 err = rtrs_clt_failover_req(clt, req);
1336                 if (err)
1337                         /* Failover failed, notify anyway */
1338                         req->conf(req->priv, err);
1339         }
1340 }
1341
1342 static void free_path_reqs(struct rtrs_clt_path *clt_path)
1343 {
1344         struct rtrs_clt_io_req *req;
1345         int i;
1346
1347         if (!clt_path->reqs)
1348                 return;
1349         for (i = 0; i < clt_path->queue_depth; ++i) {
1350                 req = &clt_path->reqs[i];
1351                 if (req->mr)
1352                         ib_dereg_mr(req->mr);
1353                 kfree(req->sge);
1354                 rtrs_iu_free(req->iu, clt_path->s.dev->ib_dev, 1);
1355         }
1356         kfree(clt_path->reqs);
1357         clt_path->reqs = NULL;
1358 }
1359
1360 static int alloc_path_reqs(struct rtrs_clt_path *clt_path)
1361 {
1362         struct rtrs_clt_io_req *req;
1363         int i, err = -ENOMEM;
1364
1365         clt_path->reqs = kcalloc(clt_path->queue_depth,
1366                                  sizeof(*clt_path->reqs),
1367                                  GFP_KERNEL);
1368         if (!clt_path->reqs)
1369                 return -ENOMEM;
1370
1371         for (i = 0; i < clt_path->queue_depth; ++i) {
1372                 req = &clt_path->reqs[i];
1373                 req->iu = rtrs_iu_alloc(1, clt_path->max_hdr_size, GFP_KERNEL,
1374                                          clt_path->s.dev->ib_dev,
1375                                          DMA_TO_DEVICE,
1376                                          rtrs_clt_rdma_done);
1377                 if (!req->iu)
1378                         goto out;
1379
1380                 req->sge = kcalloc(2, sizeof(*req->sge), GFP_KERNEL);
1381                 if (!req->sge)
1382                         goto out;
1383
1384                 req->mr = ib_alloc_mr(clt_path->s.dev->ib_pd,
1385                                       IB_MR_TYPE_MEM_REG,
1386                                       clt_path->max_pages_per_mr);
1387                 if (IS_ERR(req->mr)) {
1388                         err = PTR_ERR(req->mr);
1389                         req->mr = NULL;
1390                         pr_err("Failed to alloc clt_path->max_pages_per_mr %d\n",
1391                                clt_path->max_pages_per_mr);
1392                         goto out;
1393                 }
1394
1395                 init_completion(&req->inv_comp);
1396         }
1397
1398         return 0;
1399
1400 out:
1401         free_path_reqs(clt_path);
1402
1403         return err;
1404 }
1405
1406 static int alloc_permits(struct rtrs_clt_sess *clt)
1407 {
1408         unsigned int chunk_bits;
1409         int err, i;
1410
1411         clt->permits_map = bitmap_zalloc(clt->queue_depth, GFP_KERNEL);
1412         if (!clt->permits_map) {
1413                 err = -ENOMEM;
1414                 goto out_err;
1415         }
1416         clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1417         if (!clt->permits) {
1418                 err = -ENOMEM;
1419                 goto err_map;
1420         }
1421         chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1422         for (i = 0; i < clt->queue_depth; i++) {
1423                 struct rtrs_permit *permit;
1424
1425                 permit = get_permit(clt, i);
1426                 permit->mem_id = i;
1427                 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1428         }
1429
1430         return 0;
1431
1432 err_map:
1433         bitmap_free(clt->permits_map);
1434         clt->permits_map = NULL;
1435 out_err:
1436         return err;
1437 }
1438
1439 static void free_permits(struct rtrs_clt_sess *clt)
1440 {
1441         if (clt->permits_map)
1442                 wait_event(clt->permits_wait,
1443                            bitmap_empty(clt->permits_map, clt->queue_depth));
1444
1445         bitmap_free(clt->permits_map);
1446         clt->permits_map = NULL;
1447         kfree(clt->permits);
1448         clt->permits = NULL;
1449 }
1450
1451 static void query_fast_reg_mode(struct rtrs_clt_path *clt_path)
1452 {
1453         struct ib_device *ib_dev;
1454         u64 max_pages_per_mr;
1455         int mr_page_shift;
1456
1457         ib_dev = clt_path->s.dev->ib_dev;
1458
1459         /*
1460          * Use the smallest page size supported by the HCA, down to a
1461          * minimum of 4096 bytes. We're unlikely to build large sglists
1462          * out of smaller entries.
1463          */
1464         mr_page_shift      = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1465         max_pages_per_mr   = ib_dev->attrs.max_mr_size;
1466         do_div(max_pages_per_mr, (1ull << mr_page_shift));
1467         clt_path->max_pages_per_mr =
1468                 min3(clt_path->max_pages_per_mr, (u32)max_pages_per_mr,
1469                      ib_dev->attrs.max_fast_reg_page_list_len);
1470         clt_path->clt->max_segments =
1471                 min(clt_path->max_pages_per_mr, clt_path->clt->max_segments);
1472 }
1473
1474 static bool rtrs_clt_change_state_get_old(struct rtrs_clt_path *clt_path,
1475                                            enum rtrs_clt_state new_state,
1476                                            enum rtrs_clt_state *old_state)
1477 {
1478         bool changed;
1479
1480         spin_lock_irq(&clt_path->state_wq.lock);
1481         if (old_state)
1482                 *old_state = clt_path->state;
1483         changed = rtrs_clt_change_state(clt_path, new_state);
1484         spin_unlock_irq(&clt_path->state_wq.lock);
1485
1486         return changed;
1487 }
1488
1489 static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1490 {
1491         struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1492
1493         rtrs_rdma_error_recovery(con);
1494 }
1495
1496 static void rtrs_clt_init_hb(struct rtrs_clt_path *clt_path)
1497 {
1498         rtrs_init_hb(&clt_path->s, &io_comp_cqe,
1499                       RTRS_HB_INTERVAL_MS,
1500                       RTRS_HB_MISSED_MAX,
1501                       rtrs_clt_hb_err_handler,
1502                       rtrs_wq);
1503 }
1504
1505 static void rtrs_clt_reconnect_work(struct work_struct *work);
1506 static void rtrs_clt_close_work(struct work_struct *work);
1507
1508 static void rtrs_clt_err_recovery_work(struct work_struct *work)
1509 {
1510         struct rtrs_clt_path *clt_path;
1511         struct rtrs_clt_sess *clt;
1512         int delay_ms;
1513
1514         clt_path = container_of(work, struct rtrs_clt_path, err_recovery_work);
1515         clt = clt_path->clt;
1516         delay_ms = clt->reconnect_delay_sec * 1000;
1517         rtrs_clt_stop_and_destroy_conns(clt_path);
1518         queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork,
1519                            msecs_to_jiffies(delay_ms +
1520                                             prandom_u32() %
1521                                             RTRS_RECONNECT_SEED));
1522 }
1523
1524 static struct rtrs_clt_path *alloc_path(struct rtrs_clt_sess *clt,
1525                                         const struct rtrs_addr *path,
1526                                         size_t con_num, u32 nr_poll_queues)
1527 {
1528         struct rtrs_clt_path *clt_path;
1529         int err = -ENOMEM;
1530         int cpu;
1531         size_t total_con;
1532
1533         clt_path = kzalloc(sizeof(*clt_path), GFP_KERNEL);
1534         if (!clt_path)
1535                 goto err;
1536
1537         /*
1538          * irqmode and poll
1539          * +1: Extra connection for user messages
1540          */
1541         total_con = con_num + nr_poll_queues + 1;
1542         clt_path->s.con = kcalloc(total_con, sizeof(*clt_path->s.con),
1543                                   GFP_KERNEL);
1544         if (!clt_path->s.con)
1545                 goto err_free_path;
1546
1547         clt_path->s.con_num = total_con;
1548         clt_path->s.irq_con_num = con_num + 1;
1549
1550         clt_path->stats = kzalloc(sizeof(*clt_path->stats), GFP_KERNEL);
1551         if (!clt_path->stats)
1552                 goto err_free_con;
1553
1554         mutex_init(&clt_path->init_mutex);
1555         uuid_gen(&clt_path->s.uuid);
1556         memcpy(&clt_path->s.dst_addr, path->dst,
1557                rdma_addr_size((struct sockaddr *)path->dst));
1558
1559         /*
1560          * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1561          * checks the sa_family to be non-zero. If user passed src_addr=NULL
1562          * the sess->src_addr will contain only zeros, which is then fine.
