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
  20 #define RTRS_CONNECT_TIMEOUT_MS 30000
  21 /*
  22  * Wait a bit before trying to reconnect after a failure
  23  * in order to give server time to finish clean up which
  24  * leads to "false positives" failed reconnect attempts
  25  */
  26 #define RTRS_RECONNECT_BACKOFF 1000
  27 /*
  28  * Wait for additional random time between 0 and 8 seconds
  29  * before starting to reconnect to avoid clients reconnecting
  30  * all at once in case of a major network outage
  31  */
  32 #define RTRS_RECONNECT_SEED 8
  33
  34 #define FIRST_CONN 0x01
  35 /* limit to 128 * 4k = 512k max IO */
  36 #define RTRS_MAX_SEGMENTS          128
  37
  38 MODULE_DESCRIPTION("RDMA Transport Client");
  39 MODULE_LICENSE("GPL");
  40
  41 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
  42 static struct rtrs_rdma_dev_pd dev_pd = {
  43         .ops = &dev_pd_ops
  44 };
  45
  46 static struct workqueue_struct *rtrs_wq;
  47 static struct class *rtrs_clt_dev_class;
  48
  49 static inline bool rtrs_clt_is_connected(const struct rtrs_clt *clt)
  50 {
  51         struct rtrs_clt_sess *sess;
  52         bool connected = false;
  53
  54         rcu_read_lock();
  55         list_for_each_entry_rcu(sess, &clt->paths_list, s.entry)
  56                 connected |= READ_ONCE(sess->state) == RTRS_CLT_CONNECTED;
  57         rcu_read_unlock();
  58
  59         return connected;
  60 }
  61
  62 static struct rtrs_permit *
  63 __rtrs_get_permit(struct rtrs_clt *clt, enum rtrs_clt_con_type con_type)
  64 {
  65         size_t max_depth = clt->queue_depth;
  66         struct rtrs_permit *permit;
  67         int bit;
  68
  69         /*
  70          * Adapted from null_blk get_tag(). Callers from different cpus may
  71          * grab the same bit, since find_first_zero_bit is not atomic.
  72          * But then the test_and_set_bit_lock will fail for all the
  73          * callers but one, so that they will loop again.
  74          * This way an explicit spinlock is not required.
  75          */
  76         do {
  77                 bit = find_first_zero_bit(clt->permits_map, max_depth);
  78                 if (bit >= max_depth)
  79                         return NULL;
  80         } while (test_and_set_bit_lock(bit, clt->permits_map));
  81
  82         permit = get_permit(clt, bit);
  83         WARN_ON(permit->mem_id != bit);
  84         permit->cpu_id = raw_smp_processor_id();
  85         permit->con_type = con_type;
  86
  87         return permit;
  88 }
  89
  90 static inline void __rtrs_put_permit(struct rtrs_clt *clt,
  91                                       struct rtrs_permit *permit)
  92 {
  93         clear_bit_unlock(permit->mem_id, clt->permits_map);
  94 }
  95
  96 /**
  97  * rtrs_clt_get_permit() - allocates permit for future RDMA operation
  98  * @clt:        Current session
  99  * @con_type:   Type of connection to use with the permit
 100  * @can_wait:   Wait type
 101  *
 102  * Description:
 103  *    Allocates permit for the following RDMA operation.  Permit is used
 104  *    to preallocate all resources and to propagate memory pressure
 105  *    up earlier.
 106  *
 107  * Context:
 108  *    Can sleep if @wait == RTRS_PERMIT_WAIT
 109  */
 110 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt *clt,
 111                                           enum rtrs_clt_con_type con_type,
 112                                           enum wait_type can_wait)
 113 {
 114         struct rtrs_permit *permit;
 115         DEFINE_WAIT(wait);
 116
 117         permit = __rtrs_get_permit(clt, con_type);
 118         if (permit || !can_wait)
 119                 return permit;
 120
 121         do {
 122                 prepare_to_wait(&clt->permits_wait, &wait,
 123                                 TASK_UNINTERRUPTIBLE);
 124                 permit = __rtrs_get_permit(clt, con_type);
 125                 if (permit)
 126                         break;
 127
 128                 io_schedule();
 129         } while (1);
 130
 131         finish_wait(&clt->permits_wait, &wait);
 132
 133         return permit;
 134 }
 135 EXPORT_SYMBOL(rtrs_clt_get_permit);
 136
 137 /**
 138  * rtrs_clt_put_permit() - puts allocated permit
 139  * @clt:        Current session
 140  * @permit:     Permit to be freed
 141  *
 142  * Context:
 143  *    Does not matter
 144  */
 145 void rtrs_clt_put_permit(struct rtrs_clt *clt, struct rtrs_permit *permit)
 146 {
 147         if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
 148                 return;
 149
 150         __rtrs_put_permit(clt, permit);
 151
 152         /*
 153          * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
 154          * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
 155          * it must have added itself to &clt->permits_wait before
 156          * __rtrs_put_permit() finished.
 157          * Hence it is safe to guard wake_up() with a waitqueue_active() test.
 158          */
 159         if (waitqueue_active(&clt->permits_wait))
 160                 wake_up(&clt->permits_wait);
 161 }
 162 EXPORT_SYMBOL(rtrs_clt_put_permit);
 163
 164 /**
 165  * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
 166  * @sess: client session pointer
 167  * @permit: permit for the allocation of the RDMA buffer
 168  * Note:
 169  *     IO connection starts from 1.
 170  *     0 connection is for user messages.
 171  */
 172 static
 173 struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_sess *sess,
 174                                             struct rtrs_permit *permit)
 175 {
 176         int id = 0;
 177
 178         if (permit->con_type == RTRS_IO_CON)
 179                 id = (permit->cpu_id % (sess->s.irq_con_num - 1)) + 1;
 180
 181         return to_clt_con(sess->s.con[id]);
 182 }
 183
 184 /**
 185  * rtrs_clt_change_state() - change the session state through session state
 186  * machine.
 187  *
 188  * @sess: client session to change the state of.
 189  * @new_state: state to change to.
 190  *
 191  * returns true if sess's state is changed to new state, otherwise return false.
 192  *
 193  * Locks:
 194  * state_wq lock must be hold.
 195  */
 196 static bool rtrs_clt_change_state(struct rtrs_clt_sess *sess,
 197                                      enum rtrs_clt_state new_state)
 198 {
 199         enum rtrs_clt_state old_state;
 200         bool changed = false;
 201
 202         lockdep_assert_held(&sess->state_wq.lock);
 203
 204         old_state = sess->state;
 205         switch (new_state) {
 206         case RTRS_CLT_CONNECTING:
 207                 switch (old_state) {
 208                 case RTRS_CLT_RECONNECTING:
 209                         changed = true;
 210                         fallthrough;
 211                 default:
 212                         break;
 213                 }
 214                 break;
 215         case RTRS_CLT_RECONNECTING:
 216                 switch (old_state) {
 217                 case RTRS_CLT_CONNECTED:
 218                 case RTRS_CLT_CONNECTING_ERR:
 219                 case RTRS_CLT_CLOSED:
 220                         changed = true;
 221                         fallthrough;
 222                 default:
 223                         break;
 224                 }
 225                 break;
 226         case RTRS_CLT_CONNECTED:
 227                 switch (old_state) {
 228                 case RTRS_CLT_CONNECTING:
 229                         changed = true;
 230                         fallthrough;
 231                 default:
 232                         break;
 233                 }
 234                 break;
 235         case RTRS_CLT_CONNECTING_ERR:
 236                 switch (old_state) {
 237                 case RTRS_CLT_CONNECTING:
 238                         changed = true;
 239                         fallthrough;
 240                 default:
 241                         break;
 242                 }
 243                 break;
 244         case RTRS_CLT_CLOSING:
 245                 switch (old_state) {
 246                 case RTRS_CLT_CONNECTING:
 247                 case RTRS_CLT_CONNECTING_ERR:
 248                 case RTRS_CLT_RECONNECTING:
 249                 case RTRS_CLT_CONNECTED:
 250                         changed = true;
 251                         fallthrough;
 252                 default:
 253                         break;
 254                 }
 255                 break;
 256         case RTRS_CLT_CLOSED:
 257                 switch (old_state) {
 258                 case RTRS_CLT_CLOSING:
 259                         changed = true;
 260                         fallthrough;
 261                 default:
 262                         break;
 263                 }
 264                 break;
 265         case RTRS_CLT_DEAD:
 266                 switch (old_state) {
 267                 case RTRS_CLT_CLOSED:
 268                         changed = true;
 269                         fallthrough;
 270                 default:
 271                         break;
 272                 }
 273                 break;
 274         default:
 275                 break;
 276         }
 277         if (changed) {
 278                 sess->state = new_state;
 279                 wake_up_locked(&sess->state_wq);
 280         }
 281
 282         return changed;
 283 }
 284
 285 static bool rtrs_clt_change_state_from_to(struct rtrs_clt_sess *sess,
 286                                            enum rtrs_clt_state old_state,
 287                                            enum rtrs_clt_state new_state)
 288 {
 289         bool changed = false;
 290
 291         spin_lock_irq(&sess->state_wq.lock);
 292         if (sess->state == old_state)
 293                 changed = rtrs_clt_change_state(sess, new_state);
 294         spin_unlock_irq(&sess->state_wq.lock);
 295
 296         return changed;
 297 }
 298
 299 static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
 300 {
 301         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 302
 303         if (rtrs_clt_change_state_from_to(sess,
 304                                            RTRS_CLT_CONNECTED,
 305                                            RTRS_CLT_RECONNECTING)) {
 306                 struct rtrs_clt *clt = sess->clt;
 307                 unsigned int delay_ms;
 308
 309                 /*
 310                  * Normal scenario, reconnect if we were successfully connected
 311                  */
 312                 delay_ms = clt->reconnect_delay_sec * 1000;
 313                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork,
 314                                    msecs_to_jiffies(delay_ms +
 315                                                     prandom_u32() % RTRS_RECONNECT_SEED));
 316         } else {
 317                 /*
 318                  * Error can happen just on establishing new connection,
 319                  * so notify waiter with error state, waiter is responsible
 320                  * for cleaning the rest and reconnect if needed.
 321                  */
 322                 rtrs_clt_change_state_from_to(sess,
 323                                                RTRS_CLT_CONNECTING,
 324                                                RTRS_CLT_CONNECTING_ERR);
 325         }
 326 }
 327
 328 static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
 329 {
 330         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 331
 332         if (wc->status != IB_WC_SUCCESS) {
 333                 rtrs_err(con->c.sess, "Failed IB_WR_REG_MR: %s\n",
 334                           ib_wc_status_msg(wc->status));
 335                 rtrs_rdma_error_recovery(con);
 336         }
 337 }
 338
 339 static struct ib_cqe fast_reg_cqe = {
 340         .done = rtrs_clt_fast_reg_done
 341 };
 342
 343 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 344                               bool notify, bool can_wait);
 345
 346 static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 347 {
 348         struct rtrs_clt_io_req *req =
 349                 container_of(wc->wr_cqe, typeof(*req), inv_cqe);
 350         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 351
 352         if (wc->status != IB_WC_SUCCESS) {
 353                 rtrs_err(con->c.sess, "Failed IB_WR_LOCAL_INV: %s\n",
 354                           ib_wc_status_msg(wc->status));
 355                 rtrs_rdma_error_recovery(con);
 356         }
 357         req->need_inv = false;
 358         if (req->need_inv_comp)
 359                 complete(&req->inv_comp);
 360         else
 361                 /* Complete request from INV callback */
 362                 complete_rdma_req(req, req->inv_errno, true, false);
 363 }
 364
 365 static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
 366 {
 367         struct rtrs_clt_con *con = req->con;
 368         struct ib_send_wr wr = {
 369                 .opcode             = IB_WR_LOCAL_INV,
 370                 .wr_cqe             = &req->inv_cqe,
 371                 .send_flags         = IB_SEND_SIGNALED,
 372                 .ex.invalidate_rkey = req->mr->rkey,
 373         };
 374         req->inv_cqe.done = rtrs_clt_inv_rkey_done;
 375
 376         return ib_post_send(con->c.qp, &wr, NULL);
 377 }
 378
 379 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 380                               bool notify, bool can_wait)
 381 {
 382         struct rtrs_clt_con *con = req->con;
 383         struct rtrs_clt_sess *sess;
 384         int err;
 385
 386         if (WARN_ON(!req->in_use))
 387                 return;
 388         if (WARN_ON(!req->con))
 389                 return;
 390         sess = to_clt_sess(con->c.sess);
 391
 392         if (req->sg_cnt) {
 393                 if (req->dir == DMA_FROM_DEVICE && req->need_inv) {
 394                         /*
 395                          * We are here to invalidate read requests
 396                          * ourselves.  In normal scenario server should
 397                          * send INV for all read requests, but
 398                          * we are here, thus two things could happen:
 399                          *
 400                          *    1.  this is failover, when errno != 0
 401                          *        and can_wait == 1,
 402                          *
 403                          *    2.  something totally bad happened and
 404                          *        server forgot to send INV, so we
 405                          *        should do that ourselves.
