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 MODULE_DESCRIPTION("RDMA Transport Client");
  35 MODULE_LICENSE("GPL");
  36
  37 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
  38 static struct rtrs_rdma_dev_pd dev_pd = {
  39         .ops = &dev_pd_ops
  40 };
  41
  42 static struct workqueue_struct *rtrs_wq;
  43 static struct class *rtrs_clt_dev_class;
  44
  45 static inline bool rtrs_clt_is_connected(const struct rtrs_clt *clt)
  46 {
  47         struct rtrs_clt_sess *sess;
  48         bool connected = false;
  49
  50         rcu_read_lock();
  51         list_for_each_entry_rcu(sess, &clt->paths_list, s.entry)
  52                 connected |= READ_ONCE(sess->state) == RTRS_CLT_CONNECTED;
  53         rcu_read_unlock();
  54
  55         return connected;
  56 }
  57
  58 static struct rtrs_permit *
  59 __rtrs_get_permit(struct rtrs_clt *clt, enum rtrs_clt_con_type con_type)
  60 {
  61         size_t max_depth = clt->queue_depth;
  62         struct rtrs_permit *permit;
  63         int bit;
  64
  65         /*
  66          * Adapted from null_blk get_tag(). Callers from different cpus may
  67          * grab the same bit, since find_first_zero_bit is not atomic.
  68          * But then the test_and_set_bit_lock will fail for all the
  69          * callers but one, so that they will loop again.
  70          * This way an explicit spinlock is not required.
  71          */
  72         do {
  73                 bit = find_first_zero_bit(clt->permits_map, max_depth);
  74                 if (unlikely(bit >= max_depth))
  75                         return NULL;
  76         } while (unlikely(test_and_set_bit_lock(bit, clt->permits_map)));
  77
  78         permit = get_permit(clt, bit);
  79         WARN_ON(permit->mem_id != bit);
  80         permit->cpu_id = raw_smp_processor_id();
  81         permit->con_type = con_type;
  82
  83         return permit;
  84 }
  85
  86 static inline void __rtrs_put_permit(struct rtrs_clt *clt,
  87                                       struct rtrs_permit *permit)
  88 {
  89         clear_bit_unlock(permit->mem_id, clt->permits_map);
  90 }
  91
  92 /**
  93  * rtrs_clt_get_permit() - allocates permit for future RDMA operation
  94  * @clt:        Current session
  95  * @con_type:   Type of connection to use with the permit
  96  * @can_wait:   Wait type
  97  *
  98  * Description:
  99  *    Allocates permit for the following RDMA operation.  Permit is used
 100  *    to preallocate all resources and to propagate memory pressure
 101  *    up earlier.
 102  *
 103  * Context:
 104  *    Can sleep if @wait == RTRS_TAG_WAIT
 105  */
 106 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt *clt,
 107                                           enum rtrs_clt_con_type con_type,
 108                                           int can_wait)
 109 {
 110         struct rtrs_permit *permit;
 111         DEFINE_WAIT(wait);
 112
 113         permit = __rtrs_get_permit(clt, con_type);
 114         if (likely(permit) || !can_wait)
 115                 return permit;
 116
 117         do {
 118                 prepare_to_wait(&clt->permits_wait, &wait,
 119                                 TASK_UNINTERRUPTIBLE);
 120                 permit = __rtrs_get_permit(clt, con_type);
 121                 if (likely(permit))
 122                         break;
 123
 124                 io_schedule();
 125         } while (1);
 126
 127         finish_wait(&clt->permits_wait, &wait);
 128
 129         return permit;
 130 }
 131 EXPORT_SYMBOL(rtrs_clt_get_permit);
 132
 133 /**
 134  * rtrs_clt_put_permit() - puts allocated permit
 135  * @clt:        Current session
 136  * @permit:     Permit to be freed
 137  *
 138  * Context:
 139  *    Does not matter
 140  */
 141 void rtrs_clt_put_permit(struct rtrs_clt *clt, struct rtrs_permit *permit)
 142 {
 143         if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
 144                 return;
 145
 146         __rtrs_put_permit(clt, permit);
 147
 148         /*
 149          * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
 150          * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
 151          * it must have added itself to &clt->permits_wait before
 152          * __rtrs_put_permit() finished.
 153          * Hence it is safe to guard wake_up() with a waitqueue_active() test.
 154          */
 155         if (waitqueue_active(&clt->permits_wait))
 156                 wake_up(&clt->permits_wait);
 157 }
 158 EXPORT_SYMBOL(rtrs_clt_put_permit);
 159
 160 void *rtrs_permit_to_pdu(struct rtrs_permit *permit)
 161 {
 162         return permit + 1;
 163 }
 164 EXPORT_SYMBOL(rtrs_permit_to_pdu);
 165
 166 /**
 167  * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
 168  * @sess: client session pointer
 169  * @permit: permit for the allocation of the RDMA buffer
 170  * Note:
 171  *     IO connection starts from 1.
 172  *     0 connection is for user messages.
 173  */
 174 static
 175 struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_sess *sess,
 176                                             struct rtrs_permit *permit)
 177 {
 178         int id = 0;
 179
 180         if (likely(permit->con_type == RTRS_IO_CON))
 181                 id = (permit->cpu_id % (sess->s.con_num - 1)) + 1;
 182
 183         return to_clt_con(sess->s.con[id]);
 184 }
 185
 186 /**
 187  * __rtrs_clt_change_state() - change the session state through session state
 188  * machine.
 189  *
 190  * @sess: client session to change the state of.
 191  * @new_state: state to change to.
 192  *
 193  * returns true if successful, false if the requested state can not be set.
 194  *
 195  * Locks:
 196  * state_wq lock must be hold.
 197  */
 198 static bool __rtrs_clt_change_state(struct rtrs_clt_sess *sess,
 199                                      enum rtrs_clt_state new_state)
 200 {
 201         enum rtrs_clt_state old_state;
 202         bool changed = false;
 203
 204         lockdep_assert_held(&sess->state_wq.lock);
 205
 206         old_state = sess->state;
 207         switch (new_state) {
 208         case RTRS_CLT_CONNECTING:
 209                 switch (old_state) {
 210                 case RTRS_CLT_RECONNECTING:
 211                         changed = true;
 212                         fallthrough;
 213                 default:
 214                         break;
 215                 }
 216                 break;
 217         case RTRS_CLT_RECONNECTING:
 218                 switch (old_state) {
 219                 case RTRS_CLT_CONNECTED:
 220                 case RTRS_CLT_CONNECTING_ERR:
 221                 case RTRS_CLT_CLOSED:
 222                         changed = true;
 223                         fallthrough;
 224                 default:
 225                         break;
 226                 }
 227                 break;
 228         case RTRS_CLT_CONNECTED:
 229                 switch (old_state) {
 230                 case RTRS_CLT_CONNECTING:
 231                         changed = true;
 232                         fallthrough;
 233                 default:
 234                         break;
 235                 }
 236                 break;
 237         case RTRS_CLT_CONNECTING_ERR:
 238                 switch (old_state) {
 239                 case RTRS_CLT_CONNECTING:
 240                         changed = true;
 241                         fallthrough;
 242                 default:
 243                         break;
 244                 }
 245                 break;
 246         case RTRS_CLT_CLOSING:
 247                 switch (old_state) {
 248                 case RTRS_CLT_CONNECTING:
 249                 case RTRS_CLT_CONNECTING_ERR:
 250                 case RTRS_CLT_RECONNECTING:
 251                 case RTRS_CLT_CONNECTED:
 252                         changed = true;
 253                         fallthrough;
 254                 default:
 255                         break;
 256                 }
 257                 break;
 258         case RTRS_CLT_CLOSED:
 259                 switch (old_state) {
 260                 case RTRS_CLT_CLOSING:
 261                         changed = true;
 262                         fallthrough;
 263                 default:
 264                         break;
 265                 }
 266                 break;
 267         case RTRS_CLT_DEAD:
 268                 switch (old_state) {
 269                 case RTRS_CLT_CLOSED:
 270                         changed = true;
 271                         fallthrough;
 272                 default:
 273                         break;
 274                 }
 275                 break;
 276         default:
 277                 break;
 278         }
 279         if (changed) {
 280                 sess->state = new_state;
 281                 wake_up_locked(&sess->state_wq);
 282         }
 283
 284         return changed;
 285 }
 286
 287 static bool rtrs_clt_change_state_from_to(struct rtrs_clt_sess *sess,
 288                                            enum rtrs_clt_state old_state,
 289                                            enum rtrs_clt_state new_state)
 290 {
 291         bool changed = false;
 292
 293         spin_lock_irq(&sess->state_wq.lock);
 294         if (sess->state == old_state)
 295                 changed = __rtrs_clt_change_state(sess, new_state);
 296         spin_unlock_irq(&sess->state_wq.lock);
 297
 298         return changed;
 299 }
 300
 301 static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
 302 {
 303         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 304
 305         if (rtrs_clt_change_state_from_to(sess,
 306                                            RTRS_CLT_CONNECTED,
 307                                            RTRS_CLT_RECONNECTING)) {
 308                 struct rtrs_clt *clt = sess->clt;
 309                 unsigned int delay_ms;
 310
 311                 /*
 312                  * Normal scenario, reconnect if we were successfully connected
 313                  */
 314                 delay_ms = clt->reconnect_delay_sec * 1000;
 315                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork,
 316                                    msecs_to_jiffies(delay_ms +
 317                                                     prandom_u32() % RTRS_RECONNECT_SEED));
 318         } else {
 319                 /*
 320                  * Error can happen just on establishing new connection,
 321                  * so notify waiter with error state, waiter is responsible
 322                  * for cleaning the rest and reconnect if needed.