1563          */
1564         if (path->src)
1565                 memcpy(&clt_path->s.src_addr, path->src,
1566                        rdma_addr_size((struct sockaddr *)path->src));
1567         strscpy(clt_path->s.sessname, clt->sessname,
1568                 sizeof(clt_path->s.sessname));
1569         clt_path->clt = clt;
1570         clt_path->max_pages_per_mr = RTRS_MAX_SEGMENTS;
1571         init_waitqueue_head(&clt_path->state_wq);
1572         clt_path->state = RTRS_CLT_CONNECTING;
1573         atomic_set(&clt_path->connected_cnt, 0);
1574         INIT_WORK(&clt_path->close_work, rtrs_clt_close_work);
1575         INIT_WORK(&clt_path->err_recovery_work, rtrs_clt_err_recovery_work);
1576         INIT_DELAYED_WORK(&clt_path->reconnect_dwork, rtrs_clt_reconnect_work);
1577         rtrs_clt_init_hb(clt_path);
1578
1579         clt_path->mp_skip_entry = alloc_percpu(typeof(*clt_path->mp_skip_entry));
1580         if (!clt_path->mp_skip_entry)
1581                 goto err_free_stats;
1582
1583         for_each_possible_cpu(cpu)
1584                 INIT_LIST_HEAD(per_cpu_ptr(clt_path->mp_skip_entry, cpu));
1585
1586         err = rtrs_clt_init_stats(clt_path->stats);
1587         if (err)
1588                 goto err_free_percpu;
1589
1590         return clt_path;
1591
1592 err_free_percpu:
1593         free_percpu(clt_path->mp_skip_entry);
1594 err_free_stats:
1595         kfree(clt_path->stats);
1596 err_free_con:
1597         kfree(clt_path->s.con);
1598 err_free_path:
1599         kfree(clt_path);
1600 err:
1601         return ERR_PTR(err);
1602 }
1603
1604 void free_path(struct rtrs_clt_path *clt_path)
1605 {
1606         free_percpu(clt_path->mp_skip_entry);
1607         mutex_destroy(&clt_path->init_mutex);
1608         kfree(clt_path->s.con);
1609         kfree(clt_path->rbufs);
1610         kfree(clt_path);
1611 }
1612
1613 static int create_con(struct rtrs_clt_path *clt_path, unsigned int cid)
1614 {
1615         struct rtrs_clt_con *con;
1616
1617         con = kzalloc(sizeof(*con), GFP_KERNEL);
1618         if (!con)
1619                 return -ENOMEM;
1620
1621         /* Map first two connections to the first CPU */
1622         con->cpu  = (cid ? cid - 1 : 0) % nr_cpu_ids;
1623         con->c.cid = cid;
1624         con->c.path = &clt_path->s;
1625         /* Align with srv, init as 1 */
1626         atomic_set(&con->c.wr_cnt, 1);
1627         mutex_init(&con->con_mutex);
1628
1629         clt_path->s.con[cid] = &con->c;
1630
1631         return 0;
1632 }
1633
1634 static void destroy_con(struct rtrs_clt_con *con)
1635 {
1636         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1637
1638         clt_path->s.con[con->c.cid] = NULL;
1639         mutex_destroy(&con->con_mutex);
1640         kfree(con);
1641 }
1642
1643 static int create_con_cq_qp(struct rtrs_clt_con *con)
1644 {
1645         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1646         u32 max_send_wr, max_recv_wr, cq_num, max_send_sge, wr_limit;
1647         int err, cq_vector;
1648         struct rtrs_msg_rkey_rsp *rsp;
1649
1650         lockdep_assert_held(&con->con_mutex);
1651         if (con->c.cid == 0) {
1652                 max_send_sge = 1;
1653                 /* We must be the first here */
1654                 if (WARN_ON(clt_path->s.dev))
1655                         return -EINVAL;
1656
1657                 /*
1658                  * The whole session uses device from user connection.
1659                  * Be careful not to close user connection before ib dev
1660                  * is gracefully put.
1661                  */
1662                 clt_path->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1663                                                        &dev_pd);
1664                 if (!clt_path->s.dev) {
1665                         rtrs_wrn(clt_path->clt,
1666                                   "rtrs_ib_dev_find_get_or_add(): no memory\n");
1667                         return -ENOMEM;
1668                 }
1669                 clt_path->s.dev_ref = 1;
1670                 query_fast_reg_mode(clt_path);
1671                 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1672                 /*
1673                  * Two (request + registration) completion for send
1674                  * Two for recv if always_invalidate is set on server
1675                  * or one for recv.
1676                  * + 2 for drain and heartbeat
1677                  * in case qp gets into error state.
1678                  */
1679                 max_send_wr =
1680                         min_t(int, wr_limit, SERVICE_CON_QUEUE_DEPTH * 2 + 2);
1681                 max_recv_wr = max_send_wr;
1682         } else {
1683                 /*
1684                  * Here we assume that session members are correctly set.
1685                  * This is always true if user connection (cid == 0) is
1686                  * established first.
1687                  */
1688                 if (WARN_ON(!clt_path->s.dev))
1689                         return -EINVAL;
1690                 if (WARN_ON(!clt_path->queue_depth))
1691                         return -EINVAL;
1692
1693                 wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1694                 /* Shared between connections */
1695                 clt_path->s.dev_ref++;
1696                 max_send_wr = min_t(int, wr_limit,
1697                               /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1698                               clt_path->queue_depth * 3 + 1);
1699                 max_recv_wr = min_t(int, wr_limit,
1700                               clt_path->queue_depth * 3 + 1);
1701                 max_send_sge = 2;
1702         }
1703         atomic_set(&con->c.sq_wr_avail, max_send_wr);
1704         cq_num = max_send_wr + max_recv_wr;
1705         /* alloc iu to recv new rkey reply when server reports flags set */
1706         if (clt_path->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1707                 con->rsp_ius = rtrs_iu_alloc(cq_num, sizeof(*rsp),
1708                                               GFP_KERNEL,
1709                                               clt_path->s.dev->ib_dev,
1710                                               DMA_FROM_DEVICE,
1711                                               rtrs_clt_rdma_done);
1712                 if (!con->rsp_ius)
1713                         return -ENOMEM;
1714                 con->queue_num = cq_num;
1715         }
1716         cq_num = max_send_wr + max_recv_wr;
1717         cq_vector = con->cpu % clt_path->s.dev->ib_dev->num_comp_vectors;
1718         if (con->c.cid >= clt_path->s.irq_con_num)
1719                 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1720                                         cq_vector, cq_num, max_send_wr,
1721                                         max_recv_wr, IB_POLL_DIRECT);
1722         else
1723                 err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1724                                         cq_vector, cq_num, max_send_wr,
1725                                         max_recv_wr, IB_POLL_SOFTIRQ);
1726         /*
1727          * In case of error we do not bother to clean previous allocations,
1728          * since destroy_con_cq_qp() must be called.
1729          */
1730         return err;
1731 }
1732
1733 static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1734 {
1735         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1736
1737         /*
1738          * Be careful here: destroy_con_cq_qp() can be called even
1739          * create_con_cq_qp() failed, see comments there.