 406                          */
 407
 408                         if (can_wait) {
 409                                 req->need_inv_comp = true;
 410                         } else {
 411                                 /* This should be IO path, so always notify */
 412                                 WARN_ON(!notify);
 413                                 /* Save errno for INV callback */
 414                                 req->inv_errno = errno;
 415                         }
 416
 417                         refcount_inc(&req->ref);
 418                         err = rtrs_inv_rkey(req);
 419                         if (err) {
 420                                 rtrs_err(con->c.sess, "Send INV WR key=%#x: %d\n",
 421                                           req->mr->rkey, err);
 422                         } else if (can_wait) {
 423                                 wait_for_completion(&req->inv_comp);
 424                         } else {
 425                                 /*
 426                                  * Something went wrong, so request will be
 427                                  * completed from INV callback.
 428                                  */
 429                                 WARN_ON_ONCE(1);
 430
 431                                 return;
 432                         }
 433                         if (!refcount_dec_and_test(&req->ref))
 434                                 return;
 435                 }
 436                 ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
 437                                 req->sg_cnt, req->dir);
 438         }
 439         if (!refcount_dec_and_test(&req->ref))
 440                 return;
 441         if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
 442                 atomic_dec(&sess->stats->inflight);
 443
 444         req->in_use = false;
 445         req->con = NULL;
 446
 447         if (errno) {
 448                 rtrs_err_rl(con->c.sess, "IO request failed: error=%d path=%s [%s:%u] notify=%d\n",
 449                             errno, kobject_name(&sess->kobj), sess->hca_name,
 450                             sess->hca_port, notify);
 451         }
 452
 453         if (notify)
 454                 req->conf(req->priv, errno);
 455 }
 456
 457 static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
 458                                 struct rtrs_clt_io_req *req,
 459                                 struct rtrs_rbuf *rbuf, u32 off,
 460                                 u32 imm, struct ib_send_wr *wr)
 461 {
 462         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 463         enum ib_send_flags flags;
 464         struct ib_sge sge;
 465
 466         if (!req->sg_size) {
 467                 rtrs_wrn(con->c.sess,
 468                          "Doing RDMA Write failed, no data supplied\n");
 469                 return -EINVAL;
 470         }
 471
 472         /* user data and user message in the first list element */
 473         sge.addr   = req->iu->dma_addr;
 474         sge.length = req->sg_size;
 475         sge.lkey   = sess->s.dev->ib_pd->local_dma_lkey;
 476
 477         /*
 478          * From time to time we have to post signalled sends,
 479          * or send queue will fill up and only QP reset can help.
 480          */
 481         flags = atomic_inc_return(&con->c.wr_cnt) % sess->s.signal_interval ?
 482                         0 : IB_SEND_SIGNALED;
 483
 484         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, 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_sess *sess, u32 msg_id,
 493                            s16 errno, bool w_inval)
 494 {
 495         struct rtrs_clt_io_req *req;
 496
 497         if (WARN_ON(msg_id >= sess->queue_depth))
 498                 return;
 499
 500         req = &sess->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_sess *sess = to_clt_sess(con->c.sess);
 511
 512         WARN_ON((sess->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.sess, "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_sess *sess = to_clt_sess(con->c.sess);
 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((sess->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.sess, "rkey response is malformed: size %d\n",
 538                           wc->byte_len);
 539                 goto out;
 540         }
 541         ib_dma_sync_single_for_cpu(sess->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(sess->clt, "rkey response is malformed: type %d\n",
 546                           le16_to_cpu(msg->type));
 547                 goto out;
 548         }
 549         buf_id = le16_to_cpu(msg->buf_id);
 550         if (WARN_ON(buf_id >= sess->queue_depth))
 551                 goto out;
 552
 553         rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
 554         if (imm_type == RTRS_IO_RSP_IMM ||
 555             imm_type == RTRS_IO_RSP_W_INV_IMM) {
 556                 u32 msg_id;
 557
 558                 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 559                 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 560
 561                 if (WARN_ON(buf_id != msg_id))
 562                         goto out;
 563                 sess->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
 564                 process_io_rsp(sess, msg_id, err, w_inval);
 565         }
 566         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, iu->dma_addr,
 567                                       iu->size, DMA_FROM_DEVICE);
 568         return rtrs_clt_recv_done(con, wc);
 569 out:
 570         rtrs_rdma_error_recovery(con);
 571 }
 572
 573 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
 574
 575 static struct ib_cqe io_comp_cqe = {
 576         .done = rtrs_clt_rdma_done
 577 };
 578
 579 /*
 580  * Post x2 empty WRs: first is for this RDMA with IMM,
 581  * second is for RECV with INV, which happened earlier.
 582  */
 583 static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
 584 {
 585         struct ib_recv_wr wr_arr[2], *wr;
 586         int i;
 587
 588         memset(wr_arr, 0, sizeof(wr_arr));
 589         for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
 590                 wr = &wr_arr[i];
 591                 wr->wr_cqe  = cqe;
 592                 if (i)
 593                         /* Chain backwards */
 594                         wr->next = &wr_arr[i - 1];
 595         }
 596
 597         return ib_post_recv(con->qp, wr, NULL);
 598 }
 599
 600 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
 601 {
 602         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 603         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 604         u32 imm_type, imm_payload;
 605         bool w_inval = false;
 606         int err;
 607
 608         if (wc->status != IB_WC_SUCCESS) {
 609                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 610                         rtrs_err(sess->clt, "RDMA failed: %s\n",
 611                                   ib_wc_status_msg(wc->status));
 612                         rtrs_rdma_error_recovery(con);
 613                 }
 614                 return;
 615         }
 616         rtrs_clt_update_wc_stats(con);
 617
 618         switch (wc->opcode) {
 619         case IB_WC_RECV_RDMA_WITH_IMM:
 620                 /*
 621                  * post_recv() RDMA write completions of IO reqs (read/write)
 622                  * and hb
 623                  */
 624                 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
 625                         return;
 626                 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
 627                                &imm_type, &imm_payload);
 628                 if (imm_type == RTRS_IO_RSP_IMM ||
 629                     imm_type == RTRS_IO_RSP_W_INV_IMM) {
 630                         u32 msg_id;
 631
 632                         w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 633                         rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 634
 635                         process_io_rsp(sess, msg_id, err, w_inval);
 636                 } else if (imm_type == RTRS_HB_MSG_IMM) {
 637                         WARN_ON(con->c.cid);
 638                         rtrs_send_hb_ack(&sess->s);
 639                         if (sess->flags & RTRS_MSG_NEW_RKEY_F)
 640                                 return  rtrs_clt_recv_done(con, wc);
 641                 } else if (imm_type == RTRS_HB_ACK_IMM) {
 642                         WARN_ON(con->c.cid);
 643                         sess->s.hb_missed_cnt = 0;
 644                         sess->s.hb_cur_latency =
 645                                 ktime_sub(ktime_get(), sess->s.hb_last_sent);
 646                         if (sess->flags & RTRS_MSG_NEW_RKEY_F)
 647                                 return  rtrs_clt_recv_done(con, wc);
 648                 } else {
 649                         rtrs_wrn(con->c.sess, "Unknown IMM type %u\n",
 650                                   imm_type);
 651                 }
 652                 if (w_inval)
 653                         /*
 654                          * Post x2 empty WRs: first is for this RDMA with IMM,
 655                          * second is for RECV with INV, which happened earlier.
 656                          */
 657                         err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
 658                 else
 659                         err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 660                 if (err) {
 661                         rtrs_err(con->c.sess, "rtrs_post_recv_empty(): %d\n",
 662                                   err);
 663                         rtrs_rdma_error_recovery(con);
 664                 }
 665                 break;
 666         case IB_WC_RECV:
 667                 /*
 668                  * Key invalidations from server side
 669                  */
 670                 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
 671                           wc->wc_flags & IB_WC_WITH_IMM));
 672                 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
 673                 if (sess->flags & RTRS_MSG_NEW_RKEY_F) {
 674                         if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
 675                                 return  rtrs_clt_recv_done(con, wc);
 676
 677                         return  rtrs_clt_rkey_rsp_done(con, wc);
 678                 }
 679                 break;
 680         case IB_WC_RDMA_WRITE:
 681                 /*
 682                  * post_send() RDMA write completions of IO reqs (read/write)
 683                  * and hb.
 684                  */
 685                 break;
 686
 687         default:
 688                 rtrs_wrn(sess->clt, "Unexpected WC type: %d\n", wc->opcode);
 689                 return;
 690         }
 691 }
 692
 693 static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
 694 {
 695         int err, i;
 696         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 697
 698         for (i = 0; i < q_size; i++) {
 699                 if (sess->flags & RTRS_MSG_NEW_RKEY_F) {
 700                         struct rtrs_iu *iu = &con->rsp_ius[i];
 701
 702                         err = rtrs_iu_post_recv(&con->c, iu);
 703                 } else {
 704                         err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 705                 }
 706                 if (err)
 707                         return err;
 708         }
 709
 710         return 0;
 711 }
 712
 713 static int post_recv_sess(struct rtrs_clt_sess *sess)
 714 {
 715         size_t q_size = 0;
 716         int err, cid;
 717
 718         for (cid = 0; cid < sess->s.con_num; cid++) {
 719                 if (cid == 0)
 720                         q_size = SERVICE_CON_QUEUE_DEPTH;
 721                 else
 722                         q_size = sess->queue_depth;
 723
 724                 /*
 725                  * x2 for RDMA read responses + FR key invalidations,
 726                  * RDMA writes do not require any FR registrations.
 727                  */
 728                 q_size *= 2;
 729
 730                 err = post_recv_io(to_clt_con(sess->s.con[cid]), q_size);
 731                 if (err) {
 732                         rtrs_err(sess->clt, "post_recv_io(), err: %d\n", err);
 733                         return err;
 734                 }
 735         }
 736
 737         return 0;
 738 }
 739
 740 struct path_it {
 741         int i;
 742         struct list_head skip_list;
 743         struct rtrs_clt *clt;
 744         struct rtrs_clt_sess *(*next_path)(struct path_it *it);
 745 };
 746
 747 /**
 748  * list_next_or_null_rr_rcu - get next list element in round-robin fashion.
 749  * @head:       the head for the list.
 750  * @ptr:        the list head to take the next element from.
 751  * @type:       the type of the struct this is embedded in.
 752  * @memb:       the name of the list_head within the struct.
 753  *
 754  * Next element 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 primitive 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 #define list_next_or_null_rr_rcu(head, ptr, type, memb) \
 761 ({ \
 762         list_next_or_null_rcu(head, ptr, type, memb) ?: \
 763                 list_next_or_null_rcu(head, READ_ONCE((ptr)->next), \
 764                                       type, memb); \
 765 })
 766
 767 /**
 768  * get_next_path_rr() - Returns path in round-robin fashion.
 769  * @it: the path pointer
 770  *
 771  * Related to @MP_POLICY_RR
 772  *
 773  * Locks:
 774  *    rcu_read_lock() must be hold.
 775  */
 776 static struct rtrs_clt_sess *get_next_path_rr(struct path_it *it)
 777 {
 778         struct rtrs_clt_sess __rcu **ppcpu_path;
 779         struct rtrs_clt_sess *path;
 780         struct rtrs_clt *clt;
 781
 782         clt = it->clt;
 783
 784         /*
 785          * Here we use two RCU objects: @paths_list and @pcpu_path
 786          * pointer.  See rtrs_clt_remove_path_from_arr() for details
 787          * how that is handled.
 788          */
 789
 790         ppcpu_path = this_cpu_ptr(clt->pcpu_path);
 791         path = rcu_dereference(*ppcpu_path);
 792         if (!path)
 793                 path = list_first_or_null_rcu(&clt->paths_list,
 794                                               typeof(*path), s.entry);
 795         else
 796                 path = list_next_or_null_rr_rcu(&clt->paths_list,
 797                                                 &path->s.entry,
 798                                                 typeof(*path),
 799                                                 s.entry);
 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_sess *get_next_path_min_inflight(struct path_it *it)
 815 {
 816         struct rtrs_clt_sess *min_path = NULL;
 817         struct rtrs_clt *clt = it->clt;
 818         struct rtrs_clt_sess *sess;
 819         int min_inflight = INT_MAX;
 820         int inflight;
 821
 822         list_for_each_entry_rcu(sess, &clt->paths_list, s.entry) {
 823                 if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTED)
 824                         continue;
 825
 826                 if (!list_empty(raw_cpu_ptr(sess->mp_skip_entry)))
 827                         continue;
 828
 829                 inflight = atomic_read(&sess->stats->inflight);
 830
 831                 if (inflight < min_inflight) {
 832                         min_inflight = inflight;
 833                         min_path = sess;
 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_sess *get_next_path_min_latency(struct path_it *it)
 866 {
 867         struct rtrs_clt_sess *min_path = NULL;
 868         struct rtrs_clt *clt = it->clt;
 869         struct rtrs_clt_sess *sess;
 870         ktime_t min_latency = INT_MAX;
 871         ktime_t latency;
 872
 873         list_for_each_entry_rcu(sess, &clt->paths_list, s.entry) {
 874                 if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTED)
 875                         continue;
 876
 877                 if (!list_empty(raw_cpu_ptr(sess->mp_skip_entry)))
 878                         continue;
 879
 880                 latency = sess->s.hb_cur_latency;
 881
 882                 if (latency < min_latency) {
 883                         min_latency = latency;
 884                         min_path = sess;
 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 *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 policy.