 323                  */
 324                 rtrs_clt_change_state_from_to(sess,
 325                                                RTRS_CLT_CONNECTING,
 326                                                RTRS_CLT_CONNECTING_ERR);
 327         }
 328 }
 329
 330 static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
 331 {
 332         struct rtrs_clt_con *con = cq->cq_context;
 333
 334         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 335                 rtrs_err(con->c.sess, "Failed IB_WR_REG_MR: %s\n",
 336                           ib_wc_status_msg(wc->status));
 337                 rtrs_rdma_error_recovery(con);
 338         }
 339 }
 340
 341 static struct ib_cqe fast_reg_cqe = {
 342         .done = rtrs_clt_fast_reg_done
 343 };
 344
 345 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 346                               bool notify, bool can_wait);
 347
 348 static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 349 {
 350         struct rtrs_clt_io_req *req =
 351                 container_of(wc->wr_cqe, typeof(*req), inv_cqe);
 352         struct rtrs_clt_con *con = cq->cq_context;
 353
 354         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 355                 rtrs_err(con->c.sess, "Failed IB_WR_LOCAL_INV: %s\n",
 356                           ib_wc_status_msg(wc->status));
 357                 rtrs_rdma_error_recovery(con);
 358         }
 359         req->need_inv = false;
 360         if (likely(req->need_inv_comp))
 361                 complete(&req->inv_comp);
 362         else
 363                 /* Complete request from INV callback */
 364                 complete_rdma_req(req, req->inv_errno, true, false);
 365 }
 366
 367 static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
 368 {
 369         struct rtrs_clt_con *con = req->con;
 370         struct ib_send_wr wr = {
 371                 .opcode             = IB_WR_LOCAL_INV,
 372                 .wr_cqe             = &req->inv_cqe,
 373                 .send_flags         = IB_SEND_SIGNALED,
 374                 .ex.invalidate_rkey = req->mr->rkey,
 375         };
 376         req->inv_cqe.done = rtrs_clt_inv_rkey_done;
 377
 378         return ib_post_send(con->c.qp, &wr, NULL);
 379 }
 380
 381 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 382                               bool notify, bool can_wait)
 383 {
 384         struct rtrs_clt_con *con = req->con;
 385         struct rtrs_clt_sess *sess;
 386         int err;
 387
 388         if (WARN_ON(!req->in_use))
 389                 return;
 390         if (WARN_ON(!req->con))
 391                 return;
 392         sess = to_clt_sess(con->c.sess);
 393
 394         if (req->sg_cnt) {
 395                 if (unlikely(req->dir == DMA_FROM_DEVICE && req->need_inv)) {
 396                         /*
 397                          * We are here to invalidate read requests
 398                          * ourselves.  In normal scenario server should
 399                          * send INV for all read requests, but
 400                          * we are here, thus two things could happen:
 401                          *
 402                          *    1.  this is failover, when errno != 0
 403                          *        and can_wait == 1,
 404                          *
 405                          *    2.  something totally bad happened and
 406                          *        server forgot to send INV, so we
 407                          *        should do that ourselves.
 408                          */
 409
 410                         if (likely(can_wait)) {
 411                                 req->need_inv_comp = true;
 412                         } else {
 413                                 /* This should be IO path, so always notify */
 414                                 WARN_ON(!notify);
 415                                 /* Save errno for INV callback */
 416                                 req->inv_errno = errno;
 417                         }
 418
 419                         err = rtrs_inv_rkey(req);
 420                         if (unlikely(err)) {
 421                                 rtrs_err(con->c.sess, "Send INV WR key=%#x: %d\n",
 422                                           req->mr->rkey, err);
 423                         } else if (likely(can_wait)) {
 424                                 wait_for_completion(&req->inv_comp);
 425                         } else {
 426                                 /*
 427                                  * Something went wrong, so request will be
 428                                  * completed from INV callback.
 429                                  */
 430                                 WARN_ON_ONCE(1);
 431
 432                                 return;
 433                         }
 434                 }
 435                 ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
 436                                 req->sg_cnt, req->dir);
 437         }
 438         if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
 439                 atomic_dec(&sess->stats->inflight);
 440
 441         req->in_use = false;
 442         req->con = NULL;
 443
 444         if (notify)
 445                 req->conf(req->priv, errno);
 446 }
 447
 448 static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
 449                                 struct rtrs_clt_io_req *req,
 450                                 struct rtrs_rbuf *rbuf, u32 off,
 451                                 u32 imm, struct ib_send_wr *wr)
 452 {
 453         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 454         enum ib_send_flags flags;
 455         struct ib_sge sge;
 456
 457         if (unlikely(!req->sg_size)) {
 458                 rtrs_wrn(con->c.sess,
 459                          "Doing RDMA Write failed, no data supplied\n");
 460                 return -EINVAL;
 461         }
 462
 463         /* user data and user message in the first list element */
 464         sge.addr   = req->iu->dma_addr;
 465         sge.length = req->sg_size;
 466         sge.lkey   = sess->s.dev->ib_pd->local_dma_lkey;
 467
 468         /*
 469          * From time to time we have to post signalled sends,
 470          * or send queue will fill up and only QP reset can help.
 471          */
 472         flags = atomic_inc_return(&con->io_cnt) % sess->queue_depth ?
 473                         0 : IB_SEND_SIGNALED;
 474
 475         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, req->iu->dma_addr,
 476                                       req->sg_size, DMA_TO_DEVICE);
 477
 478         return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
 479                                             rbuf->rkey, rbuf->addr + off,
 480                                             imm, flags, wr);
 481 }
 482
 483 static void process_io_rsp(struct rtrs_clt_sess *sess, u32 msg_id,
 484                            s16 errno, bool w_inval)
 485 {
 486         struct rtrs_clt_io_req *req;
 487
 488         if (WARN_ON(msg_id >= sess->queue_depth))
 489                 return;
 490
 491         req = &sess->reqs[msg_id];
 492         /* Drop need_inv if server responded with send with invalidation */
 493         req->need_inv &= !w_inval;
 494         complete_rdma_req(req, errno, true, false);
 495 }
 496
 497 static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
 498 {
 499         struct rtrs_iu *iu;
 500         int err;
 501         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 502
 503         WARN_ON(sess->flags != RTRS_MSG_NEW_RKEY_F);
 504         iu = container_of(wc->wr_cqe, struct rtrs_iu,
 505                           cqe);
 506         err = rtrs_iu_post_recv(&con->c, iu);
 507         if (unlikely(err)) {
 508                 rtrs_err(con->c.sess, "post iu failed %d\n", err);
 509                 rtrs_rdma_error_recovery(con);
 510         }
 511 }
 512
 513 static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
 514 {
 515         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 516         struct rtrs_msg_rkey_rsp *msg;
 517         u32 imm_type, imm_payload;
 518         bool w_inval = false;
 519         struct rtrs_iu *iu;
 520         u32 buf_id;
 521         int err;
 522
 523         WARN_ON(sess->flags != RTRS_MSG_NEW_RKEY_F);
 524
 525         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
 526
 527         if (unlikely(wc->byte_len < sizeof(*msg))) {
 528                 rtrs_err(con->c.sess, "rkey response is malformed: size %d\n",
 529                           wc->byte_len);
 530                 goto out;
 531         }
 532         ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, iu->dma_addr,
 533                                    iu->size, DMA_FROM_DEVICE);
 534         msg = iu->buf;
 535         if (unlikely(le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP)) {
 536                 rtrs_err(sess->clt, "rkey response is malformed: type %d\n",
 537                           le16_to_cpu(msg->type));
 538                 goto out;
 539         }
 540         buf_id = le16_to_cpu(msg->buf_id);
 541         if (WARN_ON(buf_id >= sess->queue_depth))
 542                 goto out;
 543
 544         rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
 545         if (likely(imm_type == RTRS_IO_RSP_IMM ||
 546                    imm_type == RTRS_IO_RSP_W_INV_IMM)) {
 547                 u32 msg_id;
 548
 549                 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 550                 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 551
 552                 if (WARN_ON(buf_id != msg_id))
 553                         goto out;
 554                 sess->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
 555                 process_io_rsp(sess, msg_id, err, w_inval);
 556         }
 557         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, iu->dma_addr,
 558                                       iu->size, DMA_FROM_DEVICE);
 559         return rtrs_clt_recv_done(con, wc);
 560 out:
 561         rtrs_rdma_error_recovery(con);
 562 }
 563
 564 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
 565
 566 static struct ib_cqe io_comp_cqe = {
 567         .done = rtrs_clt_rdma_done
 568 };
 569
 570 /*
 571  * Post x2 empty WRs: first is for this RDMA with IMM,
 572  * second is for RECV with INV, which happened earlier.