1740          */
1741         lockdep_assert_held(&con->con_mutex);
1742         rtrs_cq_qp_destroy(&con->c);
1743         if (con->rsp_ius) {
1744                 rtrs_iu_free(con->rsp_ius, clt_path->s.dev->ib_dev,
1745                              con->queue_num);
1746                 con->rsp_ius = NULL;
1747                 con->queue_num = 0;
1748         }
1749         if (clt_path->s.dev_ref && !--clt_path->s.dev_ref) {
1750                 rtrs_ib_dev_put(clt_path->s.dev);
1751                 clt_path->s.dev = NULL;
1752         }
1753 }
1754
1755 static void stop_cm(struct rtrs_clt_con *con)
1756 {
1757         rdma_disconnect(con->c.cm_id);
1758         if (con->c.qp)
1759                 ib_drain_qp(con->c.qp);
1760 }
1761
1762 static void destroy_cm(struct rtrs_clt_con *con)
1763 {
1764         rdma_destroy_id(con->c.cm_id);
1765         con->c.cm_id = NULL;
1766 }
1767
1768 static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1769 {
1770         struct rtrs_path *s = con->c.path;
1771         int err;
1772
1773         mutex_lock(&con->con_mutex);
1774         err = create_con_cq_qp(con);
1775         mutex_unlock(&con->con_mutex);
1776         if (err) {
1777                 rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1778                 return err;
1779         }
1780         err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1781         if (err)
1782                 rtrs_err(s, "Resolving route failed, err: %d\n", err);
1783
1784         return err;
1785 }
1786
1787 static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1788 {
1789         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1790         struct rtrs_clt_sess *clt = clt_path->clt;
1791         struct rtrs_msg_conn_req msg;
1792         struct rdma_conn_param param;
1793
1794         int err;
1795
1796         param = (struct rdma_conn_param) {
1797                 .retry_count = 7,
1798                 .rnr_retry_count = 7,
1799                 .private_data = &msg,
1800                 .private_data_len = sizeof(msg),
1801         };
1802
1803         msg = (struct rtrs_msg_conn_req) {
1804                 .magic = cpu_to_le16(RTRS_MAGIC),
1805                 .version = cpu_to_le16(RTRS_PROTO_VER),
1806                 .cid = cpu_to_le16(con->c.cid),
1807                 .cid_num = cpu_to_le16(clt_path->s.con_num),
1808                 .recon_cnt = cpu_to_le16(clt_path->s.recon_cnt),
1809         };
1810         msg.first_conn = clt_path->for_new_clt ? FIRST_CONN : 0;
1811         uuid_copy(&msg.sess_uuid, &clt_path->s.uuid);
1812         uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1813
1814         err = rdma_connect_locked(con->c.cm_id, &param);
1815         if (err)
1816                 rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1817
1818         return err;
1819 }
1820
1821 static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1822                                        struct rdma_cm_event *ev)
1823 {
1824         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1825         struct rtrs_clt_sess *clt = clt_path->clt;
1826         const struct rtrs_msg_conn_rsp *msg;
1827         u16 version, queue_depth;
1828         int errno;
1829         u8 len;
1830
1831         msg = ev->param.conn.private_data;
1832         len = ev->param.conn.private_data_len;
1833         if (len < sizeof(*msg)) {
1834                 rtrs_err(clt, "Invalid RTRS connection response\n");
1835                 return -ECONNRESET;
1836         }
1837         if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1838                 rtrs_err(clt, "Invalid RTRS magic\n");
1839                 return -ECONNRESET;
1840         }
1841         version = le16_to_cpu(msg->version);
1842         if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1843                 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1844                           version >> 8, RTRS_PROTO_VER_MAJOR);
1845                 return -ECONNRESET;
1846         }
1847         errno = le16_to_cpu(msg->errno);
1848         if (errno) {
1849                 rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1850                           errno);
1851                 return -ECONNRESET;
1852         }
1853         if (con->c.cid == 0) {
1854                 queue_depth = le16_to_cpu(msg->queue_depth);
1855
1856                 if (clt_path->queue_depth > 0 && queue_depth != clt_path->queue_depth) {
1857                         rtrs_err(clt, "Error: queue depth changed\n");
1858
1859                         /*
1860                          * Stop any more reconnection attempts
1861                          */
1862                         clt_path->reconnect_attempts = -1;
1863                         rtrs_err(clt,
1864                                 "Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n");
1865                         return -ECONNRESET;
1866                 }
1867
1868                 if (!clt_path->rbufs) {
1869                         clt_path->rbufs = kcalloc(queue_depth,
1870                                                   sizeof(*clt_path->rbufs),
1871                                                   GFP_KERNEL);
1872                         if (!clt_path->rbufs)
1873                                 return -ENOMEM;
1874                 }
1875                 clt_path->queue_depth = queue_depth;
1876                 clt_path->s.signal_interval = min_not_zero(queue_depth,
1877                                                 (unsigned short) SERVICE_CON_QUEUE_DEPTH);
1878                 clt_path->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1879                 clt_path->max_io_size = le32_to_cpu(msg->max_io_size);
1880                 clt_path->flags = le32_to_cpu(msg->flags);
1881                 clt_path->chunk_size = clt_path->max_io_size + clt_path->max_hdr_size;
1882
1883                 /*
1884                  * Global IO size is always a minimum.
1885                  * If while a reconnection server sends us a value a bit
1886                  * higher - client does not care and uses cached minimum.
1887                  *
1888                  * Since we can have several sessions (paths) restablishing
1889                  * connections in parallel, use lock.
1890                  */
1891                 mutex_lock(&clt->paths_mutex);
1892                 clt->queue_depth = clt_path->queue_depth;
1893                 clt->max_io_size = min_not_zero(clt_path->max_io_size,
1894                                                 clt->max_io_size);
1895                 mutex_unlock(&clt->paths_mutex);
1896
1897                 /*
1898                  * Cache the hca_port and hca_name for sysfs
1899                  */
1900                 clt_path->hca_port = con->c.cm_id->port_num;
1901                 scnprintf(clt_path->hca_name, sizeof(clt_path->hca_name),
1902                           clt_path->s.dev->ib_dev->name);
1903                 clt_path->s.src_addr = con->c.cm_id->route.addr.src_addr;
1904                 /* set for_new_clt, to allow future reconnect on any path */
1905                 clt_path->for_new_clt = 1;
1906         }
1907
1908         return 0;
1909 }
1910
1911 static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1912 {
1913         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1914
1915         atomic_inc(&clt_path->connected_cnt);
1916         con->cm_err = 1;
1917 }
1918
1919 static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1920                                     struct rdma_cm_event *ev)
1921 {
1922         struct rtrs_path *s = con->c.path;
1923         const struct rtrs_msg_conn_rsp *msg;
1924         const char *rej_msg;
1925         int status, errno;
1926         u8 data_len;
1927
1928         status = ev->status;
1929         rej_msg = rdma_reject_msg(con->c.cm_id, status);
1930         msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1931
1932         if (msg && data_len >= sizeof(*msg)) {
1933                 errno = (int16_t)le16_to_cpu(msg->errno);
1934                 if (errno == -EBUSY)
1935                         rtrs_err(s,
1936                                   "Previous session is still exists on the server, please reconnect later\n");
1937                 else
1938                         rtrs_err(s,
1939                                   "Connect rejected: status %d (%s), rtrs errno %d\n",
1940                                   status, rej_msg, errno);
1941         } else {
1942                 rtrs_err(s,
1943                           "Connect rejected but with malformed message: status %d (%s)\n",
1944                           status, rej_msg);
1945         }
1946
1947         return -ECONNRESET;
1948 }
1949
1950 void rtrs_clt_close_conns(struct rtrs_clt_path *clt_path, bool wait)
1951 {
1952         trace_rtrs_clt_close_conns(clt_path);
1953
1954         if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSING, NULL))
1955                 queue_work(rtrs_wq, &clt_path->close_work);
1956         if (wait)
1957                 flush_work(&clt_path->close_work);
1958 }
1959
1960 static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1961 {
1962         if (con->cm_err == 1) {
1963                 struct rtrs_clt_path *clt_path;
1964
1965                 clt_path = to_clt_path(con->c.path);
1966                 if (atomic_dec_and_test(&clt_path->connected_cnt))
1967
1968                         wake_up(&clt_path->state_wq);
1969         }
1970         con->cm_err = cm_err;
1971 }
1972
1973 static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1974                                      struct rdma_cm_event *ev)
1975 {
1976         struct rtrs_clt_con *con = cm_id->context;
1977         struct rtrs_path *s = con->c.path;
1978         struct rtrs_clt_path *clt_path = to_clt_path(s);
1979         int cm_err = 0;
1980
1981         switch (ev->event) {
1982         case RDMA_CM_EVENT_ADDR_RESOLVED:
1983                 cm_err = rtrs_rdma_addr_resolved(con);
1984                 break;
1985         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1986                 cm_err = rtrs_rdma_route_resolved(con);
1987                 break;
1988         case RDMA_CM_EVENT_ESTABLISHED:
1989                 cm_err = rtrs_rdma_conn_established(con, ev);
1990                 if (!cm_err) {
1991                         /*
1992                          * Report success and wake up. Here we abuse state_wq,
1993                          * i.e. wake up without state change, but we set cm_err.