 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  * @sess: client session
 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_sess *sess,
 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(sess, 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
 968         iov_iter_kvec(&iter, READ, vec, 1, usr_len);
 969         len = _copy_from_iter(req->iu->buf, usr_len, &iter);
 970         WARN_ON(len != usr_len);
 971
 972         reinit_completion(&req->inv_comp);
 973 }
 974
 975 static struct rtrs_clt_io_req *
 976 rtrs_clt_get_req(struct rtrs_clt_sess *sess,
 977                  void (*conf)(void *priv, int errno),
 978                  struct rtrs_permit *permit, void *priv,
 979                  const struct kvec *vec, size_t usr_len,
 980                  struct scatterlist *sg, size_t sg_cnt,
 981                  size_t data_len, int dir)
 982 {
 983         struct rtrs_clt_io_req *req;
 984
 985         req = &sess->reqs[permit->mem_id];
 986         rtrs_clt_init_req(req, sess, conf, permit, priv, vec, usr_len,
 987                            sg, sg_cnt, data_len, dir);
 988         return req;
 989 }
 990
 991 static struct rtrs_clt_io_req *
 992 rtrs_clt_get_copy_req(struct rtrs_clt_sess *alive_sess,
 993                        struct rtrs_clt_io_req *fail_req)
 994 {
 995         struct rtrs_clt_io_req *req;
 996         struct kvec vec = {
 997                 .iov_base = fail_req->iu->buf,
 998                 .iov_len  = fail_req->usr_len
 999         };
1000
1001         req = &alive_sess->reqs[fail_req->permit->mem_id];
1002         rtrs_clt_init_req(req, alive_sess, fail_req->conf, fail_req->permit,
1003                            fail_req->priv, &vec, fail_req->usr_len,
1004                            fail_req->sglist, fail_req->sg_cnt,
1005                            fail_req->data_len, fail_req->dir);
1006         return req;
1007 }
1008
1009 static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
1010                                    struct rtrs_clt_io_req *req,
1011                                    struct rtrs_rbuf *rbuf, bool fr_en,
1012                                    u32 size, u32 imm, struct ib_send_wr *wr,
1013                                    struct ib_send_wr *tail)
1014 {
1015         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1016         struct ib_sge *sge = req->sge;
1017         enum ib_send_flags flags;
1018         struct scatterlist *sg;
1019         size_t num_sge;
1020         int i;
1021         struct ib_send_wr *ptail = NULL;
1022
1023         if (fr_en) {
1024                 i = 0;
1025                 sge[i].addr   = req->mr->iova;
1026                 sge[i].length = req->mr->length;
1027                 sge[i].lkey   = req->mr->lkey;
1028                 i++;
1029                 num_sge = 2;
1030                 ptail = tail;
1031         } else {
1032                 for_each_sg(req->sglist, sg, req->sg_cnt, i) {
1033                         sge[i].addr   = sg_dma_address(sg);
1034                         sge[i].length = sg_dma_len(sg);
1035                         sge[i].lkey   = sess->s.dev->ib_pd->local_dma_lkey;
1036                 }
1037                 num_sge = 1 + req->sg_cnt;
1038         }
1039         sge[i].addr   = req->iu->dma_addr;
1040         sge[i].length = size;
1041         sge[i].lkey   = sess->s.dev->ib_pd->local_dma_lkey;
1042
1043         /*
1044          * From time to time we have to post signalled sends,
1045          * or send queue will fill up and only QP reset can help.
1046          */
1047         flags = atomic_inc_return(&con->c.wr_cnt) % sess->s.signal_interval ?
1048                         0 : IB_SEND_SIGNALED;
1049
1050         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, req->iu->dma_addr,
1051                                       size, DMA_TO_DEVICE);
1052
1053         return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
1054                                             rbuf->rkey, rbuf->addr, imm,
1055                                             flags, wr, ptail);
1056 }
1057
1058 static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1059 {
1060         int nr;
1061
1062         /* Align the MR to a 4K page size to match the block virt boundary */
1063         nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1064         if (nr < 0)
1065                 return nr;
1066         if (nr < req->sg_cnt)
1067                 return -EINVAL;
1068         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1069
1070         return nr;
1071 }
1072
1073 static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
1074 {
1075         struct rtrs_clt_con *con = req->con;
1076         struct rtrs_sess *s = con->c.sess;
1077         struct rtrs_clt_sess *sess = to_clt_sess(s);
1078         struct rtrs_msg_rdma_write *msg;
1079
1080         struct rtrs_rbuf *rbuf;
1081         int ret, count = 0;
1082         u32 imm, buf_id;
1083         struct ib_reg_wr rwr;
1084         struct ib_send_wr inv_wr;
1085         struct ib_send_wr *wr = NULL;
1086         bool fr_en = false;
1087
1088         const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1089
1090         if (tsize > sess->chunk_size) {
1091                 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
1092                           tsize, sess->chunk_size);
1093                 return -EMSGSIZE;
1094         }
1095         if (req->sg_cnt) {
1096                 count = ib_dma_map_sg(sess->s.dev->ib_dev, req->sglist,
1097                                       req->sg_cnt, req->dir);
1098                 if (!count) {
1099                         rtrs_wrn(s, "Write request failed, map failed\n");
1100                         return -EINVAL;
1101                 }
1102         }
1103         /* put rtrs msg after sg and user message */
1104         msg = req->iu->buf + req->usr_len;
1105         msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1106         msg->usr_len = cpu_to_le16(req->usr_len);
1107
1108         /* rtrs message on server side will be after user data and message */
1109         imm = req->permit->mem_off + req->data_len + req->usr_len;
1110         imm = rtrs_to_io_req_imm(imm);
1111         buf_id = req->permit->mem_id;
1112         req->sg_size = tsize;
1113         rbuf = &sess->rbufs[buf_id];
1114
1115         if (count) {
1116                 ret = rtrs_map_sg_fr(req, count);
1117                 if (ret < 0) {
1118                         rtrs_err_rl(s,
1119                                     "Write request failed, failed to map fast reg. data, err: %d\n",
1120                                     ret);
1121                         ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
1122                                         req->sg_cnt, req->dir);
1123                         return ret;
1124                 }
1125                 inv_wr = (struct ib_send_wr) {
1126                         .opcode             = IB_WR_LOCAL_INV,
1127                         .wr_cqe             = &req->inv_cqe,
1128                         .send_flags         = IB_SEND_SIGNALED,
1129                         .ex.invalidate_rkey = req->mr->rkey,
1130                 };
1131                 req->inv_cqe.done = rtrs_clt_inv_rkey_done;
1132                 rwr = (struct ib_reg_wr) {
1133                         .wr.opcode = IB_WR_REG_MR,
1134                         .wr.wr_cqe = &fast_reg_cqe,
1135                         .mr = req->mr,
1136                         .key = req->mr->rkey,
1137                         .access = (IB_ACCESS_LOCAL_WRITE),
1138                 };
1139                 wr = &rwr.wr;
1140                 fr_en = true;
1141                 refcount_inc(&req->ref);
1142         }
1143         /*
1144          * Update stats now, after request is successfully sent it is not
1145          * safe anymore to touch it.
1146          */
1147         rtrs_clt_update_all_stats(req, WRITE);
1148
1149         ret = rtrs_post_rdma_write_sg(req->con, req, rbuf, fr_en,
1150                                       req->usr_len + sizeof(*msg),
1151                                       imm, wr, &inv_wr);
1152         if (ret) {
1153                 rtrs_err_rl(s,
1154                             "Write request failed: error=%d path=%s [%s:%u]\n",
1155                             ret, kobject_name(&sess->kobj), sess->hca_name,
1156                             sess->hca_port);
1157                 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
1158                         atomic_dec(&sess->stats->inflight);
1159                 if (req->sg_cnt)
1160                         ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
1161                                         req->sg_cnt, req->dir);
1162         }
1163
1164         return ret;
1165 }
1166
1167 static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1168 {
1169         struct rtrs_clt_con *con = req->con;
1170         struct rtrs_sess *s = con->c.sess;
1171         struct rtrs_clt_sess *sess = to_clt_sess(s);
1172         struct rtrs_msg_rdma_read *msg;
1173         struct rtrs_ib_dev *dev = sess->s.dev;
1174
1175         struct ib_reg_wr rwr;
1176         struct ib_send_wr *wr = NULL;
1177
1178         int ret, count = 0;
1179         u32 imm, buf_id;
1180
1181         const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1182
1183         if (tsize > sess->chunk_size) {
1184                 rtrs_wrn(s,
1185                           "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1186                           tsize, sess->chunk_size);
1187                 return -EMSGSIZE;
1188         }
1189
1190         if (req->sg_cnt) {
1191                 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1192                                       req->dir);
1193                 if (!count) {
1194                         rtrs_wrn(s,
1195                                   "Read request failed, dma map failed\n");
1196                         return -EINVAL;
1197                 }
1198         }
1199         /* put our message into req->buf after user message*/
1200         msg = req->iu->buf + req->usr_len;
1201         msg->type = cpu_to_le16(RTRS_MSG_READ);
1202         msg->usr_len = cpu_to_le16(req->usr_len);
1203
1204         if (count) {
1205                 ret = rtrs_map_sg_fr(req, count);
1206                 if (ret < 0) {
1207                         rtrs_err_rl(s,
1208                                      "Read request failed, failed to map  fast reg. data, err: %d\n",
1209                                      ret);
1210                         ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1211                                         req->dir);
1212                         return ret;
1213                 }
1214                 rwr = (struct ib_reg_wr) {
1215                         .wr.opcode = IB_WR_REG_MR,
1216                         .wr.wr_cqe = &fast_reg_cqe,
1217                         .mr = req->mr,
1218                         .key = req->mr->rkey,
1219                         .access = (IB_ACCESS_LOCAL_WRITE |
1220                                    IB_ACCESS_REMOTE_WRITE),
1221                 };
1222                 wr = &rwr.wr;
1223
1224                 msg->sg_cnt = cpu_to_le16(1);
1225                 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1226
1227                 msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1228                 msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1229                 msg->desc[0].len = cpu_to_le32(req->mr->length);
1230
1231                 /* Further invalidation is required */
1232                 req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1233
1234         } else {
1235                 msg->sg_cnt = 0;
1236                 msg->flags = 0;
1237         }
1238         /*
1239          * rtrs message will be after the space reserved for disk data and
1240          * user message
1241          */
1242         imm = req->permit->mem_off + req->data_len + req->usr_len;
1243         imm = rtrs_to_io_req_imm(imm);
1244         buf_id = req->permit->mem_id;
1245
1246         req->sg_size  = sizeof(*msg);
1247         req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1248         req->sg_size += req->usr_len;
1249
1250         /*
1251          * Update stats now, after request is successfully sent it is not
1252          * safe anymore to touch it.
1253          */
1254         rtrs_clt_update_all_stats(req, READ);
1255
1256         ret = rtrs_post_send_rdma(req->con, req, &sess->rbufs[buf_id],
1257                                    req->data_len, imm, wr);
1258         if (ret) {
1259                 rtrs_err_rl(s,
1260                             "Read request failed: error=%d path=%s [%s:%u]\n",
1261                             ret, kobject_name(&sess->kobj), sess->hca_name,
1262                             sess->hca_port);
1263                 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
1264                         atomic_dec(&sess->stats->inflight);
1265                 req->need_inv = false;
1266                 if (req->sg_cnt)
1267                         ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1268                                         req->sg_cnt, req->dir);
1269         }
1270
1271         return ret;
1272 }
1273
1274 /**
1275  * rtrs_clt_failover_req() - Try to find an active path for a failed request
1276  * @clt: clt context
1277  * @fail_req: a failed io request.
1278  */
1279 static int rtrs_clt_failover_req(struct rtrs_clt *clt,
1280                                  struct rtrs_clt_io_req *fail_req)
1281 {
1282         struct rtrs_clt_sess *alive_sess;
1283         struct rtrs_clt_io_req *req;
1284         int err = -ECONNABORTED;
1285         struct path_it it;
1286
1287         rcu_read_lock();
1288         for (path_it_init(&it, clt);
1289              (alive_sess = it.next_path(&it)) && it.i < it.clt->paths_num;
1290              it.i++) {
1291                 if (READ_ONCE(alive_sess->state) != RTRS_CLT_CONNECTED)
1292                         continue;
1293                 req = rtrs_clt_get_copy_req(alive_sess, fail_req);
1294                 if (req->dir == DMA_TO_DEVICE)
1295                         err = rtrs_clt_write_req(req);
1296                 else
1297                         err = rtrs_clt_read_req(req);
1298                 if (err) {
1299                         req->in_use = false;
1300                         continue;
1301                 }
1302                 /* Success path */
1303                 rtrs_clt_inc_failover_cnt(alive_sess->stats);
1304                 break;
1305         }
1306         path_it_deinit(&it);
1307         rcu_read_unlock();
1308
1309         return err;
1310 }
1311
1312 static void fail_all_outstanding_reqs(struct rtrs_clt_sess *sess)
1313 {
1314         struct rtrs_clt *clt = sess->clt;
1315         struct rtrs_clt_io_req *req;
1316         int i, err;
1317
1318         if (!sess->reqs)
1319                 return;
1320         for (i = 0; i < sess->queue_depth; ++i) {
1321                 req = &sess->reqs[i];
1322                 if (!req->in_use)
1323                         continue;
1324
1325                 /*
1326                  * Safely (without notification) complete failed request.