 573  */
 574 static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
 575 {
 576         struct ib_recv_wr wr_arr[2], *wr;
 577         int i;
 578
 579         memset(wr_arr, 0, sizeof(wr_arr));
 580         for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
 581                 wr = &wr_arr[i];
 582                 wr->wr_cqe  = cqe;
 583                 if (i)
 584                         /* Chain backwards */
 585                         wr->next = &wr_arr[i - 1];
 586         }
 587
 588         return ib_post_recv(con->qp, wr, NULL);
 589 }
 590
 591 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
 592 {
 593         struct rtrs_clt_con *con = cq->cq_context;
 594         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 595         u32 imm_type, imm_payload;
 596         bool w_inval = false;
 597         int err;
 598
 599         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 600                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 601                         rtrs_err(sess->clt, "RDMA failed: %s\n",
 602                                   ib_wc_status_msg(wc->status));
 603                         rtrs_rdma_error_recovery(con);
 604                 }
 605                 return;
 606         }
 607         rtrs_clt_update_wc_stats(con);
 608
 609         switch (wc->opcode) {
 610         case IB_WC_RECV_RDMA_WITH_IMM:
 611                 /*
 612                  * post_recv() RDMA write completions of IO reqs (read/write)
 613                  * and hb
 614                  */
 615                 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
 616                         return;
 617                 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
 618                                &imm_type, &imm_payload);
 619                 if (likely(imm_type == RTRS_IO_RSP_IMM ||
 620                            imm_type == RTRS_IO_RSP_W_INV_IMM)) {
 621                         u32 msg_id;
 622
 623                         w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 624                         rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 625
 626                         process_io_rsp(sess, msg_id, err, w_inval);
 627                 } else if (imm_type == RTRS_HB_MSG_IMM) {
 628                         WARN_ON(con->c.cid);
 629                         rtrs_send_hb_ack(&sess->s);
 630                         if (sess->flags == RTRS_MSG_NEW_RKEY_F)
 631                                 return  rtrs_clt_recv_done(con, wc);
 632                 } else if (imm_type == RTRS_HB_ACK_IMM) {
 633                         WARN_ON(con->c.cid);
 634                         sess->s.hb_missed_cnt = 0;
 635                         if (sess->flags == RTRS_MSG_NEW_RKEY_F)
 636                                 return  rtrs_clt_recv_done(con, wc);
 637                 } else {
 638                         rtrs_wrn(con->c.sess, "Unknown IMM type %u\n",
 639                                   imm_type);
 640                 }
 641                 if (w_inval)
 642                         /*
 643                          * Post x2 empty WRs: first is for this RDMA with IMM,
 644                          * second is for RECV with INV, which happened earlier.
 645                          */
 646                         err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
 647                 else
 648                         err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 649                 if (unlikely(err)) {
 650                         rtrs_err(con->c.sess, "rtrs_post_recv_empty(): %d\n",
 651                                   err);
 652                         rtrs_rdma_error_recovery(con);
 653                         break;
 654                 }
 655                 break;
 656         case IB_WC_RECV:
 657                 /*
 658                  * Key invalidations from server side
 659                  */
 660                 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
 661                           wc->wc_flags & IB_WC_WITH_IMM));
 662                 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
 663                 if (sess->flags == RTRS_MSG_NEW_RKEY_F) {
 664                         if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
 665                                 return  rtrs_clt_recv_done(con, wc);
 666
 667                         return  rtrs_clt_rkey_rsp_done(con, wc);
 668                 }
 669                 break;
 670         case IB_WC_RDMA_WRITE:
 671                 /*
 672                  * post_send() RDMA write completions of IO reqs (read/write)
 673                  * and hb
 674                  */
 675                 break;
 676
 677         default:
 678                 rtrs_wrn(sess->clt, "Unexpected WC type: %d\n", wc->opcode);
 679                 return;
 680         }
 681 }
 682
 683 static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
 684 {
 685         int err, i;
 686         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 687
 688         for (i = 0; i < q_size; i++) {
 689                 if (sess->flags == RTRS_MSG_NEW_RKEY_F) {
 690                         struct rtrs_iu *iu = &con->rsp_ius[i];
 691
 692                         err = rtrs_iu_post_recv(&con->c, iu);
 693                 } else {
 694                         err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 695                 }
 696                 if (unlikely(err))
 697                         return err;
 698         }
 699
 700         return 0;
 701 }
 702
 703 static int post_recv_sess(struct rtrs_clt_sess *sess)
 704 {
 705         size_t q_size = 0;
 706         int err, cid;
 707
 708         for (cid = 0; cid < sess->s.con_num; cid++) {
 709                 if (cid == 0)
 710                         q_size = SERVICE_CON_QUEUE_DEPTH;
 711                 else
 712                         q_size = sess->queue_depth;
 713
 714                 /*
 715                  * x2 for RDMA read responses + FR key invalidations,
 716                  * RDMA writes do not require any FR registrations.
 717                  */
 718                 q_size *= 2;
 719
 720                 err = post_recv_io(to_clt_con(sess->s.con[cid]), q_size);
 721                 if (unlikely(err)) {
 722                         rtrs_err(sess->clt, "post_recv_io(), err: %d\n", err);
 723                         return err;
 724                 }
 725         }
 726
 727         return 0;
 728 }
 729
 730 struct path_it {
 731         int i;
 732         struct list_head skip_list;
 733         struct rtrs_clt *clt;
 734         struct rtrs_clt_sess *(*next_path)(struct path_it *it);
 735 };
 736
 737 /**
 738  * list_next_or_null_rr_rcu - get next list element in round-robin fashion.
 739  * @head:       the head for the list.
 740  * @ptr:        the list head to take the next element from.
 741  * @type:       the type of the struct this is embedded in.
 742  * @memb:       the name of the list_head within the struct.
 743  *
 744  * Next element returned in round-robin fashion, i.e. head will be skipped,
 745  * but if list is observed as empty, NULL will be returned.
 746  *
 747  * This primitive may safely run concurrently with the _rcu list-mutation
 748  * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
 749  */
 750 #define list_next_or_null_rr_rcu(head, ptr, type, memb) \
 751 ({ \
 752         list_next_or_null_rcu(head, ptr, type, memb) ?: \
 753                 list_next_or_null_rcu(head, READ_ONCE((ptr)->next), \
 754                                       type, memb); \
 755 })
 756
 757 /**
 758  * get_next_path_rr() - Returns path in round-robin fashion.
 759  * @it: the path pointer
 760  *
 761  * Related to @MP_POLICY_RR
 762  *
 763  * Locks:
 764  *    rcu_read_lock() must be hold.