1994                          */
1995                         flag_success_on_conn(con);
1996                         wake_up(&clt_path->state_wq);
1997                         return 0;
1998                 }
1999                 break;
2000         case RDMA_CM_EVENT_REJECTED:
2001                 cm_err = rtrs_rdma_conn_rejected(con, ev);
2002                 break;
2003         case RDMA_CM_EVENT_DISCONNECTED:
2004                 /* No message for disconnecting */
2005                 cm_err = -ECONNRESET;
2006                 break;
2007         case RDMA_CM_EVENT_CONNECT_ERROR:
2008         case RDMA_CM_EVENT_UNREACHABLE:
2009         case RDMA_CM_EVENT_ADDR_CHANGE:
2010         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2011                 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2012                          rdma_event_msg(ev->event), ev->status);
2013                 cm_err = -ECONNRESET;
2014                 break;
2015         case RDMA_CM_EVENT_ADDR_ERROR:
2016         case RDMA_CM_EVENT_ROUTE_ERROR:
2017                 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2018                          rdma_event_msg(ev->event), ev->status);
2019                 cm_err = -EHOSTUNREACH;
2020                 break;
2021         case RDMA_CM_EVENT_DEVICE_REMOVAL:
2022                 /*
2023                  * Device removal is a special case.  Queue close and return 0.
2024                  */
2025                 rtrs_clt_close_conns(clt_path, false);
2026                 return 0;
2027         default:
2028                 rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %d)\n",
2029                          rdma_event_msg(ev->event), ev->status);
2030                 cm_err = -ECONNRESET;
2031                 break;
2032         }
2033
2034         if (cm_err) {
2035                 /*
2036                  * cm error makes sense only on connection establishing,
2037                  * in other cases we rely on normal procedure of reconnecting.
2038                  */
2039                 flag_error_on_conn(con, cm_err);
2040                 rtrs_rdma_error_recovery(con);
2041         }
2042
2043         return 0;
2044 }
2045
2046 static int create_cm(struct rtrs_clt_con *con)
2047 {
2048         struct rtrs_path *s = con->c.path;
2049         struct rtrs_clt_path *clt_path = to_clt_path(s);
2050         struct rdma_cm_id *cm_id;
2051         int err;
2052
2053         cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
2054                                clt_path->s.dst_addr.ss_family == AF_IB ?
2055                                RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
2056         if (IS_ERR(cm_id)) {
2057                 err = PTR_ERR(cm_id);
2058                 rtrs_err(s, "Failed to create CM ID, err: %d\n", err);
2059
2060                 return err;
2061         }
2062         con->c.cm_id = cm_id;
2063         con->cm_err = 0;
2064         /* allow the port to be reused */
2065         err = rdma_set_reuseaddr(cm_id, 1);
2066         if (err != 0) {
2067                 rtrs_err(s, "Set address reuse failed, err: %d\n", err);
2068                 goto destroy_cm;
2069         }
2070         err = rdma_resolve_addr(cm_id, (struct sockaddr *)&clt_path->s.src_addr,
2071                                 (struct sockaddr *)&clt_path->s.dst_addr,
2072                                 RTRS_CONNECT_TIMEOUT_MS);
2073         if (err) {
2074                 rtrs_err(s, "Failed to resolve address, err: %d\n", err);
2075                 goto destroy_cm;
2076         }
2077         /*
2078          * Combine connection status and session events. This is needed
2079          * for waiting two possible cases: cm_err has something meaningful
2080          * or session state was really changed to error by device removal.
2081          */
2082         err = wait_event_interruptible_timeout(
2083                         clt_path->state_wq,
2084                         con->cm_err || clt_path->state != RTRS_CLT_CONNECTING,
2085                         msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2086         if (err == 0 || err == -ERESTARTSYS) {
2087                 if (err == 0)
2088                         err = -ETIMEDOUT;
2089                 /* Timedout or interrupted */
2090                 goto errr;
2091         }
2092         if (con->cm_err < 0) {
2093                 err = con->cm_err;
2094                 goto errr;
2095         }
2096         if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTING) {
2097                 /* Device removal */
2098                 err = -ECONNABORTED;
2099                 goto errr;
2100         }
2101
2102         return 0;
2103
2104 errr:
2105         stop_cm(con);
2106         mutex_lock(&con->con_mutex);
2107         destroy_con_cq_qp(con);
2108         mutex_unlock(&con->con_mutex);
2109 destroy_cm:
2110         destroy_cm(con);
2111
2112         return err;
2113 }
2114
2115 static void rtrs_clt_path_up(struct rtrs_clt_path *clt_path)
2116 {
2117         struct rtrs_clt_sess *clt = clt_path->clt;
2118         int up;
2119
2120         /*
2121          * We can fire RECONNECTED event only when all paths were
2122          * connected on rtrs_clt_open(), then each was disconnected
2123          * and the first one connected again.  That's why this nasty
2124          * game with counter value.
2125          */
2126
2127         mutex_lock(&clt->paths_ev_mutex);
2128         up = ++clt->paths_up;
2129         /*
2130          * Here it is safe to access paths num directly since up counter
2131          * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
2132          * in progress, thus paths removals are impossible.
2133          */
2134         if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
2135                 clt->paths_up = clt->paths_num;
2136         else if (up == 1)
2137                 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
2138         mutex_unlock(&clt->paths_ev_mutex);
2139
2140         /* Mark session as established */
2141         clt_path->established = true;
2142         clt_path->reconnect_attempts = 0;
2143         clt_path->stats->reconnects.successful_cnt++;
2144 }
2145
2146 static void rtrs_clt_path_down(struct rtrs_clt_path *clt_path)
2147 {
2148         struct rtrs_clt_sess *clt = clt_path->clt;
2149
2150         if (!clt_path->established)
2151                 return;
2152
2153         clt_path->established = false;
2154         mutex_lock(&clt->paths_ev_mutex);
2155         WARN_ON(!clt->paths_up);
2156         if (--clt->paths_up == 0)
2157                 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2158         mutex_unlock(&clt->paths_ev_mutex);
2159 }
2160
2161 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path)
2162 {
2163         struct rtrs_clt_con *con;
2164         unsigned int cid;
2165
2166         WARN_ON(READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED);
2167
2168         /*
2169          * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2170          * exactly in between.  Start destroying after it finishes.
2171          */
2172         mutex_lock(&clt_path->init_mutex);
2173         mutex_unlock(&clt_path->init_mutex);
2174
2175         /*
2176          * All IO paths must observe !CONNECTED state before we
2177          * free everything.
2178          */
2179         synchronize_rcu();
2180
2181         rtrs_stop_hb(&clt_path->s);
2182
2183         /*
2184          * The order it utterly crucial: firstly disconnect and complete all
2185          * rdma requests with error (thus set in_use=false for requests),
2186          * then fail outstanding requests checking in_use for each, and
2187          * eventually notify upper layer about session disconnection.
2188          */
2189
2190         for (cid = 0; cid < clt_path->s.con_num; cid++) {
2191                 if (!clt_path->s.con[cid])
2192                         break;
2193                 con = to_clt_con(clt_path->s.con[cid]);
2194                 stop_cm(con);
2195         }
2196         fail_all_outstanding_reqs(clt_path);
2197         free_path_reqs(clt_path);
2198         rtrs_clt_path_down(clt_path);
2199
2200         /*
2201          * Wait for graceful shutdown, namely when peer side invokes
2202          * rdma_disconnect(). 'connected_cnt' is decremented only on
2203          * CM events, thus if other side had crashed and hb has detected
2204          * something is wrong, here we will stuck for exactly timeout ms,
2205          * since CM does not fire anything.  That is fine, we are not in
2206          * hurry.
2207          */
2208         wait_event_timeout(clt_path->state_wq,
2209                            !atomic_read(&clt_path->connected_cnt),
2210                            msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2211
2212         for (cid = 0; cid < clt_path->s.con_num; cid++) {
2213                 if (!clt_path->s.con[cid])
2214                         break;
2215                 con = to_clt_con(clt_path->s.con[cid]);
2216                 mutex_lock(&con->con_mutex);
2217                 destroy_con_cq_qp(con);
2218                 mutex_unlock(&con->con_mutex);
2219                 destroy_cm(con);
2220                 destroy_con(con);
2221         }
2222 }
2223
2224 static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_path *clt_path)
2225 {
2226         struct rtrs_clt_sess *clt = clt_path->clt;
2227         struct rtrs_clt_path *next;
2228         bool wait_for_grace = false;
2229         int cpu;
2230
2231         mutex_lock(&clt->paths_mutex);
2232         list_del_rcu(&clt_path->s.entry);
2233
2234         /* Make sure everybody observes path removal. */
2235         synchronize_rcu();
2236
2237         /*
2238          * At this point nobody sees @sess in the list, but still we have
2239          * dangling pointer @pcpu_path which _can_ point to @sess.  Since
2240          * nobody can observe @sess in the list, we guarantee that IO path
2241          * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2242          * to @sess, but can never again become @sess.