1327                  * After completion this request is still useble and can
1328                  * be failovered to another path.
1329                  */
1330                 complete_rdma_req(req, -ECONNABORTED, false, true);
1331
1332                 err = rtrs_clt_failover_req(clt, req);
1333                 if (err)
1334                         /* Failover failed, notify anyway */
1335                         req->conf(req->priv, err);
1336         }
1337 }
1338
1339 static void free_sess_reqs(struct rtrs_clt_sess *sess)
1340 {
1341         struct rtrs_clt_io_req *req;
1342         int i;
1343
1344         if (!sess->reqs)
1345                 return;
1346         for (i = 0; i < sess->queue_depth; ++i) {
1347                 req = &sess->reqs[i];
1348                 if (req->mr)
1349                         ib_dereg_mr(req->mr);
1350                 kfree(req->sge);
1351                 rtrs_iu_free(req->iu, sess->s.dev->ib_dev, 1);
1352         }
1353         kfree(sess->reqs);
1354         sess->reqs = NULL;
1355 }
1356
1357 static int alloc_sess_reqs(struct rtrs_clt_sess *sess)
1358 {
1359         struct rtrs_clt_io_req *req;
1360         int i, err = -ENOMEM;
1361
1362         sess->reqs = kcalloc(sess->queue_depth, sizeof(*sess->reqs),
1363                              GFP_KERNEL);
1364         if (!sess->reqs)
1365                 return -ENOMEM;
1366
1367         for (i = 0; i < sess->queue_depth; ++i) {
1368                 req = &sess->reqs[i];
1369                 req->iu = rtrs_iu_alloc(1, sess->max_hdr_size, GFP_KERNEL,
1370                                          sess->s.dev->ib_dev,
1371                                          DMA_TO_DEVICE,
1372                                          rtrs_clt_rdma_done);
1373                 if (!req->iu)
1374                         goto out;
1375
1376                 req->sge = kcalloc(2, sizeof(*req->sge), GFP_KERNEL);
1377                 if (!req->sge)
1378                         goto out;
1379
1380                 req->mr = ib_alloc_mr(sess->s.dev->ib_pd, IB_MR_TYPE_MEM_REG,
1381                                       sess->max_pages_per_mr);
1382                 if (IS_ERR(req->mr)) {
1383                         err = PTR_ERR(req->mr);
1384                         req->mr = NULL;
1385                         pr_err("Failed to alloc sess->max_pages_per_mr %d\n",
1386                                sess->max_pages_per_mr);
1387                         goto out;
1388                 }
1389
1390                 init_completion(&req->inv_comp);
1391         }
1392
1393         return 0;
1394
1395 out:
1396         free_sess_reqs(sess);
1397
1398         return err;
1399 }
1400
1401 static int alloc_permits(struct rtrs_clt *clt)
1402 {
1403         unsigned int chunk_bits;
1404         int err, i;
1405
1406         clt->permits_map = kcalloc(BITS_TO_LONGS(clt->queue_depth),
1407                                    sizeof(long), GFP_KERNEL);
1408         if (!clt->permits_map) {
1409                 err = -ENOMEM;
1410                 goto out_err;
1411         }
1412         clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1413         if (!clt->permits) {
1414                 err = -ENOMEM;
1415                 goto err_map;
1416         }
1417         chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1418         for (i = 0; i < clt->queue_depth; i++) {
1419                 struct rtrs_permit *permit;
1420
1421                 permit = get_permit(clt, i);
1422                 permit->mem_id = i;
1423                 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1424         }
1425
1426         return 0;
1427
1428 err_map:
1429         kfree(clt->permits_map);
1430         clt->permits_map = NULL;
1431 out_err:
1432         return err;
1433 }
1434
1435 static void free_permits(struct rtrs_clt *clt)
1436 {
1437         if (clt->permits_map) {
1438                 size_t sz = clt->queue_depth;
1439
1440                 wait_event(clt->permits_wait,
1441                            find_first_bit(clt->permits_map, sz) >= sz);
1442         }
1443         kfree(clt->permits_map);
1444         clt->permits_map = NULL;
1445         kfree(clt->permits);
1446         clt->permits = NULL;
1447 }
1448
1449 static void query_fast_reg_mode(struct rtrs_clt_sess *sess)
1450 {
1451         struct ib_device *ib_dev;
1452         u64 max_pages_per_mr;
1453         int mr_page_shift;
1454
1455         ib_dev = sess->s.dev->ib_dev;
1456
1457         /*
1458          * Use the smallest page size supported by the HCA, down to a
1459          * minimum of 4096 bytes. We're unlikely to build large sglists
1460          * out of smaller entries.
1461          */
1462         mr_page_shift      = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1463         max_pages_per_mr   = ib_dev->attrs.max_mr_size;
1464         do_div(max_pages_per_mr, (1ull << mr_page_shift));
1465         sess->max_pages_per_mr =
1466                 min3(sess->max_pages_per_mr, (u32)max_pages_per_mr,
1467                      ib_dev->attrs.max_fast_reg_page_list_len);
1468         sess->clt->max_segments =
1469                 min(sess->max_pages_per_mr, sess->clt->max_segments);
1470 }
1471
1472 static bool rtrs_clt_change_state_get_old(struct rtrs_clt_sess *sess,
1473                                            enum rtrs_clt_state new_state,
1474                                            enum rtrs_clt_state *old_state)
1475 {
1476         bool changed;
1477
1478         spin_lock_irq(&sess->state_wq.lock);
1479         if (old_state)
1480                 *old_state = sess->state;
1481         changed = rtrs_clt_change_state(sess, new_state);
1482         spin_unlock_irq(&sess->state_wq.lock);
1483
1484         return changed;
1485 }
1486
1487 static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1488 {
1489         struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1490
1491         rtrs_rdma_error_recovery(con);
1492 }
1493
1494 static void rtrs_clt_init_hb(struct rtrs_clt_sess *sess)
1495 {
1496         rtrs_init_hb(&sess->s, &io_comp_cqe,
1497                       RTRS_HB_INTERVAL_MS,
1498                       RTRS_HB_MISSED_MAX,
1499                       rtrs_clt_hb_err_handler,
1500                       rtrs_wq);
1501 }
1502
1503 static void rtrs_clt_reconnect_work(struct work_struct *work);
1504 static void rtrs_clt_close_work(struct work_struct *work);
1505
1506 static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
1507                                         const struct rtrs_addr *path,
1508                                         size_t con_num, u32 nr_poll_queues)
1509 {
1510         struct rtrs_clt_sess *sess;
1511         int err = -ENOMEM;
1512         int cpu;
1513         size_t total_con;
1514
1515         sess = kzalloc(sizeof(*sess), GFP_KERNEL);
1516         if (!sess)
1517                 goto err;
1518
1519         /*
1520          * irqmode and poll
1521          * +1: Extra connection for user messages
1522          */
1523         total_con = con_num + nr_poll_queues + 1;
1524         sess->s.con = kcalloc(total_con, sizeof(*sess->s.con), GFP_KERNEL);
1525         if (!sess->s.con)
1526                 goto err_free_sess;
1527
1528         sess->s.con_num = total_con;
1529         sess->s.irq_con_num = con_num + 1;
1530
1531         sess->stats = kzalloc(sizeof(*sess->stats), GFP_KERNEL);
1532         if (!sess->stats)
1533                 goto err_free_con;
1534
1535         mutex_init(&sess->init_mutex);
1536         uuid_gen(&sess->s.uuid);
1537         memcpy(&sess->s.dst_addr, path->dst,
1538                rdma_addr_size((struct sockaddr *)path->dst));
1539
1540         /*
1541          * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1542          * checks the sa_family to be non-zero. If user passed src_addr=NULL
1543          * the sess->src_addr will contain only zeros, which is then fine.
1544          */
1545         if (path->src)
1546                 memcpy(&sess->s.src_addr, path->src,
1547                        rdma_addr_size((struct sockaddr *)path->src));
1548         strscpy(sess->s.sessname, clt->sessname, sizeof(sess->s.sessname));
1549         sess->clt = clt;
1550         sess->max_pages_per_mr = RTRS_MAX_SEGMENTS;
1551         init_waitqueue_head(&sess->state_wq);
1552         sess->state = RTRS_CLT_CONNECTING;
1553         atomic_set(&sess->connected_cnt, 0);
1554         INIT_WORK(&sess->close_work, rtrs_clt_close_work);
1555         INIT_DELAYED_WORK(&sess->reconnect_dwork, rtrs_clt_reconnect_work);
1556         rtrs_clt_init_hb(sess);
1557
1558         sess->mp_skip_entry = alloc_percpu(typeof(*sess->mp_skip_entry));
1559         if (!sess->mp_skip_entry)
1560                 goto err_free_stats;
1561
1562         for_each_possible_cpu(cpu)
1563                 INIT_LIST_HEAD(per_cpu_ptr(sess->mp_skip_entry, cpu));
1564
1565         err = rtrs_clt_init_stats(sess->stats);
1566         if (err)
1567                 goto err_free_percpu;
1568
1569         return sess;
1570
1571 err_free_percpu:
1572         free_percpu(sess->mp_skip_entry);
1573 err_free_stats:
1574         kfree(sess->stats);
1575 err_free_con:
1576         kfree(sess->s.con);
1577 err_free_sess:
1578         kfree(sess);
1579 err:
1580         return ERR_PTR(err);
1581 }
1582
1583 void free_sess(struct rtrs_clt_sess *sess)
1584 {
1585         free_percpu(sess->mp_skip_entry);
1586         mutex_destroy(&sess->init_mutex);
1587         kfree(sess->s.con);
1588         kfree(sess->rbufs);
1589         kfree(sess);
1590 }
1591
1592 static int create_con(struct rtrs_clt_sess *sess, unsigned int cid)
1593 {
1594         struct rtrs_clt_con *con;
1595
1596         con = kzalloc(sizeof(*con), GFP_KERNEL);
1597         if (!con)
1598                 return -ENOMEM;
1599
1600         /* Map first two connections to the first CPU */
1601         con->cpu  = (cid ? cid - 1 : 0) % nr_cpu_ids;
1602         con->c.cid = cid;
1603         con->c.sess = &sess->s;
1604         /* Align with srv, init as 1 */
1605         atomic_set(&con->c.wr_cnt, 1);
1606         mutex_init(&con->con_mutex);
1607
1608         sess->s.con[cid] = &con->c;
1609
1610         return 0;
1611 }
1612
1613 static void destroy_con(struct rtrs_clt_con *con)
1614 {
1615         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1616
1617         sess->s.con[con->c.cid] = NULL;
1618         mutex_destroy(&con->con_mutex);
1619         kfree(con);
1620 }
1621
1622 static int create_con_cq_qp(struct rtrs_clt_con *con)
1623 {
1624         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1625         u32 max_send_wr, max_recv_wr, cq_num, max_send_sge, wr_limit;
1626         int err, cq_vector;
1627         struct rtrs_msg_rkey_rsp *rsp;
1628
1629         lockdep_assert_held(&con->con_mutex);
1630         if (con->c.cid == 0) {
1631                 max_send_sge = 1;
1632                 /* We must be the first here */
1633                 if (WARN_ON(sess->s.dev))
1634                         return -EINVAL;
1635
1636                 /*
1637                  * The whole session uses device from user connection.
1638                  * Be careful not to close user connection before ib dev
1639                  * is gracefully put.
1640                  */
1641                 sess->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1642                                                        &dev_pd);
1643                 if (!sess->s.dev) {
1644                         rtrs_wrn(sess->clt,
1645                                   "rtrs_ib_dev_find_get_or_add(): no memory\n");
1646                         return -ENOMEM;
1647                 }
1648                 sess->s.dev_ref = 1;
1649                 query_fast_reg_mode(sess);
1650                 wr_limit = sess->s.dev->ib_dev->attrs.max_qp_wr;
1651                 /*
1652                  * Two (request + registration) completion for send
1653                  * Two for recv if always_invalidate is set on server
1654                  * or one for recv.
1655                  * + 2 for drain and heartbeat
1656                  * in case qp gets into error state.
1657                  */
1658                 max_send_wr =
1659                         min_t(int, wr_limit, SERVICE_CON_QUEUE_DEPTH * 2 + 2);
1660                 max_recv_wr = max_send_wr;
1661         } else {
1662                 /*
1663                  * Here we assume that session members are correctly set.
1664                  * This is always true if user connection (cid == 0) is
1665                  * established first.