 765  */
 766 static struct rtrs_clt_sess *get_next_path_rr(struct path_it *it)
 767 {
 768         struct rtrs_clt_sess __rcu **ppcpu_path;
 769         struct rtrs_clt_sess *path;
 770         struct rtrs_clt *clt;
 771
 772         clt = it->clt;
 773
 774         /*
 775          * Here we use two RCU objects: @paths_list and @pcpu_path
 776          * pointer.  See rtrs_clt_remove_path_from_arr() for details
 777          * how that is handled.
 778          */
 779
 780         ppcpu_path = this_cpu_ptr(clt->pcpu_path);
 781         path = rcu_dereference(*ppcpu_path);
 782         if (unlikely(!path))
 783                 path = list_first_or_null_rcu(&clt->paths_list,
 784                                               typeof(*path), s.entry);
 785         else
 786                 path = list_next_or_null_rr_rcu(&clt->paths_list,
 787                                                 &path->s.entry,
 788                                                 typeof(*path),
 789                                                 s.entry);
 790         rcu_assign_pointer(*ppcpu_path, path);
 791
 792         return path;
 793 }
 794
 795 /**
 796  * get_next_path_min_inflight() - Returns path with minimal inflight count.
 797  * @it: the path pointer
 798  *
 799  * Related to @MP_POLICY_MIN_INFLIGHT
 800  *
 801  * Locks:
 802  *    rcu_read_lock() must be hold.
 803  */
 804 static struct rtrs_clt_sess *get_next_path_min_inflight(struct path_it *it)
 805 {
 806         struct rtrs_clt_sess *min_path = NULL;
 807         struct rtrs_clt *clt = it->clt;
 808         struct rtrs_clt_sess *sess;
 809         int min_inflight = INT_MAX;
 810         int inflight;
 811
 812         list_for_each_entry_rcu(sess, &clt->paths_list, s.entry) {
 813                 if (unlikely(!list_empty(raw_cpu_ptr(sess->mp_skip_entry))))
 814                         continue;
 815
 816                 inflight = atomic_read(&sess->stats->inflight);
 817
 818                 if (inflight < min_inflight) {
 819                         min_inflight = inflight;
 820                         min_path = sess;
 821                 }
 822         }
 823
 824         /*
 825          * add the path to the skip list, so that next time we can get
 826          * a different one
 827          */
 828         if (min_path)
 829                 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
 830
 831         return min_path;
 832 }
 833
 834 static inline void path_it_init(struct path_it *it, struct rtrs_clt *clt)
 835 {
 836         INIT_LIST_HEAD(&it->skip_list);
 837         it->clt = clt;
 838         it->i = 0;
 839
 840         if (clt->mp_policy == MP_POLICY_RR)
 841                 it->next_path = get_next_path_rr;
 842         else
 843                 it->next_path = get_next_path_min_inflight;
 844 }
 845
 846 static inline void path_it_deinit(struct path_it *it)
 847 {
 848         struct list_head *skip, *tmp;
 849         /*
 850          * The skip_list is used only for the MIN_INFLIGHT policy.
 851          * We need to remove paths from it, so that next IO can insert
 852          * paths (->mp_skip_entry) into a skip_list again.
 853          */
 854         list_for_each_safe(skip, tmp, &it->skip_list)
 855                 list_del_init(skip);
 856 }
 857
 858 /**
 859  * rtrs_clt_init_req() Initialize an rtrs_clt_io_req holding information
 860  * about an inflight IO.
 861  * The user buffer holding user control message (not data) is copied into
 862  * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
 863  * also hold the control message of rtrs.
 864  * @req: an io request holding information about IO.
 865  * @sess: client session
 866  * @conf: conformation callback function to notify upper layer.
 867  * @permit: permit for allocation of RDMA remote buffer
 868  * @priv: private pointer
 869  * @vec: kernel vector containing control message
 870  * @usr_len: length of the user message
 871  * @sg: scater list for IO data
 872  * @sg_cnt: number of scater list entries
 873  * @data_len: length of the IO data
 874  * @dir: direction of the IO.
 875  */
 876 static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
 877                               struct rtrs_clt_sess *sess,
 878                               void (*conf)(void *priv, int errno),
 879                               struct rtrs_permit *permit, void *priv,
 880                               const struct kvec *vec, size_t usr_len,
 881                               struct scatterlist *sg, size_t sg_cnt,
 882                               size_t data_len, int dir)
 883 {
 884         struct iov_iter iter;
 885         size_t len;
 886
 887         req->permit = permit;
 888         req->in_use = true;
 889         req->usr_len = usr_len;
 890         req->data_len = data_len;
 891         req->sglist = sg;
 892         req->sg_cnt = sg_cnt;
 893         req->priv = priv;
 894         req->dir = dir;
 895         req->con = rtrs_permit_to_clt_con(sess, permit);
 896         req->conf = conf;
 897         req->need_inv = false;
 898         req->need_inv_comp = false;
 899         req->inv_errno = 0;
 900
 901         iov_iter_kvec(&iter, READ, vec, 1, usr_len);
 902         len = _copy_from_iter(req->iu->buf, usr_len, &iter);
 903         WARN_ON(len != usr_len);
 904
 905         reinit_completion(&req->inv_comp);
 906 }
 907
 908 static struct rtrs_clt_io_req *
 909 rtrs_clt_get_req(struct rtrs_clt_sess *sess,
 910                  void (*conf)(void *priv, int errno),
 911                  struct rtrs_permit *permit, void *priv,
 912                  const struct kvec *vec, size_t usr_len,
 913                  struct scatterlist *sg, size_t sg_cnt,
 914                  size_t data_len, int dir)
 915 {
 916         struct rtrs_clt_io_req *req;
 917
 918         req = &sess->reqs[permit->mem_id];
 919         rtrs_clt_init_req(req, sess, conf, permit, priv, vec, usr_len,
 920                            sg, sg_cnt, data_len, dir);
 921         return req;
 922 }
 923
 924 static struct rtrs_clt_io_req *
 925 rtrs_clt_get_copy_req(struct rtrs_clt_sess *alive_sess,
 926                        struct rtrs_clt_io_req *fail_req)
 927 {
 928         struct rtrs_clt_io_req *req;
 929         struct kvec vec = {
 930                 .iov_base = fail_req->iu->buf,
 931                 .iov_len  = fail_req->usr_len
 932         };
 933
 934         req = &alive_sess->reqs[fail_req->permit->mem_id];
 935         rtrs_clt_init_req(req, alive_sess, fail_req->conf, fail_req->permit,
 936                            fail_req->priv, &vec, fail_req->usr_len,
 937                            fail_req->sglist, fail_req->sg_cnt,
 938                            fail_req->data_len, fail_req->dir);
 939         return req;
 940 }
 941
 942 static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
 943                                     struct rtrs_clt_io_req *req,
 944                                     struct rtrs_rbuf *rbuf,
 945                                     u32 size, u32 imm)
 946 {
 947         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 948         struct ib_sge *sge = req->sge;
 949         enum ib_send_flags flags;
 950         struct scatterlist *sg;
 951         size_t num_sge;
 952         int i;
 953
 954         for_each_sg(req->sglist, sg, req->sg_cnt, i) {
 955                 sge[i].addr   = sg_dma_address(sg);
 956                 sge[i].length = sg_dma_len(sg);
 957                 sge[i].lkey   = sess->s.dev->ib_pd->local_dma_lkey;
 958         }
 959         sge[i].addr   = req->iu->dma_addr;
 960         sge[i].length = size;
 961         sge[i].lkey   = sess->s.dev->ib_pd->local_dma_lkey;
 962
 963         num_sge = 1 + req->sg_cnt;
 964
 965         /*
 966          * From time to time we have to post signalled sends,
 967          * or send queue will fill up and only QP reset can help.
 968          */
 969         flags = atomic_inc_return(&con->io_cnt) % sess->queue_depth ?