2243          */
2244
2245         /*
2246          * Decrement paths number only after grace period, because
2247          * caller of do_each_path() must firstly observe list without
2248          * path and only then decremented paths number.
2249          *
2250          * Otherwise there can be the following situation:
2251          *    o Two paths exist and IO is coming.
2252          *    o One path is removed:
2253          *      CPU#0                          CPU#1
2254          *      do_each_path():                rtrs_clt_remove_path_from_arr():
2255          *          path = get_next_path()
2256          *          ^^^                            list_del_rcu(path)
2257          *          [!CONNECTED path]              clt->paths_num--
2258          *                                              ^^^^^^^^^
2259          *          load clt->paths_num                 from 2 to 1
2260          *                    ^^^^^^^^^
2261          *                    sees 1
2262          *
2263          *      path is observed as !CONNECTED, but do_each_path() loop
2264          *      ends, because expression i < clt->paths_num is false.
2265          */
2266         clt->paths_num--;
2267
2268         /*
2269          * Get @next connection from current @sess which is going to be
2270          * removed.  If @sess is the last element, then @next is NULL.
2271          */
2272         rcu_read_lock();
2273         next = rtrs_clt_get_next_path_or_null(&clt->paths_list, clt_path);
2274         rcu_read_unlock();
2275
2276         /*
2277          * @pcpu paths can still point to the path which is going to be
2278          * removed, so change the pointer manually.
2279          */
2280         for_each_possible_cpu(cpu) {
2281                 struct rtrs_clt_path __rcu **ppcpu_path;
2282
2283                 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2284                 if (rcu_dereference_protected(*ppcpu_path,
2285                         lockdep_is_held(&clt->paths_mutex)) != clt_path)
2286                         /*
2287                          * synchronize_rcu() was called just after deleting
2288                          * entry from the list, thus IO code path cannot
2289                          * change pointer back to the pointer which is going
2290                          * to be removed, we are safe here.
2291                          */
2292                         continue;
2293
2294                 /*
2295                  * We race with IO code path, which also changes pointer,
2296                  * thus we have to be careful not to overwrite it.
2297                  */
2298                 if (try_cmpxchg((struct rtrs_clt_path **)ppcpu_path, &clt_path,
2299                                 next))
2300                         /*
2301                          * @ppcpu_path was successfully replaced with @next,
2302                          * that means that someone could also pick up the
2303                          * @sess and dereferencing it right now, so wait for
2304                          * a grace period is required.
2305                          */
2306                         wait_for_grace = true;
2307         }
2308         if (wait_for_grace)
2309                 synchronize_rcu();
2310
2311         mutex_unlock(&clt->paths_mutex);
2312 }
2313
2314 static void rtrs_clt_add_path_to_arr(struct rtrs_clt_path *clt_path)
2315 {
2316         struct rtrs_clt_sess *clt = clt_path->clt;
2317
2318         mutex_lock(&clt->paths_mutex);
2319         clt->paths_num++;
2320
2321         list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2322         mutex_unlock(&clt->paths_mutex);
2323 }
2324
2325 static void rtrs_clt_close_work(struct work_struct *work)
2326 {
2327         struct rtrs_clt_path *clt_path;
2328
2329         clt_path = container_of(work, struct rtrs_clt_path, close_work);
2330
2331         cancel_work_sync(&clt_path->err_recovery_work);
2332         cancel_delayed_work_sync(&clt_path->reconnect_dwork);
2333         rtrs_clt_stop_and_destroy_conns(clt_path);
2334         rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSED, NULL);
2335 }
2336
2337 static int init_conns(struct rtrs_clt_path *clt_path)
2338 {
2339         unsigned int cid;
2340         int err;
2341
2342         /*
2343          * On every new session connections increase reconnect counter
2344          * to avoid clashes with previous sessions not yet closed
2345          * sessions on a server side.
2346          */
2347         clt_path->s.recon_cnt++;
2348
2349         /* Establish all RDMA connections  */
2350         for (cid = 0; cid < clt_path->s.con_num; cid++) {
2351                 err = create_con(clt_path, cid);
2352                 if (err)
2353                         goto destroy;
2354
2355                 err = create_cm(to_clt_con(clt_path->s.con[cid]));
2356                 if (err) {
2357                         destroy_con(to_clt_con(clt_path->s.con[cid]));
2358                         goto destroy;
2359                 }
2360         }
2361         err = alloc_path_reqs(clt_path);
2362         if (err)
2363                 goto destroy;
2364
2365         rtrs_start_hb(&clt_path->s);
2366
2367         return 0;
2368
2369 destroy:
2370         while (cid--) {
2371                 struct rtrs_clt_con *con = to_clt_con(clt_path->s.con[cid]);
2372
2373                 stop_cm(con);
2374
2375                 mutex_lock(&con->con_mutex);
2376                 destroy_con_cq_qp(con);
2377                 mutex_unlock(&con->con_mutex);
2378                 destroy_cm(con);
2379                 destroy_con(con);
2380         }
2381         /*
2382          * If we've never taken async path and got an error, say,
2383          * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2384          * manually to keep reconnecting.
2385          */
2386         rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2387
2388         return err;
2389 }
2390
2391 static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2392 {
2393         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2394         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2395         struct rtrs_iu *iu;
2396
2397         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2398         rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2399
2400         if (wc->status != IB_WC_SUCCESS) {
2401                 rtrs_err(clt_path->clt, "Path info request send failed: %s\n",
2402                           ib_wc_status_msg(wc->status));
2403                 rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2404                 return;
2405         }
2406
2407         rtrs_clt_update_wc_stats(con);
2408 }
2409
2410 static int process_info_rsp(struct rtrs_clt_path *clt_path,
2411                             const struct rtrs_msg_info_rsp *msg)
2412 {
2413         unsigned int sg_cnt, total_len;
2414         int i, sgi;
2415
2416         sg_cnt = le16_to_cpu(msg->sg_cnt);
2417         if (!sg_cnt || (clt_path->queue_depth % sg_cnt)) {
2418                 rtrs_err(clt_path->clt,
2419                           "Incorrect sg_cnt %d, is not multiple\n",
2420                           sg_cnt);
2421                 return -EINVAL;
2422         }
2423
2424         /*
2425          * Check if IB immediate data size is enough to hold the mem_id and
2426          * the offset inside the memory chunk.