1666                  */
1667                 if (WARN_ON(!sess->s.dev))
1668                         return -EINVAL;
1669                 if (WARN_ON(!sess->queue_depth))
1670                         return -EINVAL;
1671
1672                 wr_limit = sess->s.dev->ib_dev->attrs.max_qp_wr;
1673                 /* Shared between connections */
1674                 sess->s.dev_ref++;
1675                 max_send_wr = min_t(int, wr_limit,
1676                               /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1677                               sess->queue_depth * 3 + 1);
1678                 max_recv_wr = min_t(int, wr_limit,
1679                               sess->queue_depth * 3 + 1);
1680                 max_send_sge = 2;
1681         }
1682         atomic_set(&con->c.sq_wr_avail, max_send_wr);
1683         cq_num = max_send_wr + max_recv_wr;
1684         /* alloc iu to recv new rkey reply when server reports flags set */
1685         if (sess->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1686                 con->rsp_ius = rtrs_iu_alloc(cq_num, sizeof(*rsp),
1687                                               GFP_KERNEL, sess->s.dev->ib_dev,
1688                                               DMA_FROM_DEVICE,
1689                                               rtrs_clt_rdma_done);
1690                 if (!con->rsp_ius)
1691                         return -ENOMEM;
1692                 con->queue_num = cq_num;
1693         }
1694         cq_num = max_send_wr + max_recv_wr;
1695         cq_vector = con->cpu % sess->s.dev->ib_dev->num_comp_vectors;
1696         if (con->c.cid >= sess->s.irq_con_num)
1697                 err = rtrs_cq_qp_create(&sess->s, &con->c, max_send_sge,
1698                                         cq_vector, cq_num, max_send_wr,
1699                                         max_recv_wr, IB_POLL_DIRECT);
1700         else
1701                 err = rtrs_cq_qp_create(&sess->s, &con->c, max_send_sge,
1702                                         cq_vector, cq_num, max_send_wr,
1703                                         max_recv_wr, IB_POLL_SOFTIRQ);
1704         /*
1705          * In case of error we do not bother to clean previous allocations,
1706          * since destroy_con_cq_qp() must be called.
1707          */
1708         return err;
1709 }
1710
1711 static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1712 {
1713         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1714
1715         /*
1716          * Be careful here: destroy_con_cq_qp() can be called even
1717          * create_con_cq_qp() failed, see comments there.
1718          */
1719         lockdep_assert_held(&con->con_mutex);
1720         rtrs_cq_qp_destroy(&con->c);
1721         if (con->rsp_ius) {
1722                 rtrs_iu_free(con->rsp_ius, sess->s.dev->ib_dev, con->queue_num);
1723                 con->rsp_ius = NULL;
1724                 con->queue_num = 0;
1725         }
1726         if (sess->s.dev_ref && !--sess->s.dev_ref) {
1727                 rtrs_ib_dev_put(sess->s.dev);
1728                 sess->s.dev = NULL;
1729         }
1730 }
1731
1732 static void stop_cm(struct rtrs_clt_con *con)
1733 {
1734         rdma_disconnect(con->c.cm_id);
1735         if (con->c.qp)
1736                 ib_drain_qp(con->c.qp);
1737 }
1738
1739 static void destroy_cm(struct rtrs_clt_con *con)
1740 {
1741         rdma_destroy_id(con->c.cm_id);
1742         con->c.cm_id = NULL;
1743 }
1744
1745 static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1746 {
1747         struct rtrs_sess *s = con->c.sess;
1748         int err;
1749
1750         mutex_lock(&con->con_mutex);
1751         err = create_con_cq_qp(con);
1752         mutex_unlock(&con->con_mutex);
1753         if (err) {
1754                 rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1755                 return err;
1756         }
1757         err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1758         if (err)
1759                 rtrs_err(s, "Resolving route failed, err: %d\n", err);
1760
1761         return err;
1762 }
1763
1764 static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1765 {
1766         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1767         struct rtrs_clt *clt = sess->clt;
1768         struct rtrs_msg_conn_req msg;
1769         struct rdma_conn_param param;
1770
1771         int err;
1772
1773         param = (struct rdma_conn_param) {
1774                 .retry_count = 7,
1775                 .rnr_retry_count = 7,
1776                 .private_data = &msg,
1777                 .private_data_len = sizeof(msg),
1778         };
1779
1780         msg = (struct rtrs_msg_conn_req) {
1781                 .magic = cpu_to_le16(RTRS_MAGIC),
1782                 .version = cpu_to_le16(RTRS_PROTO_VER),
1783                 .cid = cpu_to_le16(con->c.cid),
1784                 .cid_num = cpu_to_le16(sess->s.con_num),
1785                 .recon_cnt = cpu_to_le16(sess->s.recon_cnt),
1786         };
1787         msg.first_conn = sess->for_new_clt ? FIRST_CONN : 0;
1788         uuid_copy(&msg.sess_uuid, &sess->s.uuid);
1789         uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1790
1791         err = rdma_connect_locked(con->c.cm_id, &param);
1792         if (err)
1793                 rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1794
1795         return err;
1796 }
1797
1798 static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1799                                        struct rdma_cm_event *ev)
1800 {
1801         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1802         struct rtrs_clt *clt = sess->clt;
1803         const struct rtrs_msg_conn_rsp *msg;
1804         u16 version, queue_depth;
1805         int errno;
1806         u8 len;
1807
1808         msg = ev->param.conn.private_data;
1809         len = ev->param.conn.private_data_len;
1810         if (len < sizeof(*msg)) {
1811                 rtrs_err(clt, "Invalid RTRS connection response\n");
1812                 return -ECONNRESET;
1813         }
1814         if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1815                 rtrs_err(clt, "Invalid RTRS magic\n");
1816                 return -ECONNRESET;
1817         }
1818         version = le16_to_cpu(msg->version);
1819         if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1820                 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1821                           version >> 8, RTRS_PROTO_VER_MAJOR);
1822                 return -ECONNRESET;
1823         }
1824         errno = le16_to_cpu(msg->errno);
1825         if (errno) {
1826                 rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1827                           errno);
1828                 return -ECONNRESET;
1829         }
1830         if (con->c.cid == 0) {
1831                 queue_depth = le16_to_cpu(msg->queue_depth);
1832
1833                 if (sess->queue_depth > 0 && queue_depth != sess->queue_depth) {
1834                         rtrs_err(clt, "Error: queue depth changed\n");
1835
1836                         /*
1837                          * Stop any more reconnection attempts
1838                          */
1839                         sess->reconnect_attempts = -1;
1840                         rtrs_err(clt,
1841                                 "Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n");
1842                         return -ECONNRESET;
1843                 }
1844
1845                 if (!sess->rbufs) {
1846                         sess->rbufs = kcalloc(queue_depth, sizeof(*sess->rbufs),
1847                                               GFP_KERNEL);
1848                         if (!sess->rbufs)
1849                                 return -ENOMEM;
1850                 }
1851                 sess->queue_depth = queue_depth;
1852                 sess->s.signal_interval = min_not_zero(queue_depth,
1853                                                 (unsigned short) SERVICE_CON_QUEUE_DEPTH);
1854                 sess->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1855                 sess->max_io_size = le32_to_cpu(msg->max_io_size);
1856                 sess->flags = le32_to_cpu(msg->flags);
1857                 sess->chunk_size = sess->max_io_size + sess->max_hdr_size;
1858
1859                 /*
1860                  * Global IO size is always a minimum.
1861                  * If while a reconnection server sends us a value a bit
1862                  * higher - client does not care and uses cached minimum.
1863                  *
1864                  * Since we can have several sessions (paths) restablishing
1865                  * connections in parallel, use lock.
1866                  */
1867                 mutex_lock(&clt->paths_mutex);
1868                 clt->queue_depth = sess->queue_depth;
1869                 clt->max_io_size = min_not_zero(sess->max_io_size,
1870                                                 clt->max_io_size);
1871                 mutex_unlock(&clt->paths_mutex);
1872
1873                 /*
1874                  * Cache the hca_port and hca_name for sysfs
1875                  */
1876                 sess->hca_port = con->c.cm_id->port_num;
1877                 scnprintf(sess->hca_name, sizeof(sess->hca_name),
1878                           sess->s.dev->ib_dev->name);
1879                 sess->s.src_addr = con->c.cm_id->route.addr.src_addr;
1880                 /* set for_new_clt, to allow future reconnect on any path */
1881                 sess->for_new_clt = 1;
1882         }
1883
1884         return 0;
1885 }
1886
1887 static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1888 {
1889         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1890
1891         atomic_inc(&sess->connected_cnt);
1892         con->cm_err = 1;
1893 }
1894
1895 static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1896                                     struct rdma_cm_event *ev)
1897 {
1898         struct rtrs_sess *s = con->c.sess;
1899         const struct rtrs_msg_conn_rsp *msg;
1900         const char *rej_msg;
1901         int status, errno;
1902         u8 data_len;
1903
1904         status = ev->status;
1905         rej_msg = rdma_reject_msg(con->c.cm_id, status);
1906         msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1907
1908         if (msg && data_len >= sizeof(*msg)) {
1909                 errno = (int16_t)le16_to_cpu(msg->errno);
1910                 if (errno == -EBUSY)
1911                         rtrs_err(s,
1912                                   "Previous session is still exists on the server, please reconnect later\n");
1913                 else
1914                         rtrs_err(s,
1915                                   "Connect rejected: status %d (%s), rtrs errno %d\n",
1916                                   status, rej_msg, errno);
1917         } else {
1918                 rtrs_err(s,
1919                           "Connect rejected but with malformed message: status %d (%s)\n",
1920                           status, rej_msg);
1921         }
1922
1923         return -ECONNRESET;
1924 }
1925
1926 void rtrs_clt_close_conns(struct rtrs_clt_sess *sess, bool wait)
1927 {
1928         if (rtrs_clt_change_state_get_old(sess, RTRS_CLT_CLOSING, NULL))
1929                 queue_work(rtrs_wq, &sess->close_work);
1930         if (wait)
1931                 flush_work(&sess->close_work);
1932 }
1933
1934 static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1935 {
1936         if (con->cm_err == 1) {
1937                 struct rtrs_clt_sess *sess;
1938
1939                 sess = to_clt_sess(con->c.sess);
1940                 if (atomic_dec_and_test(&sess->connected_cnt))
1941
1942                         wake_up(&sess->state_wq);
1943         }
1944         con->cm_err = cm_err;
1945 }
1946
1947 static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1948                                      struct rdma_cm_event *ev)
1949 {
1950         struct rtrs_clt_con *con = cm_id->context;
1951         struct rtrs_sess *s = con->c.sess;
1952         struct rtrs_clt_sess *sess = to_clt_sess(s);
1953         int cm_err = 0;
1954
1955         switch (ev->event) {
1956         case RDMA_CM_EVENT_ADDR_RESOLVED:
1957                 cm_err = rtrs_rdma_addr_resolved(con);
1958                 break;
1959         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1960                 cm_err = rtrs_rdma_route_resolved(con);
1961                 break;
1962         case RDMA_CM_EVENT_ESTABLISHED:
1963                 cm_err = rtrs_rdma_conn_established(con, ev);
1964                 if (!cm_err) {
1965                         /*
1966                          * Report success and wake up. Here we abuse state_wq,
1967                          * i.e. wake up without state change, but we set cm_err.
1968                          */
1969                         flag_success_on_conn(con);
1970                         wake_up(&sess->state_wq);
1971                         return 0;
1972                 }
1973                 break;
1974         case RDMA_CM_EVENT_REJECTED:
1975                 cm_err = rtrs_rdma_conn_rejected(con, ev);
1976                 break;
1977         case RDMA_CM_EVENT_DISCONNECTED:
1978                 /* No message for disconnecting */
1979                 cm_err = -ECONNRESET;
1980                 break;
1981         case RDMA_CM_EVENT_CONNECT_ERROR:
1982         case RDMA_CM_EVENT_UNREACHABLE:
1983         case RDMA_CM_EVENT_ADDR_CHANGE:
1984         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1985                 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
1986                          rdma_event_msg(ev->event), ev->status);
1987                 cm_err = -ECONNRESET;
1988                 break;
1989         case RDMA_CM_EVENT_ADDR_ERROR:
1990         case RDMA_CM_EVENT_ROUTE_ERROR:
1991                 rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
1992                          rdma_event_msg(ev->event), ev->status);
1993                 cm_err = -EHOSTUNREACH;
1994                 break;
1995         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1996                 /*
1997                  * Device removal is a special case.  Queue close and return 0.
1998                  */
1999                 rtrs_clt_close_conns(sess, false);
2000                 return 0;
2001         default:
2002                 rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %d)\n",
2003                          rdma_event_msg(ev->event), ev->status);
2004                 cm_err = -ECONNRESET;
2005                 break;
2006         }
2007
2008         if (cm_err) {
2009                 /*
2010                  * cm error makes sense only on connection establishing,
2011                  * in other cases we rely on normal procedure of reconnecting.
2012                  */
2013                 flag_error_on_conn(con, cm_err);
2014                 rtrs_rdma_error_recovery(con);
2015         }
2016
2017         return 0;
2018 }
2019
2020 static int create_cm(struct rtrs_clt_con *con)
2021 {
2022         struct rtrs_sess *s = con->c.sess;
2023         struct rtrs_clt_sess *sess = to_clt_sess(s);
2024         struct rdma_cm_id *cm_id;
2025         int err;
2026
2027         cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
2028                                sess->s.dst_addr.ss_family == AF_IB ?
2029                                RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
2030         if (IS_ERR(cm_id)) {
2031                 err = PTR_ERR(cm_id);
2032                 rtrs_err(s, "Failed to create CM ID, err: %d\n", err);
2033
2034                 return err;
2035         }
2036         con->c.cm_id = cm_id;
2037         con->cm_err = 0;
2038         /* allow the port to be reused */
2039         err = rdma_set_reuseaddr(cm_id, 1);
2040         if (err != 0) {
2041                 rtrs_err(s, "Set address reuse failed, err: %d\n", err);
2042                 goto destroy_cm;
2043         }
2044         err = rdma_resolve_addr(cm_id, (struct sockaddr *)&sess->s.src_addr,
2045                                 (struct sockaddr *)&sess->s.dst_addr,
2046                                 RTRS_CONNECT_TIMEOUT_MS);
2047         if (err) {
2048                 rtrs_err(s, "Failed to resolve address, err: %d\n", err);
2049                 goto destroy_cm;
2050         }
2051         /*
2052          * Combine connection status and session events. This is needed
2053          * for waiting two possible cases: cm_err has something meaningful
2054          * or session state was really changed to error by device removal.