 970                         0 : IB_SEND_SIGNALED;
 971
 972         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, req->iu->dma_addr,
 973                                       size, DMA_TO_DEVICE);
 974
 975         return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
 976                                             rbuf->rkey, rbuf->addr, imm,
 977                                             flags, NULL);
 978 }
 979
 980 static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
 981 {
 982         struct rtrs_clt_con *con = req->con;
 983         struct rtrs_sess *s = con->c.sess;
 984         struct rtrs_clt_sess *sess = to_clt_sess(s);
 985         struct rtrs_msg_rdma_write *msg;
 986
 987         struct rtrs_rbuf *rbuf;
 988         int ret, count = 0;
 989         u32 imm, buf_id;
 990
 991         const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
 992
 993         if (unlikely(tsize > sess->chunk_size)) {
 994                 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
 995                           tsize, sess->chunk_size);
 996                 return -EMSGSIZE;
 997         }
 998         if (req->sg_cnt) {
 999                 count = ib_dma_map_sg(sess->s.dev->ib_dev, req->sglist,
1000                                       req->sg_cnt, req->dir);
1001                 if (unlikely(!count)) {
1002                         rtrs_wrn(s, "Write request failed, map failed\n");
1003                         return -EINVAL;
1004                 }
1005         }
1006         /* put rtrs msg after sg and user message */
1007         msg = req->iu->buf + req->usr_len;
1008         msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1009         msg->usr_len = cpu_to_le16(req->usr_len);
1010
1011         /* rtrs message on server side will be after user data and message */
1012         imm = req->permit->mem_off + req->data_len + req->usr_len;
1013         imm = rtrs_to_io_req_imm(imm);
1014         buf_id = req->permit->mem_id;
1015         req->sg_size = tsize;
1016         rbuf = &sess->rbufs[buf_id];
1017
1018         /*
1019          * Update stats now, after request is successfully sent it is not
1020          * safe anymore to touch it.
1021          */
1022         rtrs_clt_update_all_stats(req, WRITE);
1023
1024         ret = rtrs_post_rdma_write_sg(req->con, req, rbuf,
1025                                        req->usr_len + sizeof(*msg),
1026                                        imm);
1027         if (unlikely(ret)) {
1028                 rtrs_err(s, "Write request failed: %d\n", ret);
1029                 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
1030                         atomic_dec(&sess->stats->inflight);
1031                 if (req->sg_cnt)
1032                         ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
1033                                         req->sg_cnt, req->dir);
1034         }
1035
1036         return ret;
1037 }
1038
1039 static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1040 {
1041         int nr;
1042
1043         /* Align the MR to a 4K page size to match the block virt boundary */
1044         nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1045         if (nr < 0)
1046                 return nr;
1047         if (unlikely(nr < req->sg_cnt))
1048                 return -EINVAL;
1049         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1050
1051         return nr;
1052 }
1053
1054 static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1055 {
1056         struct rtrs_clt_con *con = req->con;
1057         struct rtrs_sess *s = con->c.sess;
1058         struct rtrs_clt_sess *sess = to_clt_sess(s);
1059         struct rtrs_msg_rdma_read *msg;
1060         struct rtrs_ib_dev *dev;
1061
1062         struct ib_reg_wr rwr;
1063         struct ib_send_wr *wr = NULL;
1064
1065         int ret, count = 0;
1066         u32 imm, buf_id;
1067
1068         const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1069
1070         s = &sess->s;
1071         dev = sess->s.dev;
1072
1073         if (unlikely(tsize > sess->chunk_size)) {
1074                 rtrs_wrn(s,
1075                           "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1076                           tsize, sess->chunk_size);
1077                 return -EMSGSIZE;
1078         }
1079
1080         if (req->sg_cnt) {
1081                 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1082                                       req->dir);
1083                 if (unlikely(!count)) {
1084                         rtrs_wrn(s,
1085                                   "Read request failed, dma map failed\n");
1086                         return -EINVAL;
1087                 }
1088         }
1089         /* put our message into req->buf after user message*/
1090         msg = req->iu->buf + req->usr_len;
1091         msg->type = cpu_to_le16(RTRS_MSG_READ);
1092         msg->usr_len = cpu_to_le16(req->usr_len);
1093
1094         if (count) {
1095                 ret = rtrs_map_sg_fr(req, count);
1096                 if (ret < 0) {
1097                         rtrs_err_rl(s,
1098                                      "Read request failed, failed to map  fast reg. data, err: %d\n",
1099                                      ret);
1100                         ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1101                                         req->dir);
1102                         return ret;
1103                 }
1104                 rwr = (struct ib_reg_wr) {
1105                         .wr.opcode = IB_WR_REG_MR,
1106                         .wr.wr_cqe = &fast_reg_cqe,
1107                         .mr = req->mr,
1108                         .key = req->mr->rkey,
1109                         .access = (IB_ACCESS_LOCAL_WRITE |
1110                                    IB_ACCESS_REMOTE_WRITE),
1111                 };
1112                 wr = &rwr.wr;
1113
1114                 msg->sg_cnt = cpu_to_le16(1);
1115                 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1116
1117                 msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1118                 msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1119                 msg->desc[0].len = cpu_to_le32(req->mr->length);
1120
1121                 /* Further invalidation is required */
1122                 req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1123
1124         } else {
1125                 msg->sg_cnt = 0;
1126                 msg->flags = 0;
1127         }
1128         /*
1129          * rtrs message will be after the space reserved for disk data and
1130          * user message
1131          */
1132         imm = req->permit->mem_off + req->data_len + req->usr_len;
1133         imm = rtrs_to_io_req_imm(imm);
1134         buf_id = req->permit->mem_id;
1135
1136         req->sg_size  = sizeof(*msg);
1137         req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1138         req->sg_size += req->usr_len;
1139
1140         /*
1141          * Update stats now, after request is successfully sent it is not
1142          * safe anymore to touch it.
1143          */
1144         rtrs_clt_update_all_stats(req, READ);
1145
1146         ret = rtrs_post_send_rdma(req->con, req, &sess->rbufs[buf_id],
1147                                    req->data_len, imm, wr);
1148         if (unlikely(ret)) {
1149                 rtrs_err(s, "Read request failed: %d\n", ret);
1150                 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
1151                         atomic_dec(&sess->stats->inflight);
1152                 req->need_inv = false;
1153                 if (req->sg_cnt)
1154                         ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1155                                         req->sg_cnt, req->dir);
1156         }
1157
1158         return ret;
1159 }
1160
1161 /**
1162  * rtrs_clt_failover_req() Try to find an active path for a failed request
1163  * @clt: clt context
1164  * @fail_req: a failed io request.
1165  */
1166 static int rtrs_clt_failover_req(struct rtrs_clt *clt,
1167                                  struct rtrs_clt_io_req *fail_req)
1168 {
1169         struct rtrs_clt_sess *alive_sess;
1170         struct rtrs_clt_io_req *req;
1171         int err = -ECONNABORTED;
1172         struct path_it it;
1173
1174         rcu_read_lock();
1175         for (path_it_init(&it, clt);
1176              (alive_sess = it.next_path(&it)) && it.i < it.clt->paths_num;
1177              it.i++) {
1178                 if (unlikely(READ_ONCE(alive_sess->state) !=
1179                              RTRS_CLT_CONNECTED))
1180                         continue;
1181                 req = rtrs_clt_get_copy_req(alive_sess, fail_req);
1182                 if (req->dir == DMA_TO_DEVICE)
1183                         err = rtrs_clt_write_req(req);
1184                 else
1185                         err = rtrs_clt_read_req(req);
1186                 if (unlikely(err)) {
1187                         req->in_use = false;
1188                         continue;
1189                 }
1190                 /* Success path */
1191                 rtrs_clt_inc_failover_cnt(alive_sess->stats);
1192                 break;
1193         }
1194         path_it_deinit(&it);
1195         rcu_read_unlock();
1196
1197         return err;
1198 }
1199
1200 static void fail_all_outstanding_reqs(struct rtrs_clt_sess *sess)
1201 {
1202         struct rtrs_clt *clt = sess->clt;
1203         struct rtrs_clt_io_req *req;
1204         int i, err;
1205
1206         if (!sess->reqs)
1207                 return;
1208         for (i = 0; i < sess->queue_depth; ++i) {
1209                 req = &sess->reqs[i];
1210                 if (!req->in_use)
1211                         continue;
1212
1213                 /*
1214                  * Safely (without notification) complete failed request.
1215                  * After completion this request is still useble and can
1216                  * be failovered to another path.