2427          */
2428         if ((ilog2(sg_cnt - 1) + 1) + (ilog2(clt_path->chunk_size - 1) + 1) >
2429             MAX_IMM_PAYL_BITS) {
2430                 rtrs_err(clt_path->clt,
2431                           "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2432                           MAX_IMM_PAYL_BITS, sg_cnt, clt_path->chunk_size);
2433                 return -EINVAL;
2434         }
2435         total_len = 0;
2436         for (sgi = 0, i = 0; sgi < sg_cnt && i < clt_path->queue_depth; sgi++) {
2437                 const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2438                 u32 len, rkey;
2439                 u64 addr;
2440
2441                 addr = le64_to_cpu(desc->addr);
2442                 rkey = le32_to_cpu(desc->key);
2443                 len  = le32_to_cpu(desc->len);
2444
2445                 total_len += len;
2446
2447                 if (!len || (len % clt_path->chunk_size)) {
2448                         rtrs_err(clt_path->clt, "Incorrect [%d].len %d\n",
2449                                   sgi,
2450                                   len);
2451                         return -EINVAL;
2452                 }
2453                 for ( ; len && i < clt_path->queue_depth; i++) {
2454                         clt_path->rbufs[i].addr = addr;
2455                         clt_path->rbufs[i].rkey = rkey;
2456
2457                         len  -= clt_path->chunk_size;
2458                         addr += clt_path->chunk_size;
2459                 }
2460         }
2461         /* Sanity check */
2462         if (sgi != sg_cnt || i != clt_path->queue_depth) {
2463                 rtrs_err(clt_path->clt,
2464                          "Incorrect sg vector, not fully mapped\n");
2465                 return -EINVAL;
2466         }
2467         if (total_len != clt_path->chunk_size * clt_path->queue_depth) {
2468                 rtrs_err(clt_path->clt, "Incorrect total_len %d\n", total_len);
2469                 return -EINVAL;
2470         }
2471
2472         return 0;
2473 }
2474
2475 static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2476 {
2477         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2478         struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2479         struct rtrs_msg_info_rsp *msg;
2480         enum rtrs_clt_state state;
2481         struct rtrs_iu *iu;
2482         size_t rx_sz;
2483         int err;
2484
2485         state = RTRS_CLT_CONNECTING_ERR;
2486
2487         WARN_ON(con->c.cid);
2488         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2489         if (wc->status != IB_WC_SUCCESS) {
2490                 rtrs_err(clt_path->clt, "Path info response recv failed: %s\n",
2491                           ib_wc_status_msg(wc->status));
2492                 goto out;
2493         }
2494         WARN_ON(wc->opcode != IB_WC_RECV);
2495
2496         if (wc->byte_len < sizeof(*msg)) {
2497                 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2498                           wc->byte_len);
2499                 goto out;
2500         }
2501         ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
2502                                    iu->size, DMA_FROM_DEVICE);
2503         msg = iu->buf;
2504         if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP) {
2505                 rtrs_err(clt_path->clt, "Path info response is malformed: type %d\n",
2506                           le16_to_cpu(msg->type));
2507                 goto out;
2508         }
2509         rx_sz  = sizeof(*msg);
2510         rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2511         if (wc->byte_len < rx_sz) {
2512                 rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2513                           wc->byte_len);
2514                 goto out;
2515         }
2516         err = process_info_rsp(clt_path, msg);
2517         if (err)
2518                 goto out;
2519
2520         err = post_recv_path(clt_path);
2521         if (err)
2522                 goto out;
2523
2524         state = RTRS_CLT_CONNECTED;
2525
2526 out:
2527         rtrs_clt_update_wc_stats(con);
2528         rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2529         rtrs_clt_change_state_get_old(clt_path, state, NULL);
2530 }
2531
2532 static int rtrs_send_path_info(struct rtrs_clt_path *clt_path)
2533 {
2534         struct rtrs_clt_con *usr_con = to_clt_con(clt_path->s.con[0]);
2535         struct rtrs_msg_info_req *msg;
2536         struct rtrs_iu *tx_iu, *rx_iu;
2537         size_t rx_sz;
2538         int err;
2539
2540         rx_sz  = sizeof(struct rtrs_msg_info_rsp);
2541         rx_sz += sizeof(struct rtrs_sg_desc) * clt_path->queue_depth;
2542
2543         tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2544                                clt_path->s.dev->ib_dev, DMA_TO_DEVICE,
2545                                rtrs_clt_info_req_done);
2546         rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, clt_path->s.dev->ib_dev,
2547                                DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2548         if (!tx_iu || !rx_iu) {
2549                 err = -ENOMEM;
2550                 goto out;
2551         }
2552         /* Prepare for getting info response */
2553         err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2554         if (err) {
2555                 rtrs_err(clt_path->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2556                 goto out;
2557         }
2558         rx_iu = NULL;
2559
2560         msg = tx_iu->buf;
2561         msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2562         memcpy(msg->pathname, clt_path->s.sessname, sizeof(msg->pathname));
2563
2564         ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
2565                                       tx_iu->dma_addr,
2566                                       tx_iu->size, DMA_TO_DEVICE);
2567
2568         /* Send info request */
2569         err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2570         if (err) {
2571                 rtrs_err(clt_path->clt, "rtrs_iu_post_send(), err: %d\n", err);
2572                 goto out;
2573         }
2574         tx_iu = NULL;
2575
2576         /* Wait for state change */
2577         wait_event_interruptible_timeout(clt_path->state_wq,
2578                                          clt_path->state != RTRS_CLT_CONNECTING,
2579                                          msecs_to_jiffies(
2580                                                  RTRS_CONNECT_TIMEOUT_MS));
2581         if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) {
2582                 if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTING_ERR)
2583                         err = -ECONNRESET;
2584                 else
2585                         err = -ETIMEDOUT;
2586         }
2587
2588 out:
2589         if (tx_iu)
2590                 rtrs_iu_free(tx_iu, clt_path->s.dev->ib_dev, 1);
2591         if (rx_iu)
2592                 rtrs_iu_free(rx_iu, clt_path->s.dev->ib_dev, 1);
2593         if (err)
2594                 /* If we've never taken async path because of malloc problems */
2595                 rtrs_clt_change_state_get_old(clt_path,
2596                                               RTRS_CLT_CONNECTING_ERR, NULL);
2597
2598         return err;
2599 }
2600
2601 /**
2602  * init_path() - establishes all path connections and does handshake
2603  * @clt_path: client path.
2604  * In case of error full close or reconnect procedure should be taken,
2605  * because reconnect or close async works can be started.
2606  */
2607 static int init_path(struct rtrs_clt_path *clt_path)
2608 {
2609         int err;
2610         char str[NAME_MAX];
2611         struct rtrs_addr path = {
2612                 .src = &clt_path->s.src_addr,
2613                 .dst = &clt_path->s.dst_addr,
2614         };
2615
2616         rtrs_addr_to_str(&path, str, sizeof(str));
2617
2618         mutex_lock(&clt_path->init_mutex);
2619         err = init_conns(clt_path);
2620         if (err) {
2621                 rtrs_err(clt_path->clt,
2622                          "init_conns() failed: err=%d path=%s [%s:%u]\n", err,
2623                          str, clt_path->hca_name, clt_path->hca_port);
2624                 goto out;
2625         }
2626         err = rtrs_send_path_info(clt_path);
2627         if (err) {
2628                 rtrs_err(clt_path->clt,
2629                          "rtrs_send_path_info() failed: err=%d path=%s [%s:%u]\n",
2630                          err, str, clt_path->hca_name, clt_path->hca_port);
2631                 goto out;
2632         }
2633         rtrs_clt_path_up(clt_path);
2634 out:
2635         mutex_unlock(&clt_path->init_mutex);
2636
2637         return err;
2638 }
2639
2640 static void rtrs_clt_reconnect_work(struct work_struct *work)
2641 {
2642         struct rtrs_clt_path *clt_path;
2643         struct rtrs_clt_sess *clt;
2644         int err;
2645
2646         clt_path = container_of(to_delayed_work(work), struct rtrs_clt_path,
2647                                 reconnect_dwork);
2648         clt = clt_path->clt;
2649
2650         trace_rtrs_clt_reconnect_work(clt_path);
2651
2652         if (READ_ONCE(clt_path->state) != RTRS_CLT_RECONNECTING)
2653                 return;
2654
2655         if (clt_path->reconnect_attempts >= clt->max_reconnect_attempts) {
2656                 /* Close a path completely if max attempts is reached */
2657                 rtrs_clt_close_conns(clt_path, false);
2658                 return;
2659         }
2660         clt_path->reconnect_attempts++;
2661
2662         msleep(RTRS_RECONNECT_BACKOFF);
2663         if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING, NULL)) {
2664                 err = init_path(clt_path);
2665                 if (err)
2666                         goto reconnect_again;
2667         }
2668
2669         return;
2670
2671 reconnect_again:
2672         if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_RECONNECTING, NULL)) {
2673                 clt_path->stats->reconnects.fail_cnt++;
2674                 queue_work(rtrs_wq, &clt_path->err_recovery_work);
2675         }
2676 }
2677
2678 static void rtrs_clt_dev_release(struct device *dev)
2679 {
2680         struct rtrs_clt_sess *clt = container_of(dev, struct rtrs_clt_sess,
2681                                                  dev);
2682
2683         mutex_destroy(&clt->paths_ev_mutex);
2684         mutex_destroy(&clt->paths_mutex);
2685         kfree(clt);
2686 }
2687
2688 static struct rtrs_clt_sess *alloc_clt(const char *sessname, size_t paths_num,
2689                                   u16 port, size_t pdu_sz, void *priv,
2690                                   void  (*link_ev)(void *priv,
2691                                                    enum rtrs_clt_link_ev ev),
2692                                   unsigned int reconnect_delay_sec,
2693                                   unsigned int max_reconnect_attempts)
2694 {
2695         struct rtrs_clt_sess *clt;
2696         int err;
2697
2698         if (!paths_num || paths_num > MAX_PATHS_NUM)
2699                 return ERR_PTR(-EINVAL);
2700
2701         if (strlen(sessname) >= sizeof(clt->sessname))
2702                 return ERR_PTR(-EINVAL);
2703
2704         clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2705         if (!clt)
2706                 return ERR_PTR(-ENOMEM);
2707
2708         clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2709         if (!clt->pcpu_path) {
2710                 kfree(clt);
2711                 return ERR_PTR(-ENOMEM);
2712         }
2713
2714         clt->dev.class = rtrs_clt_dev_class;
2715         clt->dev.release = rtrs_clt_dev_release;
2716         uuid_gen(&clt->paths_uuid);
2717         INIT_LIST_HEAD_RCU(&clt->paths_list);
2718         clt->paths_num = paths_num;
2719         clt->paths_up = MAX_PATHS_NUM;
2720         clt->port = port;
2721         clt->pdu_sz = pdu_sz;
2722         clt->max_segments = RTRS_MAX_SEGMENTS;
2723         clt->reconnect_delay_sec = reconnect_delay_sec;
2724         clt->max_reconnect_attempts = max_reconnect_attempts;
2725         clt->priv = priv;
2726         clt->link_ev = link_ev;
2727         clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2728         strscpy(clt->sessname, sessname, sizeof(clt->sessname));
2729         init_waitqueue_head(&clt->permits_wait);
2730         mutex_init(&clt->paths_ev_mutex);
2731         mutex_init(&clt->paths_mutex);
2732         device_initialize(&clt->dev);
2733
2734         err = dev_set_name(&clt->dev, "%s", sessname);
2735         if (err)
2736                 goto err_put;
2737
2738         /*
2739          * Suppress user space notification until
2740          * sysfs files are created
2741          */
2742         dev_set_uevent_suppress(&clt->dev, true);
2743         err = device_add(&clt->dev);
2744         if (err)
2745                 goto err_put;
2746
2747         clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2748         if (!clt->kobj_paths) {
2749                 err = -ENOMEM;
2750                 goto err_del;
2751         }
2752         err = rtrs_clt_create_sysfs_root_files(clt);
2753         if (err) {
2754                 kobject_del(clt->kobj_paths);
2755                 kobject_put(clt->kobj_paths);
2756                 goto err_del;
2757         }
2758         dev_set_uevent_suppress(&clt->dev, false);
2759         kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2760
2761         return clt;
2762 err_del:
2763         device_del(&clt->dev);
2764 err_put:
2765         free_percpu(clt->pcpu_path);
2766         put_device(&clt->dev);
2767         return ERR_PTR(err);
2768 }
2769
2770 static void free_clt(struct rtrs_clt_sess *clt)
2771 {
2772         free_percpu(clt->pcpu_path);
2773
2774         /*
2775          * release callback will free clt and destroy mutexes in last put
2776          */
2777         device_unregister(&clt->dev);
2778 }
2779
2780 /**
2781  * rtrs_clt_open() - Open a path to an RTRS server
2782  * @ops: holds the link event callback and the private pointer.