2055          */
2056         err = wait_event_interruptible_timeout(
2057                         sess->state_wq,
2058                         con->cm_err || sess->state != RTRS_CLT_CONNECTING,
2059                         msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2060         if (err == 0 || err == -ERESTARTSYS) {
2061                 if (err == 0)
2062                         err = -ETIMEDOUT;
2063                 /* Timedout or interrupted */
2064                 goto errr;
2065         }
2066         if (con->cm_err < 0) {
2067                 err = con->cm_err;
2068                 goto errr;
2069         }
2070         if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTING) {
2071                 /* Device removal */
2072                 err = -ECONNABORTED;
2073                 goto errr;
2074         }
2075
2076         return 0;
2077
2078 errr:
2079         stop_cm(con);
2080         mutex_lock(&con->con_mutex);
2081         destroy_con_cq_qp(con);
2082         mutex_unlock(&con->con_mutex);
2083 destroy_cm:
2084         destroy_cm(con);
2085
2086         return err;
2087 }
2088
2089 static void rtrs_clt_sess_up(struct rtrs_clt_sess *sess)
2090 {
2091         struct rtrs_clt *clt = sess->clt;
2092         int up;
2093
2094         /*
2095          * We can fire RECONNECTED event only when all paths were
2096          * connected on rtrs_clt_open(), then each was disconnected
2097          * and the first one connected again.  That's why this nasty
2098          * game with counter value.
2099          */
2100
2101         mutex_lock(&clt->paths_ev_mutex);
2102         up = ++clt->paths_up;
2103         /*
2104          * Here it is safe to access paths num directly since up counter
2105          * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
2106          * in progress, thus paths removals are impossible.
2107          */
2108         if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
2109                 clt->paths_up = clt->paths_num;
2110         else if (up == 1)
2111                 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
2112         mutex_unlock(&clt->paths_ev_mutex);
2113
2114         /* Mark session as established */
2115         sess->established = true;
2116         sess->reconnect_attempts = 0;
2117         sess->stats->reconnects.successful_cnt++;
2118 }
2119
2120 static void rtrs_clt_sess_down(struct rtrs_clt_sess *sess)
2121 {
2122         struct rtrs_clt *clt = sess->clt;
2123
2124         if (!sess->established)
2125                 return;
2126
2127         sess->established = false;
2128         mutex_lock(&clt->paths_ev_mutex);
2129         WARN_ON(!clt->paths_up);
2130         if (--clt->paths_up == 0)
2131                 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2132         mutex_unlock(&clt->paths_ev_mutex);
2133 }
2134
2135 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_sess *sess)
2136 {
2137         struct rtrs_clt_con *con;
2138         unsigned int cid;
2139
2140         WARN_ON(READ_ONCE(sess->state) == RTRS_CLT_CONNECTED);
2141
2142         /*
2143          * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2144          * exactly in between.  Start destroying after it finishes.
2145          */
2146         mutex_lock(&sess->init_mutex);
2147         mutex_unlock(&sess->init_mutex);
2148
2149         /*
2150          * All IO paths must observe !CONNECTED state before we
2151          * free everything.
2152          */
2153         synchronize_rcu();
2154
2155         rtrs_stop_hb(&sess->s);
2156
2157         /*
2158          * The order it utterly crucial: firstly disconnect and complete all
2159          * rdma requests with error (thus set in_use=false for requests),
2160          * then fail outstanding requests checking in_use for each, and
2161          * eventually notify upper layer about session disconnection.
2162          */
2163
2164         for (cid = 0; cid < sess->s.con_num; cid++) {
2165                 if (!sess->s.con[cid])
2166                         break;
2167                 con = to_clt_con(sess->s.con[cid]);
2168                 stop_cm(con);
2169         }
2170         fail_all_outstanding_reqs(sess);
2171         free_sess_reqs(sess);
2172         rtrs_clt_sess_down(sess);
2173
2174         /*
2175          * Wait for graceful shutdown, namely when peer side invokes
2176          * rdma_disconnect(). 'connected_cnt' is decremented only on
2177          * CM events, thus if other side had crashed and hb has detected
2178          * something is wrong, here we will stuck for exactly timeout ms,
2179          * since CM does not fire anything.  That is fine, we are not in
2180          * hurry.
2181          */
2182         wait_event_timeout(sess->state_wq, !atomic_read(&sess->connected_cnt),
2183                            msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2184
2185         for (cid = 0; cid < sess->s.con_num; cid++) {
2186                 if (!sess->s.con[cid])
2187                         break;
2188                 con = to_clt_con(sess->s.con[cid]);
2189                 mutex_lock(&con->con_mutex);
2190                 destroy_con_cq_qp(con);
2191                 mutex_unlock(&con->con_mutex);
2192                 destroy_cm(con);
2193                 destroy_con(con);
2194         }
2195 }
2196
2197 static inline bool xchg_sessions(struct rtrs_clt_sess __rcu **rcu_ppcpu_path,
2198                                  struct rtrs_clt_sess *sess,
2199                                  struct rtrs_clt_sess *next)
2200 {
2201         struct rtrs_clt_sess **ppcpu_path;
2202
2203         /* Call cmpxchg() without sparse warnings */
2204         ppcpu_path = (typeof(ppcpu_path))rcu_ppcpu_path;
2205         return sess == cmpxchg(ppcpu_path, sess, next);
2206 }
2207
2208 static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_sess *sess)
2209 {
2210         struct rtrs_clt *clt = sess->clt;
2211         struct rtrs_clt_sess *next;
2212         bool wait_for_grace = false;
2213         int cpu;
2214
2215         mutex_lock(&clt->paths_mutex);
2216         list_del_rcu(&sess->s.entry);
2217
2218         /* Make sure everybody observes path removal. */
2219         synchronize_rcu();
2220
2221         /*
2222          * At this point nobody sees @sess in the list, but still we have
2223          * dangling pointer @pcpu_path which _can_ point to @sess.  Since
2224          * nobody can observe @sess in the list, we guarantee that IO path
2225          * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2226          * to @sess, but can never again become @sess.
2227          */
2228
2229         /*
2230          * Decrement paths number only after grace period, because
2231          * caller of do_each_path() must firstly observe list without
2232          * path and only then decremented paths number.
2233          *
2234          * Otherwise there can be the following situation:
2235          *    o Two paths exist and IO is coming.
2236          *    o One path is removed:
2237          *      CPU#0                          CPU#1
2238          *      do_each_path():                rtrs_clt_remove_path_from_arr():
2239          *          path = get_next_path()
2240          *          ^^^                            list_del_rcu(path)
2241          *          [!CONNECTED path]              clt->paths_num--
2242          *                                              ^^^^^^^^^
2243          *          load clt->paths_num                 from 2 to 1
2244          *                    ^^^^^^^^^
2245          *                    sees 1
2246          *
2247          *      path is observed as !CONNECTED, but do_each_path() loop
2248          *      ends, because expression i < clt->paths_num is false.
2249          */
2250         clt->paths_num--;
2251
2252         /*
2253          * Get @next connection from current @sess which is going to be
2254          * removed.  If @sess is the last element, then @next is NULL.
2255          */
2256         rcu_read_lock();
2257         next = list_next_or_null_rr_rcu(&clt->paths_list, &sess->s.entry,
2258                                         typeof(*next), s.entry);
2259         rcu_read_unlock();
2260
2261         /*
2262          * @pcpu paths can still point to the path which is going to be
2263          * removed, so change the pointer manually.
2264          */
2265         for_each_possible_cpu(cpu) {
2266                 struct rtrs_clt_sess __rcu **ppcpu_path;
2267
2268                 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2269                 if (rcu_dereference_protected(*ppcpu_path,
2270                         lockdep_is_held(&clt->paths_mutex)) != sess)
2271                         /*
2272                          * synchronize_rcu() was called just after deleting
2273                          * entry from the list, thus IO code path cannot
2274                          * change pointer back to the pointer which is going
2275                          * to be removed, we are safe here.
2276                          */
2277                         continue;
2278
2279                 /*
2280                  * We race with IO code path, which also changes pointer,
2281                  * thus we have to be careful not to overwrite it.
2282                  */
2283                 if (xchg_sessions(ppcpu_path, sess, next))
2284                         /*
2285                          * @ppcpu_path was successfully replaced with @next,
2286                          * that means that someone could also pick up the
2287                          * @sess and dereferencing it right now, so wait for
2288                          * a grace period is required.
2289                          */
2290                         wait_for_grace = true;
2291         }
2292         if (wait_for_grace)
2293                 synchronize_rcu();
2294
2295         mutex_unlock(&clt->paths_mutex);
2296 }
2297
2298 static void rtrs_clt_add_path_to_arr(struct rtrs_clt_sess *sess)
2299 {
2300         struct rtrs_clt *clt = sess->clt;
2301
2302         mutex_lock(&clt->paths_mutex);
2303         clt->paths_num++;
2304
2305         list_add_tail_rcu(&sess->s.entry, &clt->paths_list);
2306         mutex_unlock(&clt->paths_mutex);
2307 }
2308
2309 static void rtrs_clt_close_work(struct work_struct *work)
2310 {
2311         struct rtrs_clt_sess *sess;
2312
2313         sess = container_of(work, struct rtrs_clt_sess, close_work);
2314
2315         cancel_delayed_work_sync(&sess->reconnect_dwork);
2316         rtrs_clt_stop_and_destroy_conns(sess);
2317         rtrs_clt_change_state_get_old(sess, RTRS_CLT_CLOSED, NULL);
2318 }
2319
2320 static int init_conns(struct rtrs_clt_sess *sess)
2321 {
2322         unsigned int cid;
2323         int err;
2324
2325         /*
2326          * On every new session connections increase reconnect counter
2327          * to avoid clashes with previous sessions not yet closed
2328          * sessions on a server side.
2329          */
2330         sess->s.recon_cnt++;
2331
2332         /* Establish all RDMA connections  */
2333         for (cid = 0; cid < sess->s.con_num; cid++) {
2334                 err = create_con(sess, cid);
2335                 if (err)
2336                         goto destroy;
2337
2338                 err = create_cm(to_clt_con(sess->s.con[cid]));
2339                 if (err) {
2340                         destroy_con(to_clt_con(sess->s.con[cid]));
2341                         goto destroy;
2342                 }
2343         }
2344         err = alloc_sess_reqs(sess);
2345         if (err)
2346                 goto destroy;
2347
2348         rtrs_start_hb(&sess->s);
2349
2350         return 0;
2351
2352 destroy:
2353         while (cid--) {
2354                 struct rtrs_clt_con *con = to_clt_con(sess->s.con[cid]);
2355
2356                 stop_cm(con);
2357
2358                 mutex_lock(&con->con_mutex);
2359                 destroy_con_cq_qp(con);
2360                 mutex_unlock(&con->con_mutex);
2361                 destroy_cm(con);
2362                 destroy_con(con);
2363         }
2364         /*
2365          * If we've never taken async path and got an error, say,
2366          * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2367          * manually to keep reconnecting.
2368          */
2369         rtrs_clt_change_state_get_old(sess, RTRS_CLT_CONNECTING_ERR, NULL);
2370
2371         return err;
2372 }
2373
2374 static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2375 {
2376         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2377         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
2378         struct rtrs_iu *iu;
2379
2380         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2381         rtrs_iu_free(iu, sess->s.dev->ib_dev, 1);
2382
2383         if (wc->status != IB_WC_SUCCESS) {
2384                 rtrs_err(sess->clt, "Sess info request send failed: %s\n",
2385                           ib_wc_status_msg(wc->status));
2386                 rtrs_clt_change_state_get_old(sess, RTRS_CLT_CONNECTING_ERR, NULL);
2387                 return;
2388         }
2389
2390         rtrs_clt_update_wc_stats(con);
2391 }
2392
2393 static int process_info_rsp(struct rtrs_clt_sess *sess,
2394                             const struct rtrs_msg_info_rsp *msg)
2395 {
2396         unsigned int sg_cnt, total_len;
2397         int i, sgi;
2398
2399         sg_cnt = le16_to_cpu(msg->sg_cnt);
2400         if (!sg_cnt || (sess->queue_depth % sg_cnt)) {
2401                 rtrs_err(sess->clt, "Incorrect sg_cnt %d, is not multiple\n",
2402                           sg_cnt);
2403                 return -EINVAL;
2404         }
2405
2406         /*
2407          * Check if IB immediate data size is enough to hold the mem_id and
2408          * the offset inside the memory chunk.