1217                  */
1218                 complete_rdma_req(req, -ECONNABORTED, false, true);
1219
1220                 err = rtrs_clt_failover_req(clt, req);
1221                 if (unlikely(err))
1222                         /* Failover failed, notify anyway */
1223                         req->conf(req->priv, err);
1224         }
1225 }
1226
1227 static void free_sess_reqs(struct rtrs_clt_sess *sess)
1228 {
1229         struct rtrs_clt_io_req *req;
1230         int i;
1231
1232         if (!sess->reqs)
1233                 return;
1234         for (i = 0; i < sess->queue_depth; ++i) {
1235                 req = &sess->reqs[i];
1236                 if (req->mr)
1237                         ib_dereg_mr(req->mr);
1238                 kfree(req->sge);
1239                 rtrs_iu_free(req->iu, DMA_TO_DEVICE,
1240                               sess->s.dev->ib_dev, 1);
1241         }
1242         kfree(sess->reqs);
1243         sess->reqs = NULL;
1244 }
1245
1246 static int alloc_sess_reqs(struct rtrs_clt_sess *sess)
1247 {
1248         struct rtrs_clt_io_req *req;
1249         struct rtrs_clt *clt = sess->clt;
1250         int i, err = -ENOMEM;
1251
1252         sess->reqs = kcalloc(sess->queue_depth, sizeof(*sess->reqs),
1253                              GFP_KERNEL);
1254         if (!sess->reqs)
1255                 return -ENOMEM;
1256
1257         for (i = 0; i < sess->queue_depth; ++i) {
1258                 req = &sess->reqs[i];
1259                 req->iu = rtrs_iu_alloc(1, sess->max_hdr_size, GFP_KERNEL,
1260                                          sess->s.dev->ib_dev,
1261                                          DMA_TO_DEVICE,
1262                                          rtrs_clt_rdma_done);
1263                 if (!req->iu)
1264                         goto out;
1265
1266                 req->sge = kmalloc_array(clt->max_segments + 1,
1267                                          sizeof(*req->sge), GFP_KERNEL);
1268                 if (!req->sge)
1269                         goto out;
1270
1271                 req->mr = ib_alloc_mr(sess->s.dev->ib_pd, IB_MR_TYPE_MEM_REG,
1272                                       sess->max_pages_per_mr);
1273                 if (IS_ERR(req->mr)) {
1274                         err = PTR_ERR(req->mr);
1275                         req->mr = NULL;
1276                         pr_err("Failed to alloc sess->max_pages_per_mr %d\n",
1277                                sess->max_pages_per_mr);
1278                         goto out;
1279                 }
1280
1281                 init_completion(&req->inv_comp);
1282         }
1283
1284         return 0;
1285
1286 out:
1287         free_sess_reqs(sess);
1288
1289         return err;
1290 }
1291
1292 static int alloc_permits(struct rtrs_clt *clt)
1293 {
1294         unsigned int chunk_bits;
1295         int err, i;
1296
1297         clt->permits_map = kcalloc(BITS_TO_LONGS(clt->queue_depth),
1298                                    sizeof(long), GFP_KERNEL);
1299         if (!clt->permits_map) {
1300                 err = -ENOMEM;
1301                 goto out_err;
1302         }
1303         clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1304         if (!clt->permits) {
1305                 err = -ENOMEM;
1306                 goto err_map;
1307         }
1308         chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1309         for (i = 0; i < clt->queue_depth; i++) {
1310                 struct rtrs_permit *permit;
1311
1312                 permit = get_permit(clt, i);
1313                 permit->mem_id = i;
1314                 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1315         }
1316
1317         return 0;
1318
1319 err_map:
1320         kfree(clt->permits_map);
1321         clt->permits_map = NULL;
1322 out_err:
1323         return err;
1324 }
1325
1326 static void free_permits(struct rtrs_clt *clt)
1327 {
1328         kfree(clt->permits_map);
1329         clt->permits_map = NULL;
1330         kfree(clt->permits);
1331         clt->permits = NULL;
1332 }
1333
1334 static void query_fast_reg_mode(struct rtrs_clt_sess *sess)
1335 {
1336         struct ib_device *ib_dev;
1337         u64 max_pages_per_mr;
1338         int mr_page_shift;
1339
1340         ib_dev = sess->s.dev->ib_dev;
1341
1342         /*
1343          * Use the smallest page size supported by the HCA, down to a
1344          * minimum of 4096 bytes. We're unlikely to build large sglists
1345          * out of smaller entries.
1346          */
1347         mr_page_shift      = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1348         max_pages_per_mr   = ib_dev->attrs.max_mr_size;
1349         do_div(max_pages_per_mr, (1ull << mr_page_shift));
1350         sess->max_pages_per_mr =
1351                 min3(sess->max_pages_per_mr, (u32)max_pages_per_mr,
1352                      ib_dev->attrs.max_fast_reg_page_list_len);
1353         sess->max_send_sge = ib_dev->attrs.max_send_sge;
1354 }
1355
1356 static bool rtrs_clt_change_state_get_old(struct rtrs_clt_sess *sess,
1357                                            enum rtrs_clt_state new_state,
1358                                            enum rtrs_clt_state *old_state)
1359 {
1360         bool changed;
1361
1362         spin_lock_irq(&sess->state_wq.lock);
1363         *old_state = sess->state;
1364         changed = __rtrs_clt_change_state(sess, new_state);
1365         spin_unlock_irq(&sess->state_wq.lock);
1366
1367         return changed;
1368 }
1369
1370 static bool rtrs_clt_change_state(struct rtrs_clt_sess *sess,
1371                                    enum rtrs_clt_state new_state)
1372 {
1373         enum rtrs_clt_state old_state;
1374
1375         return rtrs_clt_change_state_get_old(sess, new_state, &old_state);
1376 }
1377
1378 static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1379 {
1380         struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1381
1382         rtrs_rdma_error_recovery(con);
1383 }
1384
1385 static void rtrs_clt_init_hb(struct rtrs_clt_sess *sess)
1386 {
1387         rtrs_init_hb(&sess->s, &io_comp_cqe,
1388                       RTRS_HB_INTERVAL_MS,
1389                       RTRS_HB_MISSED_MAX,
1390                       rtrs_clt_hb_err_handler,
1391                       rtrs_wq);
1392 }
1393
1394 static void rtrs_clt_start_hb(struct rtrs_clt_sess *sess)
1395 {
1396         rtrs_start_hb(&sess->s);
1397 }
1398
1399 static void rtrs_clt_stop_hb(struct rtrs_clt_sess *sess)
1400 {
1401         rtrs_stop_hb(&sess->s);
1402 }
1403
1404 static void rtrs_clt_reconnect_work(struct work_struct *work);
1405 static void rtrs_clt_close_work(struct work_struct *work);
1406
1407 static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
1408                                          const struct rtrs_addr *path,
1409                                          size_t con_num, u16 max_segments,
1410                                          size_t max_segment_size)
1411 {
1412         struct rtrs_clt_sess *sess;
1413         int err = -ENOMEM;
1414         int cpu;
1415
1416         sess = kzalloc(sizeof(*sess), GFP_KERNEL);
1417         if (!sess)
1418                 goto err;
1419
1420         /* Extra connection for user messages */
1421         con_num += 1;
1422
1423         sess->s.con = kcalloc(con_num, sizeof(*sess->s.con), GFP_KERNEL);
1424         if (!sess->s.con)
1425                 goto err_free_sess;
1426
1427         sess->stats = kzalloc(sizeof(*sess->stats), GFP_KERNEL);
1428         if (!sess->stats)
1429                 goto err_free_con;
1430
1431         mutex_init(&sess->init_mutex);
1432         uuid_gen(&sess->s.uuid);
1433         memcpy(&sess->s.dst_addr, path->dst,
1434                rdma_addr_size((struct sockaddr *)path->dst));
1435
1436         /*
1437          * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1438          * checks the sa_family to be non-zero. If user passed src_addr=NULL
1439          * the sess->src_addr will contain only zeros, which is then fine.