2783  * @pathname: name of the path to an RTRS server
2784  * @paths: Paths to be established defined by their src and dst addresses
2785  * @paths_num: Number of elements in the @paths array
2786  * @port: port to be used by the RTRS session
2787  * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2788  * @reconnect_delay_sec: time between reconnect tries
2789  * @max_reconnect_attempts: Number of times to reconnect on error before giving
2790  *                          up, 0 for * disabled, -1 for forever
2791  * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
2792  *
2793  * Starts session establishment with the rtrs_server. The function can block
2794  * up to ~2000ms before it returns.
2795  *
2796  * Return a valid pointer on success otherwise PTR_ERR.
2797  */
2798 struct rtrs_clt_sess *rtrs_clt_open(struct rtrs_clt_ops *ops,
2799                                  const char *pathname,
2800                                  const struct rtrs_addr *paths,
2801                                  size_t paths_num, u16 port,
2802                                  size_t pdu_sz, u8 reconnect_delay_sec,
2803                                  s16 max_reconnect_attempts, u32 nr_poll_queues)
2804 {
2805         struct rtrs_clt_path *clt_path, *tmp;
2806         struct rtrs_clt_sess *clt;
2807         int err, i;
2808
2809         if (strchr(pathname, '/') || strchr(pathname, '.')) {
2810                 pr_err("pathname cannot contain / and .\n");
2811                 err = -EINVAL;
2812                 goto out;
2813         }
2814
2815         clt = alloc_clt(pathname, paths_num, port, pdu_sz, ops->priv,
2816                         ops->link_ev,
2817                         reconnect_delay_sec,
2818                         max_reconnect_attempts);
2819         if (IS_ERR(clt)) {
2820                 err = PTR_ERR(clt);
2821                 goto out;
2822         }
2823         for (i = 0; i < paths_num; i++) {
2824                 struct rtrs_clt_path *clt_path;
2825
2826                 clt_path = alloc_path(clt, &paths[i], nr_cpu_ids,
2827                                   nr_poll_queues);
2828                 if (IS_ERR(clt_path)) {
2829                         err = PTR_ERR(clt_path);
2830                         goto close_all_path;
2831                 }
2832                 if (!i)
2833                         clt_path->for_new_clt = 1;
2834                 list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2835
2836                 err = init_path(clt_path);
2837                 if (err) {
2838                         list_del_rcu(&clt_path->s.entry);
2839                         rtrs_clt_close_conns(clt_path, true);
2840                         free_percpu(clt_path->stats->pcpu_stats);
2841                         kfree(clt_path->stats);
2842                         free_path(clt_path);
2843                         goto close_all_path;
2844                 }
2845
2846                 err = rtrs_clt_create_path_files(clt_path);
2847                 if (err) {
2848                         list_del_rcu(&clt_path->s.entry);
2849                         rtrs_clt_close_conns(clt_path, true);
2850                         free_percpu(clt_path->stats->pcpu_stats);
2851                         kfree(clt_path->stats);
2852                         free_path(clt_path);
2853                         goto close_all_path;
2854                 }
2855         }
2856         err = alloc_permits(clt);
2857         if (err)
2858                 goto close_all_path;
2859
2860         return clt;
2861
2862 close_all_path:
2863         list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2864                 rtrs_clt_destroy_path_files(clt_path, NULL);
2865                 rtrs_clt_close_conns(clt_path, true);
2866                 kobject_put(&clt_path->kobj);
2867         }
2868         rtrs_clt_destroy_sysfs_root(clt);
2869         free_clt(clt);
2870
2871 out:
2872         return ERR_PTR(err);
2873 }
2874 EXPORT_SYMBOL(rtrs_clt_open);
2875
2876 /**
2877  * rtrs_clt_close() - Close a path
2878  * @clt: Session handle. Session is freed upon return.
2879  */
2880 void rtrs_clt_close(struct rtrs_clt_sess *clt)
2881 {
2882         struct rtrs_clt_path *clt_path, *tmp;
2883
2884         /* Firstly forbid sysfs access */
2885         rtrs_clt_destroy_sysfs_root(clt);
2886
2887         /* Now it is safe to iterate over all paths without locks */
2888         list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2889                 rtrs_clt_close_conns(clt_path, true);
2890                 rtrs_clt_destroy_path_files(clt_path, NULL);
2891                 kobject_put(&clt_path->kobj);
2892         }
2893         free_permits(clt);
2894         free_clt(clt);
2895 }
2896 EXPORT_SYMBOL(rtrs_clt_close);
2897
2898 int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_path *clt_path)
2899 {
2900         enum rtrs_clt_state old_state;
2901         int err = -EBUSY;
2902         bool changed;
2903
2904         changed = rtrs_clt_change_state_get_old(clt_path,
2905                                                  RTRS_CLT_RECONNECTING,
2906                                                  &old_state);
2907         if (changed) {
2908                 clt_path->reconnect_attempts = 0;
2909                 rtrs_clt_stop_and_destroy_conns(clt_path);
2910                 queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork, 0);
2911         }
2912         if (changed || old_state == RTRS_CLT_RECONNECTING) {
2913                 /*
2914                  * flush_delayed_work() queues pending work for immediate
2915                  * execution, so do the flush if we have queued something
2916                  * right now or work is pending.
2917                  */
2918                 flush_delayed_work(&clt_path->reconnect_dwork);
2919                 err = (READ_ONCE(clt_path->state) ==
2920                        RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2921         }
2922
2923         return err;
2924 }
2925
2926 int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_path *clt_path,
2927                                      const struct attribute *sysfs_self)
2928 {
2929         enum rtrs_clt_state old_state;
2930         bool changed;
2931
2932         /*
2933          * Continue stopping path till state was changed to DEAD or
2934          * state was observed as DEAD:
2935          * 1. State was changed to DEAD - we were fast and nobody
2936          *    invoked rtrs_clt_reconnect(), which can again start
2937          *    reconnecting.
2938          * 2. State was observed as DEAD - we have someone in parallel
2939          *    removing the path.