2409          */
2410         if ((ilog2(sg_cnt - 1) + 1) + (ilog2(sess->chunk_size - 1) + 1) >
2411             MAX_IMM_PAYL_BITS) {
2412                 rtrs_err(sess->clt,
2413                           "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2414                           MAX_IMM_PAYL_BITS, sg_cnt, sess->chunk_size);
2415                 return -EINVAL;
2416         }
2417         total_len = 0;
2418         for (sgi = 0, i = 0; sgi < sg_cnt && i < sess->queue_depth; sgi++) {
2419                 const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2420                 u32 len, rkey;
2421                 u64 addr;
2422
2423                 addr = le64_to_cpu(desc->addr);
2424                 rkey = le32_to_cpu(desc->key);
2425                 len  = le32_to_cpu(desc->len);
2426
2427                 total_len += len;
2428
2429                 if (!len || (len % sess->chunk_size)) {
2430                         rtrs_err(sess->clt, "Incorrect [%d].len %d\n", sgi,
2431                                   len);
2432                         return -EINVAL;
2433                 }
2434                 for ( ; len && i < sess->queue_depth; i++) {
2435                         sess->rbufs[i].addr = addr;
2436                         sess->rbufs[i].rkey = rkey;
2437
2438                         len  -= sess->chunk_size;
2439                         addr += sess->chunk_size;
2440                 }
2441         }
2442         /* Sanity check */
2443         if (sgi != sg_cnt || i != sess->queue_depth) {
2444                 rtrs_err(sess->clt, "Incorrect sg vector, not fully mapped\n");
2445                 return -EINVAL;
2446         }
2447         if (total_len != sess->chunk_size * sess->queue_depth) {
2448                 rtrs_err(sess->clt, "Incorrect total_len %d\n", total_len);
2449                 return -EINVAL;
2450         }
2451
2452         return 0;
2453 }
2454
2455 static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2456 {
2457         struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2458         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
2459         struct rtrs_msg_info_rsp *msg;
2460         enum rtrs_clt_state state;
2461         struct rtrs_iu *iu;
2462         size_t rx_sz;
2463         int err;
2464
2465         state = RTRS_CLT_CONNECTING_ERR;
2466
2467         WARN_ON(con->c.cid);
2468         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2469         if (wc->status != IB_WC_SUCCESS) {
2470                 rtrs_err(sess->clt, "Sess info response recv failed: %s\n",
2471                           ib_wc_status_msg(wc->status));
2472                 goto out;
2473         }
2474         WARN_ON(wc->opcode != IB_WC_RECV);
2475
2476         if (wc->byte_len < sizeof(*msg)) {
2477                 rtrs_err(sess->clt, "Sess info response is malformed: size %d\n",
2478                           wc->byte_len);
2479                 goto out;
2480         }
2481         ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, iu->dma_addr,
2482                                    iu->size, DMA_FROM_DEVICE);
2483         msg = iu->buf;
2484         if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP) {
2485                 rtrs_err(sess->clt, "Sess info response is malformed: type %d\n",
2486                           le16_to_cpu(msg->type));
2487                 goto out;
2488         }
2489         rx_sz  = sizeof(*msg);
2490         rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2491         if (wc->byte_len < rx_sz) {
2492                 rtrs_err(sess->clt, "Sess info response is malformed: size %d\n",
2493                           wc->byte_len);
2494                 goto out;
2495         }
2496         err = process_info_rsp(sess, msg);
2497         if (err)
2498                 goto out;
2499
2500         err = post_recv_sess(sess);
2501         if (err)
2502                 goto out;
2503
2504         state = RTRS_CLT_CONNECTED;
2505
2506 out:
2507         rtrs_clt_update_wc_stats(con);
2508         rtrs_iu_free(iu, sess->s.dev->ib_dev, 1);
2509         rtrs_clt_change_state_get_old(sess, state, NULL);
2510 }
2511
2512 static int rtrs_send_sess_info(struct rtrs_clt_sess *sess)
2513 {
2514         struct rtrs_clt_con *usr_con = to_clt_con(sess->s.con[0]);
2515         struct rtrs_msg_info_req *msg;
2516         struct rtrs_iu *tx_iu, *rx_iu;
2517         size_t rx_sz;
2518         int err;
2519
2520         rx_sz  = sizeof(struct rtrs_msg_info_rsp);
2521         rx_sz += sizeof(struct rtrs_sg_desc) * sess->queue_depth;
2522
2523         tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2524                                sess->s.dev->ib_dev, DMA_TO_DEVICE,
2525                                rtrs_clt_info_req_done);
2526         rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, sess->s.dev->ib_dev,
2527                                DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2528         if (!tx_iu || !rx_iu) {
2529                 err = -ENOMEM;
2530                 goto out;
2531         }
2532         /* Prepare for getting info response */
2533         err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2534         if (err) {
2535                 rtrs_err(sess->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2536                 goto out;
2537         }
2538         rx_iu = NULL;
2539
2540         msg = tx_iu->buf;
2541         msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2542         memcpy(msg->sessname, sess->s.sessname, sizeof(msg->sessname));
2543
2544         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, tx_iu->dma_addr,
2545                                       tx_iu->size, DMA_TO_DEVICE);
2546
2547         /* Send info request */
2548         err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2549         if (err) {
2550                 rtrs_err(sess->clt, "rtrs_iu_post_send(), err: %d\n", err);
2551                 goto out;
2552         }
2553         tx_iu = NULL;
2554
2555         /* Wait for state change */
2556         wait_event_interruptible_timeout(sess->state_wq,
2557                                          sess->state != RTRS_CLT_CONNECTING,
2558                                          msecs_to_jiffies(
2559                                                  RTRS_CONNECT_TIMEOUT_MS));
2560         if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTED) {
2561                 if (READ_ONCE(sess->state) == RTRS_CLT_CONNECTING_ERR)
2562                         err = -ECONNRESET;
2563                 else
2564                         err = -ETIMEDOUT;
2565         }
2566
2567 out:
2568         if (tx_iu)
2569                 rtrs_iu_free(tx_iu, sess->s.dev->ib_dev, 1);
2570         if (rx_iu)
2571                 rtrs_iu_free(rx_iu, sess->s.dev->ib_dev, 1);
2572         if (err)
2573                 /* If we've never taken async path because of malloc problems */
2574                 rtrs_clt_change_state_get_old(sess, RTRS_CLT_CONNECTING_ERR, NULL);
2575
2576         return err;
2577 }
2578
2579 /**
2580  * init_sess() - establishes all session connections and does handshake
2581  * @sess: client session.
2582  * In case of error full close or reconnect procedure should be taken,
2583  * because reconnect or close async works can be started.
2584  */
2585 static int init_sess(struct rtrs_clt_sess *sess)
2586 {
2587         int err;
2588         char str[NAME_MAX];
2589         struct rtrs_addr path = {
2590                 .src = &sess->s.src_addr,
2591                 .dst = &sess->s.dst_addr,
2592         };
2593
2594         rtrs_addr_to_str(&path, str, sizeof(str));
2595
2596         mutex_lock(&sess->init_mutex);
2597         err = init_conns(sess);
2598         if (err) {
2599                 rtrs_err(sess->clt,
2600                          "init_conns() failed: err=%d path=%s [%s:%u]\n", err,
2601                          str, sess->hca_name, sess->hca_port);
2602                 goto out;
2603         }
2604         err = rtrs_send_sess_info(sess);
2605         if (err) {
2606                 rtrs_err(
2607                         sess->clt,
2608                         "rtrs_send_sess_info() failed: err=%d path=%s [%s:%u]\n",
2609                         err, str, sess->hca_name, sess->hca_port);
2610                 goto out;
2611         }
2612         rtrs_clt_sess_up(sess);
2613 out:
2614         mutex_unlock(&sess->init_mutex);
2615
2616         return err;
2617 }
2618
2619 static void rtrs_clt_reconnect_work(struct work_struct *work)
2620 {
2621         struct rtrs_clt_sess *sess;
2622         struct rtrs_clt *clt;
2623         unsigned int delay_ms;
2624         int err;
2625
2626         sess = container_of(to_delayed_work(work), struct rtrs_clt_sess,
2627                             reconnect_dwork);
2628         clt = sess->clt;
2629
2630         if (READ_ONCE(sess->state) != RTRS_CLT_RECONNECTING)
2631                 return;
2632
2633         if (sess->reconnect_attempts >= clt->max_reconnect_attempts) {
2634                 /* Close a session completely if max attempts is reached */
2635                 rtrs_clt_close_conns(sess, false);
2636                 return;
2637         }
2638         sess->reconnect_attempts++;
2639
2640         /* Stop everything */
2641         rtrs_clt_stop_and_destroy_conns(sess);
2642         msleep(RTRS_RECONNECT_BACKOFF);
2643         if (rtrs_clt_change_state_get_old(sess, RTRS_CLT_CONNECTING, NULL)) {
2644                 err = init_sess(sess);
2645                 if (err)
2646                         goto reconnect_again;
2647         }
2648
2649         return;
2650
2651 reconnect_again:
2652         if (rtrs_clt_change_state_get_old(sess, RTRS_CLT_RECONNECTING, NULL)) {
2653                 sess->stats->reconnects.fail_cnt++;
2654                 delay_ms = clt->reconnect_delay_sec * 1000;
2655                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork,
2656                                    msecs_to_jiffies(delay_ms +
2657                                                     prandom_u32() %
2658                                                     RTRS_RECONNECT_SEED));
2659         }
2660 }
2661
2662 static void rtrs_clt_dev_release(struct device *dev)
2663 {
2664         struct rtrs_clt *clt = container_of(dev, struct rtrs_clt, dev);
2665
2666         kfree(clt);
2667 }
2668
2669 static struct rtrs_clt *alloc_clt(const char *sessname, size_t paths_num,
2670                                   u16 port, size_t pdu_sz, void *priv,
2671                                   void  (*link_ev)(void *priv,
2672                                                    enum rtrs_clt_link_ev ev),
2673                                   unsigned int reconnect_delay_sec,
2674                                   unsigned int max_reconnect_attempts)
2675 {
2676         struct rtrs_clt *clt;
2677         int err;
2678
2679         if (!paths_num || paths_num > MAX_PATHS_NUM)
2680                 return ERR_PTR(-EINVAL);
2681
2682         if (strlen(sessname) >= sizeof(clt->sessname))
2683                 return ERR_PTR(-EINVAL);
2684
2685         clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2686         if (!clt)
2687                 return ERR_PTR(-ENOMEM);
2688
2689         clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2690         if (!clt->pcpu_path) {
2691                 kfree(clt);
2692                 return ERR_PTR(-ENOMEM);
2693         }
2694
2695         uuid_gen(&clt->paths_uuid);
2696         INIT_LIST_HEAD_RCU(&clt->paths_list);
2697         clt->paths_num = paths_num;
2698         clt->paths_up = MAX_PATHS_NUM;
2699         clt->port = port;
2700         clt->pdu_sz = pdu_sz;
2701         clt->max_segments = RTRS_MAX_SEGMENTS;
2702         clt->reconnect_delay_sec = reconnect_delay_sec;
2703         clt->max_reconnect_attempts = max_reconnect_attempts;
2704         clt->priv = priv;
2705         clt->link_ev = link_ev;
2706         clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2707         strscpy(clt->sessname, sessname, sizeof(clt->sessname));
2708         init_waitqueue_head(&clt->permits_wait);
2709         mutex_init(&clt->paths_ev_mutex);
2710         mutex_init(&clt->paths_mutex);
2711
2712         clt->dev.class = rtrs_clt_dev_class;
2713         clt->dev.release = rtrs_clt_dev_release;
2714         err = dev_set_name(&clt->dev, "%s", sessname);
2715         if (err)
2716                 goto err;
2717         /*
2718          * Suppress user space notification until
2719          * sysfs files are created
2720          */
2721         dev_set_uevent_suppress(&clt->dev, true);
2722         err = device_register(&clt->dev);
2723         if (err) {
2724                 put_device(&clt->dev);
2725                 goto err;
2726         }
2727
2728         clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2729         if (!clt->kobj_paths) {
2730                 err = -ENOMEM;
2731                 goto err_dev;
2732         }
2733         err = rtrs_clt_create_sysfs_root_files(clt);
2734         if (err) {
2735                 kobject_del(clt->kobj_paths);
2736                 kobject_put(clt->kobj_paths);
2737                 goto err_dev;
2738         }
2739         dev_set_uevent_suppress(&clt->dev, false);
2740         kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2741
2742         return clt;
2743 err_dev:
2744         device_unregister(&clt->dev);
2745 err:
2746         free_percpu(clt->pcpu_path);
2747         kfree(clt);
2748         return ERR_PTR(err);
2749 }
2750
2751 static void free_clt(struct rtrs_clt *clt)
2752 {
2753         free_permits(clt);
2754         free_percpu(clt->pcpu_path);
2755         mutex_destroy(&clt->paths_ev_mutex);
2756         mutex_destroy(&clt->paths_mutex);
2757         /* release callback will free clt in last put */
2758         device_unregister(&clt->dev);
2759 }
2760
2761 /**
2762  * rtrs_clt_open() - Open a session to an RTRS server
2763  * @ops: holds the link event callback and the private pointer.
2764  * @sessname: name of the session
2765  * @paths: Paths to be established defined by their src and dst addresses
2766  * @paths_num: Number of elements in the @paths array
2767  * @port: port to be used by the RTRS session
2768  * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2769  * @reconnect_delay_sec: time between reconnect tries
2770  * @max_reconnect_attempts: Number of times to reconnect on error before giving
2771  *                          up, 0 for * disabled, -1 for forever
2772  * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
2773  *
2774  * Starts session establishment with the rtrs_server. The function can block
2775  * up to ~2000ms before it returns.
2776  *
2777  * Return a valid pointer on success otherwise PTR_ERR.