1440          */
1441         if (path->src)
1442                 memcpy(&sess->s.src_addr, path->src,
1443                        rdma_addr_size((struct sockaddr *)path->src));
1444         strlcpy(sess->s.sessname, clt->sessname, sizeof(sess->s.sessname));
1445         sess->s.con_num = con_num;
1446         sess->clt = clt;
1447         sess->max_pages_per_mr = max_segments * max_segment_size >> 12;
1448         init_waitqueue_head(&sess->state_wq);
1449         sess->state = RTRS_CLT_CONNECTING;
1450         atomic_set(&sess->connected_cnt, 0);
1451         INIT_WORK(&sess->close_work, rtrs_clt_close_work);
1452         INIT_DELAYED_WORK(&sess->reconnect_dwork, rtrs_clt_reconnect_work);
1453         rtrs_clt_init_hb(sess);
1454
1455         sess->mp_skip_entry = alloc_percpu(typeof(*sess->mp_skip_entry));
1456         if (!sess->mp_skip_entry)
1457                 goto err_free_stats;
1458
1459         for_each_possible_cpu(cpu)
1460                 INIT_LIST_HEAD(per_cpu_ptr(sess->mp_skip_entry, cpu));
1461
1462         err = rtrs_clt_init_stats(sess->stats);
1463         if (err)
1464                 goto err_free_percpu;
1465
1466         return sess;
1467
1468 err_free_percpu:
1469         free_percpu(sess->mp_skip_entry);
1470 err_free_stats:
1471         kfree(sess->stats);
1472 err_free_con:
1473         kfree(sess->s.con);
1474 err_free_sess:
1475         kfree(sess);
1476 err:
1477         return ERR_PTR(err);
1478 }
1479
1480 void free_sess(struct rtrs_clt_sess *sess)
1481 {
1482         free_percpu(sess->mp_skip_entry);
1483         mutex_destroy(&sess->init_mutex);
1484         kfree(sess->s.con);
1485         kfree(sess->rbufs);
1486         kfree(sess);
1487 }
1488
1489 static int create_con(struct rtrs_clt_sess *sess, unsigned int cid)
1490 {
1491         struct rtrs_clt_con *con;
1492
1493         con = kzalloc(sizeof(*con), GFP_KERNEL);
1494         if (!con)
1495                 return -ENOMEM;
1496
1497         /* Map first two connections to the first CPU */
1498         con->cpu  = (cid ? cid - 1 : 0) % nr_cpu_ids;
1499         con->c.cid = cid;
1500         con->c.sess = &sess->s;
1501         atomic_set(&con->io_cnt, 0);
1502
1503         sess->s.con[cid] = &con->c;
1504
1505         return 0;
1506 }
1507
1508 static void destroy_con(struct rtrs_clt_con *con)
1509 {
1510         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1511
1512         sess->s.con[con->c.cid] = NULL;
1513         kfree(con);
1514 }
1515
1516 static int create_con_cq_qp(struct rtrs_clt_con *con)
1517 {
1518         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1519         u16 wr_queue_size;
1520         int err, cq_vector;
1521         struct rtrs_msg_rkey_rsp *rsp;
1522
1523         /*
1524          * This function can fail, but still destroy_con_cq_qp() should
1525          * be called, this is because create_con_cq_qp() is called on cm
1526          * event path, thus caller/waiter never knows: have we failed before
1527          * create_con_cq_qp() or after.  To solve this dilemma without
1528          * creating any additional flags just allow destroy_con_cq_qp() be
1529          * called many times.
1530          */
1531
1532         if (con->c.cid == 0) {
1533                 /*
1534                  * One completion for each receive and two for each send
1535                  * (send request + registration)
1536                  * + 2 for drain and heartbeat
1537                  * in case qp gets into error state
1538                  */
1539                 wr_queue_size = SERVICE_CON_QUEUE_DEPTH * 3 + 2;
1540                 /* We must be the first here */
1541                 if (WARN_ON(sess->s.dev))
1542                         return -EINVAL;
1543
1544                 /*
1545                  * The whole session uses device from user connection.
1546                  * Be careful not to close user connection before ib dev
1547                  * is gracefully put.
1548                  */
1549                 sess->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1550                                                        &dev_pd);
1551                 if (!sess->s.dev) {
1552                         rtrs_wrn(sess->clt,
1553                                   "rtrs_ib_dev_find_get_or_add(): no memory\n");
1554                         return -ENOMEM;
1555                 }
1556                 sess->s.dev_ref = 1;
1557                 query_fast_reg_mode(sess);
1558         } else {
1559                 /*
1560                  * Here we assume that session members are correctly set.
1561                  * This is always true if user connection (cid == 0) is
1562                  * established first.
1563                  */
1564                 if (WARN_ON(!sess->s.dev))
1565                         return -EINVAL;
1566                 if (WARN_ON(!sess->queue_depth))
1567                         return -EINVAL;
1568
1569                 /* Shared between connections */
1570                 sess->s.dev_ref++;
1571                 wr_queue_size =
1572                         min_t(int, sess->s.dev->ib_dev->attrs.max_qp_wr,
1573                               /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1574                               sess->queue_depth * 3 + 1);
1575         }
1576         /* alloc iu to recv new rkey reply when server reports flags set */
1577         if (sess->flags == RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1578                 con->rsp_ius = rtrs_iu_alloc(wr_queue_size, sizeof(*rsp),
1579                                               GFP_KERNEL, sess->s.dev->ib_dev,
1580                                               DMA_FROM_DEVICE,
1581                                               rtrs_clt_rdma_done);
1582                 if (!con->rsp_ius)
1583                         return -ENOMEM;
1584                 con->queue_size = wr_queue_size;
1585         }
1586         cq_vector = con->cpu % sess->s.dev->ib_dev->num_comp_vectors;
1587         err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
1588                                  cq_vector, wr_queue_size, wr_queue_size,
1589                                  IB_POLL_SOFTIRQ);
1590         /*
1591          * In case of error we do not bother to clean previous allocations,
1592          * since destroy_con_cq_qp() must be called.
1593          */
1594         return err;
1595 }
1596
1597 static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1598 {
1599         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1600
1601         /*
1602          * Be careful here: destroy_con_cq_qp() can be called even
1603          * create_con_cq_qp() failed, see comments there.
1604          */
1605
1606         rtrs_cq_qp_destroy(&con->c);
1607         if (con->rsp_ius) {
1608                 rtrs_iu_free(con->rsp_ius, DMA_FROM_DEVICE,
1609                               sess->s.dev->ib_dev, con->queue_size);
1610                 con->rsp_ius = NULL;
1611                 con->queue_size = 0;
1612         }
1613         if (sess->s.dev_ref && !--sess->s.dev_ref) {
1614                 rtrs_ib_dev_put(sess->s.dev);
1615                 sess->s.dev = NULL;
1616         }
1617 }
1618
1619 static void stop_cm(struct rtrs_clt_con *con)
1620 {
1621         rdma_disconnect(con->c.cm_id);
1622         if (con->c.qp)
1623                 ib_drain_qp(con->c.qp);
1624 }
1625
1626 static void destroy_cm(struct rtrs_clt_con *con)
1627 {
1628         rdma_destroy_id(con->c.cm_id);
1629         con->c.cm_id = NULL;
1630 }
1631
1632 static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1633 {
1634         struct rtrs_sess *s = con->c.sess;
1635         int err;
1636
1637         err = create_con_cq_qp(con);
1638         if (err) {
1639                 rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1640                 return err;
1641         }
1642         err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1643         if (err) {
1644                 rtrs_err(s, "Resolving route failed, err: %d\n", err);
1645                 destroy_con_cq_qp(con);
1646         }
1647
1648         return err;
1649 }
1650
1651 static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1652 {
1653         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1654         struct rtrs_clt *clt = sess->clt;
1655         struct rtrs_msg_conn_req msg;
1656         struct rdma_conn_param param;
1657
1658         int err;
1659
1660         param = (struct rdma_conn_param) {
1661                 .retry_count = 7,
1662                 .rnr_retry_count = 7,
1663                 .private_data = &msg,
1664                 .private_data_len = sizeof(msg),
1665         };
1666
1667         msg = (struct rtrs_msg_conn_req) {
1668                 .magic = cpu_to_le16(RTRS_MAGIC),
1669                 .version = cpu_to_le16(RTRS_PROTO_VER),
1670                 .cid = cpu_to_le16(con->c.cid),
1671                 .cid_num = cpu_to_le16(sess->s.con_num),
1672                 .recon_cnt = cpu_to_le16(sess->s.recon_cnt),
1673         };
1674         uuid_copy(&msg.sess_uuid, &sess->s.uuid);
1675         uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1676
1677         err = rdma_connect(con->c.cm_id, &param);
1678         if (err)
1679                 rtrs_err(clt, "rdma_connect(): %d\n", err);
1680
1681         return err;
1682 }
1683
1684 static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1685                                        struct rdma_cm_event *ev)
1686 {
1687         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1688         struct rtrs_clt *clt = sess->clt;
1689         const struct rtrs_msg_conn_rsp *msg;
1690         u16 version, queue_depth;
1691         int errno;
1692         u8 len;
1693
1694         msg = ev->param.conn.private_data;
1695         len = ev->param.conn.private_data_len;
1696         if (len < sizeof(*msg)) {
1697                 rtrs_err(clt, "Invalid RTRS connection response\n");
1698                 return -ECONNRESET;
1699         }
1700         if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1701                 rtrs_err(clt, "Invalid RTRS magic\n");
1702                 return -ECONNRESET;
1703         }
1704         version = le16_to_cpu(msg->version);
1705         if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1706                 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1707                           version >> 8, RTRS_PROTO_VER_MAJOR);
1708                 return -ECONNRESET;
1709         }
1710         errno = le16_to_cpu(msg->errno);
1711         if (errno) {
1712                 rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1713                           errno);
1714                 return -ECONNRESET;
1715         }
1716         if (con->c.cid == 0) {
1717                 queue_depth = le16_to_cpu(msg->queue_depth);
1718
1719                 if (queue_depth > MAX_SESS_QUEUE_DEPTH) {
1720                         rtrs_err(clt, "Invalid RTRS message: queue=%d\n",
1721                                   queue_depth);
1722                         return -ECONNRESET;
1723                 }
1724                 if (!sess->rbufs || sess->queue_depth < queue_depth) {
1725                         kfree(sess->rbufs);
1726                         sess->rbufs = kcalloc(queue_depth, sizeof(*sess->rbufs),
1727                                               GFP_KERNEL);
1728                         if (!sess->rbufs)
1729                                 return -ENOMEM;
1730                 }
1731                 sess->queue_depth = queue_depth;
1732                 sess->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1733                 sess->max_io_size = le32_to_cpu(msg->max_io_size);
1734                 sess->flags = le32_to_cpu(msg->flags);
1735                 sess->chunk_size = sess->max_io_size + sess->max_hdr_size;
1736
1737                 /*
1738                  * Global queue depth and IO size is always a minimum.