2940          */
2941         do {
2942                 rtrs_clt_close_conns(clt_path, true);
2943                 changed = rtrs_clt_change_state_get_old(clt_path,
2944                                                         RTRS_CLT_DEAD,
2945                                                         &old_state);
2946         } while (!changed && old_state != RTRS_CLT_DEAD);
2947
2948         if (changed) {
2949                 rtrs_clt_remove_path_from_arr(clt_path);
2950                 rtrs_clt_destroy_path_files(clt_path, sysfs_self);
2951                 kobject_put(&clt_path->kobj);
2952         }
2953
2954         return 0;
2955 }
2956
2957 void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt_sess *clt, int value)
2958 {
2959         clt->max_reconnect_attempts = (unsigned int)value;
2960 }
2961
2962 int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt_sess *clt)
2963 {
2964         return (int)clt->max_reconnect_attempts;
2965 }
2966
2967 /**
2968  * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2969  *
2970  * @dir:        READ/WRITE
2971  * @ops:        callback function to be called as confirmation, and the pointer.
2972  * @clt:        Session
2973  * @permit:     Preallocated permit
2974  * @vec:        Message that is sent to server together with the request.
2975  *              Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2976  *              Since the msg is copied internally it can be allocated on stack.
2977  * @nr:         Number of elements in @vec.
2978  * @data_len:   length of data sent to/from server
2979  * @sg:         Pages to be sent/received to/from server.
2980  * @sg_cnt:     Number of elements in the @sg
2981  *
2982  * Return:
2983  * 0:           Success
2984  * <0:          Error
2985  *
2986  * On dir=READ rtrs client will request a data transfer from Server to client.
2987  * The data that the server will respond with will be stored in @sg when
2988  * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2989  * On dir=WRITE rtrs client will rdma write data in sg to server side.
2990  */
2991 int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2992                      struct rtrs_clt_sess *clt, struct rtrs_permit *permit,
2993                      const struct kvec *vec, size_t nr, size_t data_len,
2994                      struct scatterlist *sg, unsigned int sg_cnt)
2995 {
2996         struct rtrs_clt_io_req *req;
2997         struct rtrs_clt_path *clt_path;
2998
2999         enum dma_data_direction dma_dir;
3000         int err = -ECONNABORTED, i;
3001         size_t usr_len, hdr_len;
3002         struct path_it it;
3003
3004         /* Get kvec length */
3005         for (i = 0, usr_len = 0; i < nr; i++)
3006                 usr_len += vec[i].iov_len;
3007
3008         if (dir == READ) {
3009                 hdr_len = sizeof(struct rtrs_msg_rdma_read) +
3010                           sg_cnt * sizeof(struct rtrs_sg_desc);
3011                 dma_dir = DMA_FROM_DEVICE;
3012         } else {
3013                 hdr_len = sizeof(struct rtrs_msg_rdma_write);
3014                 dma_dir = DMA_TO_DEVICE;
3015         }
3016
3017         rcu_read_lock();
3018         for (path_it_init(&it, clt);
3019              (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3020                 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3021                         continue;
3022
3023                 if (usr_len + hdr_len > clt_path->max_hdr_size) {
3024                         rtrs_wrn_rl(clt_path->clt,
3025                                      "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
3026                                      dir == READ ? "Read" : "Write",
3027                                      usr_len, hdr_len, clt_path->max_hdr_size);
3028                         err = -EMSGSIZE;
3029                         break;
3030                 }
3031                 req = rtrs_clt_get_req(clt_path, ops->conf_fn, permit, ops->priv,
3032                                        vec, usr_len, sg, sg_cnt, data_len,
3033                                        dma_dir);
3034                 if (dir == READ)
3035                         err = rtrs_clt_read_req(req);
3036                 else
3037                         err = rtrs_clt_write_req(req);
3038                 if (err) {
3039                         req->in_use = false;
3040                         continue;
3041                 }
3042                 /* Success path */
3043                 break;
3044         }
3045         path_it_deinit(&it);
3046         rcu_read_unlock();
3047
3048         return err;
3049 }
3050 EXPORT_SYMBOL(rtrs_clt_request);
3051
3052 int rtrs_clt_rdma_cq_direct(struct rtrs_clt_sess *clt, unsigned int index)
3053 {
3054         /* If no path, return -1 for block layer not to try again */
3055         int cnt = -1;
3056         struct rtrs_con *con;
3057         struct rtrs_clt_path *clt_path;
3058         struct path_it it;
3059
3060         rcu_read_lock();
3061         for (path_it_init(&it, clt);
3062              (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3063                 if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3064                         continue;
3065
3066                 con = clt_path->s.con[index + 1];
3067                 cnt = ib_process_cq_direct(con->cq, -1);
3068                 if (cnt)
3069                         break;
3070         }
3071         path_it_deinit(&it);
3072         rcu_read_unlock();
3073
3074         return cnt;
3075 }
3076 EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
3077
3078 /**
3079  * rtrs_clt_query() - queries RTRS session attributes
3080  *@clt: session pointer
3081  *@attr: query results for session attributes.
3082  * Returns:
3083  *    0 on success
3084  *    -ECOMM            no connection to the server
3085  */
3086 int rtrs_clt_query(struct rtrs_clt_sess *clt, struct rtrs_attrs *attr)
3087 {
3088         if (!rtrs_clt_is_connected(clt))
3089                 return -ECOMM;
3090
3091         attr->queue_depth      = clt->queue_depth;
3092         attr->max_segments     = clt->max_segments;
3093         /* Cap max_io_size to min of remote buffer size and the fr pages */
3094         attr->max_io_size = min_t(int, clt->max_io_size,
3095                                   clt->max_segments * SZ_4K);
3096
3097         return 0;
3098 }
3099 EXPORT_SYMBOL(rtrs_clt_query);
3100
3101 int rtrs_clt_create_path_from_sysfs(struct rtrs_clt_sess *clt,
3102                                      struct rtrs_addr *addr)
3103 {
3104         struct rtrs_clt_path *clt_path;
3105         int err;
3106
3107         clt_path = alloc_path(clt, addr, nr_cpu_ids, 0);
3108         if (IS_ERR(clt_path))
3109                 return PTR_ERR(clt_path);
3110
3111         mutex_lock(&clt->paths_mutex);
3112         if (clt->paths_num == 0) {
3113                 /*
3114                  * When all the paths are removed for a session,
3115                  * the addition of the first path is like a new session for
3116                  * the storage server
3117                  */
3118                 clt_path->for_new_clt = 1;
3119         }
3120
3121         mutex_unlock(&clt->paths_mutex);
3122
3123         /*
3124          * It is totally safe to add path in CONNECTING state: coming
3125          * IO will never grab it.  Also it is very important to add
3126          * path before init, since init fires LINK_CONNECTED event.
3127          */
3128         rtrs_clt_add_path_to_arr(clt_path);
3129
3130         err = init_path(clt_path);
3131         if (err)
3132                 goto close_path;
3133
3134         err = rtrs_clt_create_path_files(clt_path);
3135         if (err)
3136                 goto close_path;
3137
3138         return 0;
3139
3140 close_path:
3141         rtrs_clt_remove_path_from_arr(clt_path);
3142         rtrs_clt_close_conns(clt_path, true);
3143         free_percpu(clt_path->stats->pcpu_stats);
3144         kfree(clt_path->stats);
3145         free_path(clt_path);
3146
3147         return err;
3148 }
3149
3150 static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
3151 {
3152         if (!(dev->ib_dev->attrs.device_cap_flags &
3153               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
3154                 pr_err("Memory registrations not supported.\n");
3155                 return -ENOTSUPP;
3156         }
3157
3158         return 0;
3159 }
3160
3161 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
3162         .init = rtrs_clt_ib_dev_init
3163 };
3164
3165 static int __init rtrs_client_init(void)
3166 {
3167         rtrs_rdma_dev_pd_init(0, &dev_pd);
3168
3169         rtrs_clt_dev_class = class_create(THIS_MODULE, "rtrs-client");
3170         if (IS_ERR(rtrs_clt_dev_class)) {
3171                 pr_err("Failed to create rtrs-client dev class\n");
3172                 return PTR_ERR(rtrs_clt_dev_class);
3173         }
3174         rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
3175         if (!rtrs_wq) {
3176                 class_destroy(rtrs_clt_dev_class);
3177                 return -ENOMEM;
3178         }
3179
3180         return 0;
3181 }
3182
3183 static void __exit rtrs_client_exit(void)
3184 {
3185         destroy_workqueue(rtrs_wq);
3186         class_destroy(rtrs_clt_dev_class);
3187         rtrs_rdma_dev_pd_deinit(&dev_pd);
3188 }
3189
3190 module_init(rtrs_client_init);
3191 module_exit(rtrs_client_exit);