2778  */
2779 struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
2780                                  const char *sessname,
2781                                  const struct rtrs_addr *paths,
2782                                  size_t paths_num, u16 port,
2783                                  size_t pdu_sz, u8 reconnect_delay_sec,
2784                                  s16 max_reconnect_attempts, u32 nr_poll_queues)
2785 {
2786         struct rtrs_clt_sess *sess, *tmp;
2787         struct rtrs_clt *clt;
2788         int err, i;
2789
2790         clt = alloc_clt(sessname, paths_num, port, pdu_sz, ops->priv,
2791                         ops->link_ev,
2792                         reconnect_delay_sec,
2793                         max_reconnect_attempts);
2794         if (IS_ERR(clt)) {
2795                 err = PTR_ERR(clt);
2796                 goto out;
2797         }
2798         for (i = 0; i < paths_num; i++) {
2799                 struct rtrs_clt_sess *sess;
2800
2801                 sess = alloc_sess(clt, &paths[i], nr_cpu_ids,
2802                                   nr_poll_queues);
2803                 if (IS_ERR(sess)) {
2804                         err = PTR_ERR(sess);
2805                         goto close_all_sess;
2806                 }
2807                 if (!i)
2808                         sess->for_new_clt = 1;
2809                 list_add_tail_rcu(&sess->s.entry, &clt->paths_list);
2810
2811                 err = init_sess(sess);
2812                 if (err) {
2813                         list_del_rcu(&sess->s.entry);
2814                         rtrs_clt_close_conns(sess, true);
2815                         free_percpu(sess->stats->pcpu_stats);
2816                         kfree(sess->stats);
2817                         free_sess(sess);
2818                         goto close_all_sess;
2819                 }
2820
2821                 err = rtrs_clt_create_sess_files(sess);
2822                 if (err) {
2823                         list_del_rcu(&sess->s.entry);
2824                         rtrs_clt_close_conns(sess, true);
2825                         free_percpu(sess->stats->pcpu_stats);
2826                         kfree(sess->stats);
2827                         free_sess(sess);
2828                         goto close_all_sess;
2829                 }
2830         }
2831         err = alloc_permits(clt);
2832         if (err)
2833                 goto close_all_sess;
2834
2835         return clt;
2836
2837 close_all_sess:
2838         list_for_each_entry_safe(sess, tmp, &clt->paths_list, s.entry) {
2839                 rtrs_clt_destroy_sess_files(sess, NULL);
2840                 rtrs_clt_close_conns(sess, true);
2841                 kobject_put(&sess->kobj);
2842         }
2843         rtrs_clt_destroy_sysfs_root(clt);
2844         free_clt(clt);
2845
2846 out:
2847         return ERR_PTR(err);
2848 }
2849 EXPORT_SYMBOL(rtrs_clt_open);
2850
2851 /**
2852  * rtrs_clt_close() - Close a session
2853  * @clt: Session handle. Session is freed upon return.
2854  */
2855 void rtrs_clt_close(struct rtrs_clt *clt)
2856 {
2857         struct rtrs_clt_sess *sess, *tmp;
2858
2859         /* Firstly forbid sysfs access */
2860         rtrs_clt_destroy_sysfs_root(clt);
2861
2862         /* Now it is safe to iterate over all paths without locks */
2863         list_for_each_entry_safe(sess, tmp, &clt->paths_list, s.entry) {
2864                 rtrs_clt_close_conns(sess, true);
2865                 rtrs_clt_destroy_sess_files(sess, NULL);
2866                 kobject_put(&sess->kobj);
2867         }
2868         free_clt(clt);
2869 }
2870 EXPORT_SYMBOL(rtrs_clt_close);
2871
2872 int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_sess *sess)
2873 {
2874         enum rtrs_clt_state old_state;
2875         int err = -EBUSY;
2876         bool changed;
2877
2878         changed = rtrs_clt_change_state_get_old(sess, RTRS_CLT_RECONNECTING,
2879                                                  &old_state);
2880         if (changed) {
2881                 sess->reconnect_attempts = 0;
2882                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork, 0);
2883         }
2884         if (changed || old_state == RTRS_CLT_RECONNECTING) {
2885                 /*
2886                  * flush_delayed_work() queues pending work for immediate
2887                  * execution, so do the flush if we have queued something
2888                  * right now or work is pending.
2889                  */
2890                 flush_delayed_work(&sess->reconnect_dwork);
2891                 err = (READ_ONCE(sess->state) ==
2892                        RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2893         }
2894
2895         return err;
2896 }
2897
2898 int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_sess *sess,
2899                                      const struct attribute *sysfs_self)
2900 {
2901         enum rtrs_clt_state old_state;
2902         bool changed;
2903
2904         /*
2905          * Continue stopping path till state was changed to DEAD or
2906          * state was observed as DEAD:
2907          * 1. State was changed to DEAD - we were fast and nobody
2908          *    invoked rtrs_clt_reconnect(), which can again start
2909          *    reconnecting.
2910          * 2. State was observed as DEAD - we have someone in parallel
2911          *    removing the path.
2912          */
2913         do {
2914                 rtrs_clt_close_conns(sess, true);
2915                 changed = rtrs_clt_change_state_get_old(sess,
2916                                                         RTRS_CLT_DEAD,
2917                                                         &old_state);
2918         } while (!changed && old_state != RTRS_CLT_DEAD);
2919
2920         if (changed) {
2921                 rtrs_clt_remove_path_from_arr(sess);
2922                 rtrs_clt_destroy_sess_files(sess, sysfs_self);
2923                 kobject_put(&sess->kobj);
2924         }
2925
2926         return 0;
2927 }
2928
2929 void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt *clt, int value)
2930 {
2931         clt->max_reconnect_attempts = (unsigned int)value;
2932 }
2933
2934 int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt *clt)
2935 {
2936         return (int)clt->max_reconnect_attempts;
2937 }
2938
2939 /**
2940  * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2941  *
2942  * @dir:        READ/WRITE
2943  * @ops:        callback function to be called as confirmation, and the pointer.
2944  * @clt:        Session
2945  * @permit:     Preallocated permit
2946  * @vec:        Message that is sent to server together with the request.
2947  *              Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2948  *              Since the msg is copied internally it can be allocated on stack.
2949  * @nr:         Number of elements in @vec.
2950  * @data_len:   length of data sent to/from server
2951  * @sg:         Pages to be sent/received to/from server.
2952  * @sg_cnt:     Number of elements in the @sg
2953  *
2954  * Return:
2955  * 0:           Success
2956  * <0:          Error
2957  *
2958  * On dir=READ rtrs client will request a data transfer from Server to client.
2959  * The data that the server will respond with will be stored in @sg when
2960  * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2961  * On dir=WRITE rtrs client will rdma write data in sg to server side.
2962  */
2963 int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2964                      struct rtrs_clt *clt, struct rtrs_permit *permit,
2965                       const struct kvec *vec, size_t nr, size_t data_len,
2966                       struct scatterlist *sg, unsigned int sg_cnt)
2967 {
2968         struct rtrs_clt_io_req *req;
2969         struct rtrs_clt_sess *sess;
2970
2971         enum dma_data_direction dma_dir;
2972         int err = -ECONNABORTED, i;
2973         size_t usr_len, hdr_len;
2974         struct path_it it;
2975
2976         /* Get kvec length */
2977         for (i = 0, usr_len = 0; i < nr; i++)
2978                 usr_len += vec[i].iov_len;
2979
2980         if (dir == READ) {
2981                 hdr_len = sizeof(struct rtrs_msg_rdma_read) +
2982                           sg_cnt * sizeof(struct rtrs_sg_desc);
2983                 dma_dir = DMA_FROM_DEVICE;
2984         } else {
2985                 hdr_len = sizeof(struct rtrs_msg_rdma_write);
2986                 dma_dir = DMA_TO_DEVICE;
2987         }
2988
2989         rcu_read_lock();
2990         for (path_it_init(&it, clt);
2991              (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
2992                 if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTED)
2993                         continue;
2994
2995                 if (usr_len + hdr_len > sess->max_hdr_size) {
2996                         rtrs_wrn_rl(sess->clt,
2997                                      "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
2998                                      dir == READ ? "Read" : "Write",
2999                                      usr_len, hdr_len, sess->max_hdr_size);
3000                         err = -EMSGSIZE;
3001                         break;
3002                 }
3003                 req = rtrs_clt_get_req(sess, ops->conf_fn, permit, ops->priv,
3004                                        vec, usr_len, sg, sg_cnt, data_len,
3005                                        dma_dir);
3006                 if (dir == READ)
3007                         err = rtrs_clt_read_req(req);
3008                 else
3009                         err = rtrs_clt_write_req(req);
3010                 if (err) {
3011                         req->in_use = false;
3012                         continue;
3013                 }
3014                 /* Success path */
3015                 break;
3016         }
3017         path_it_deinit(&it);
3018         rcu_read_unlock();
3019
3020         return err;
3021 }
3022 EXPORT_SYMBOL(rtrs_clt_request);
3023
3024 int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
3025 {
3026         /* If no path, return -1 for block layer not to try again */
3027         int cnt = -1;
3028         struct rtrs_con *con;
3029         struct rtrs_clt_sess *sess;
3030         struct path_it it;
3031
3032         rcu_read_lock();
3033         for (path_it_init(&it, clt);
3034              (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3035                 if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTED)
3036                         continue;
3037
3038                 con = sess->s.con[index + 1];
3039                 cnt = ib_process_cq_direct(con->cq, -1);
3040                 if (cnt)
3041                         break;
3042         }
3043         path_it_deinit(&it);
3044         rcu_read_unlock();
3045
3046         return cnt;
3047 }
3048 EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
3049
3050 /**
3051  * rtrs_clt_query() - queries RTRS session attributes
3052  *@clt: session pointer
3053  *@attr: query results for session attributes.
3054  * Returns:
3055  *    0 on success
3056  *    -ECOMM            no connection to the server
3057  */
3058 int rtrs_clt_query(struct rtrs_clt *clt, struct rtrs_attrs *attr)
3059 {
3060         if (!rtrs_clt_is_connected(clt))
3061                 return -ECOMM;
3062
3063         attr->queue_depth      = clt->queue_depth;
3064         attr->max_segments     = clt->max_segments;
3065         /* Cap max_io_size to min of remote buffer size and the fr pages */
3066         attr->max_io_size = min_t(int, clt->max_io_size,
3067                                   clt->max_segments * SZ_4K);
3068
3069         return 0;
3070 }
3071 EXPORT_SYMBOL(rtrs_clt_query);
3072
3073 int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
3074                                      struct rtrs_addr *addr)
3075 {
3076         struct rtrs_clt_sess *sess;
3077         int err;
3078
3079         sess = alloc_sess(clt, addr, nr_cpu_ids, 0);
3080         if (IS_ERR(sess))
3081                 return PTR_ERR(sess);
3082
3083         mutex_lock(&clt->paths_mutex);
3084         if (clt->paths_num == 0) {
3085                 /*
3086                  * When all the paths are removed for a session,
3087                  * the addition of the first path is like a new session for
3088                  * the storage server
3089                  */
3090                 sess->for_new_clt = 1;
3091         }
3092
3093         mutex_unlock(&clt->paths_mutex);
3094
3095         /*
3096          * It is totally safe to add path in CONNECTING state: coming
3097          * IO will never grab it.  Also it is very important to add
3098          * path before init, since init fires LINK_CONNECTED event.
3099          */
3100         rtrs_clt_add_path_to_arr(sess);
3101
3102         err = init_sess(sess);
3103         if (err)
3104                 goto close_sess;
3105
3106         err = rtrs_clt_create_sess_files(sess);
3107         if (err)
3108                 goto close_sess;
3109
3110         return 0;
3111
3112 close_sess:
3113         rtrs_clt_remove_path_from_arr(sess);
3114         rtrs_clt_close_conns(sess, true);
3115         free_percpu(sess->stats->pcpu_stats);
3116         kfree(sess->stats);
3117         free_sess(sess);
3118
3119         return err;
3120 }
3121
3122 static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
3123 {
3124         if (!(dev->ib_dev->attrs.device_cap_flags &
3125               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
3126                 pr_err("Memory registrations not supported.\n");
3127                 return -ENOTSUPP;
3128         }
3129
3130         return 0;
3131 }
3132
3133 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
3134         .init = rtrs_clt_ib_dev_init
3135 };
3136
3137 static int __init rtrs_client_init(void)
3138 {
3139         rtrs_rdma_dev_pd_init(0, &dev_pd);
3140
3141         rtrs_clt_dev_class = class_create(THIS_MODULE, "rtrs-client");
3142         if (IS_ERR(rtrs_clt_dev_class)) {
3143                 pr_err("Failed to create rtrs-client dev class\n");
3144                 return PTR_ERR(rtrs_clt_dev_class);
3145         }
3146         rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
3147         if (!rtrs_wq) {
3148                 class_destroy(rtrs_clt_dev_class);
3149                 return -ENOMEM;
3150         }
3151
3152         return 0;
3153 }
3154
3155 static void __exit rtrs_client_exit(void)
3156 {
3157         destroy_workqueue(rtrs_wq);
3158         class_destroy(rtrs_clt_dev_class);
3159         rtrs_rdma_dev_pd_deinit(&dev_pd);
3160 }
3161
3162 module_init(rtrs_client_init);
3163 module_exit(rtrs_client_exit);