1739                  * If while a reconnection server sends us a value a bit
1740                  * higher - client does not care and uses cached minimum.
1741                  *
1742                  * Since we can have several sessions (paths) restablishing
1743                  * connections in parallel, use lock.
1744                  */
1745                 mutex_lock(&clt->paths_mutex);
1746                 clt->queue_depth = min_not_zero(sess->queue_depth,
1747                                                 clt->queue_depth);
1748                 clt->max_io_size = min_not_zero(sess->max_io_size,
1749                                                 clt->max_io_size);
1750                 mutex_unlock(&clt->paths_mutex);
1751
1752                 /*
1753                  * Cache the hca_port and hca_name for sysfs
1754                  */
1755                 sess->hca_port = con->c.cm_id->port_num;
1756                 scnprintf(sess->hca_name, sizeof(sess->hca_name),
1757                           sess->s.dev->ib_dev->name);
1758                 sess->s.src_addr = con->c.cm_id->route.addr.src_addr;
1759         }
1760
1761         return 0;
1762 }
1763
1764 static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1765 {
1766         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1767
1768         atomic_inc(&sess->connected_cnt);
1769         con->cm_err = 1;
1770 }
1771
1772 static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1773                                     struct rdma_cm_event *ev)
1774 {
1775         struct rtrs_sess *s = con->c.sess;
1776         const struct rtrs_msg_conn_rsp *msg;
1777         const char *rej_msg;
1778         int status, errno;
1779         u8 data_len;
1780
1781         status = ev->status;
1782         rej_msg = rdma_reject_msg(con->c.cm_id, status);
1783         msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1784
1785         if (msg && data_len >= sizeof(*msg)) {
1786                 errno = (int16_t)le16_to_cpu(msg->errno);
1787                 if (errno == -EBUSY)
1788                         rtrs_err(s,
1789                                   "Previous session is still exists on the server, please reconnect later\n");
1790                 else
1791                         rtrs_err(s,
1792                                   "Connect rejected: status %d (%s), rtrs errno %d\n",
1793                                   status, rej_msg, errno);
1794         } else {
1795                 rtrs_err(s,
1796                           "Connect rejected but with malformed message: status %d (%s)\n",
1797                           status, rej_msg);
1798         }
1799
1800         return -ECONNRESET;
1801 }
1802
1803 static void rtrs_clt_close_conns(struct rtrs_clt_sess *sess, bool wait)
1804 {
1805         if (rtrs_clt_change_state(sess, RTRS_CLT_CLOSING))
1806                 queue_work(rtrs_wq, &sess->close_work);
1807         if (wait)
1808                 flush_work(&sess->close_work);
1809 }
1810
1811 static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1812 {
1813         if (con->cm_err == 1) {
1814                 struct rtrs_clt_sess *sess;
1815
1816                 sess = to_clt_sess(con->c.sess);
1817                 if (atomic_dec_and_test(&sess->connected_cnt))
1818
1819                         wake_up(&sess->state_wq);
1820         }
1821         con->cm_err = cm_err;
1822 }
1823
1824 static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1825                                      struct rdma_cm_event *ev)
1826 {
1827         struct rtrs_clt_con *con = cm_id->context;
1828         struct rtrs_sess *s = con->c.sess;
1829         struct rtrs_clt_sess *sess = to_clt_sess(s);
1830         int cm_err = 0;
1831
1832         switch (ev->event) {
1833         case RDMA_CM_EVENT_ADDR_RESOLVED:
1834                 cm_err = rtrs_rdma_addr_resolved(con);
1835                 break;
1836         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1837                 cm_err = rtrs_rdma_route_resolved(con);
1838                 break;
1839         case RDMA_CM_EVENT_ESTABLISHED:
1840                 con->cm_err = rtrs_rdma_conn_established(con, ev);
1841                 if (likely(!con->cm_err)) {
1842                         /*
1843                          * Report success and wake up. Here we abuse state_wq,
1844                          * i.e. wake up without state change, but we set cm_err.
1845                          */
1846                         flag_success_on_conn(con);
1847                         wake_up(&sess->state_wq);
1848                         return 0;
1849                 }
1850                 break;
1851         case RDMA_CM_EVENT_REJECTED:
1852                 cm_err = rtrs_rdma_conn_rejected(con, ev);
1853                 break;
1854         case RDMA_CM_EVENT_CONNECT_ERROR:
1855         case RDMA_CM_EVENT_UNREACHABLE:
1856                 rtrs_wrn(s, "CM error event %d\n", ev->event);
1857                 cm_err = -ECONNRESET;
1858                 break;
1859         case RDMA_CM_EVENT_ADDR_ERROR:
1860         case RDMA_CM_EVENT_ROUTE_ERROR:
1861                 cm_err = -EHOSTUNREACH;
1862                 break;
1863         case RDMA_CM_EVENT_DISCONNECTED:
1864         case RDMA_CM_EVENT_ADDR_CHANGE:
1865         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1866                 cm_err = -ECONNRESET;
1867                 break;
1868         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1869                 /*
1870                  * Device removal is a special case.  Queue close and return 0.
1871                  */
1872                 rtrs_clt_close_conns(sess, false);
1873                 return 0;
1874         default:
1875                 rtrs_err(s, "Unexpected RDMA CM event (%d)\n", ev->event);
1876                 cm_err = -ECONNRESET;
1877                 break;
1878         }
1879
1880         if (cm_err) {
1881                 /*
1882                  * cm error makes sense only on connection establishing,
1883                  * in other cases we rely on normal procedure of reconnecting.
1884                  */
1885                 flag_error_on_conn(con, cm_err);
1886                 rtrs_rdma_error_recovery(con);
1887         }
1888
1889         return 0;
1890 }
1891
1892 static int create_cm(struct rtrs_clt_con *con)
1893 {
1894         struct rtrs_sess *s = con->c.sess;
1895         struct rtrs_clt_sess *sess = to_clt_sess(s);
1896         struct rdma_cm_id *cm_id;
1897         int err;
1898
1899         cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
1900                                sess->s.dst_addr.ss_family == AF_IB ?
1901                                RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
1902         if (IS_ERR(cm_id)) {
1903                 err = PTR_ERR(cm_id);
1904                 rtrs_err(s, "Failed to create CM ID, err: %d\n", err);
1905
1906                 return err;
1907         }
1908         con->c.cm_id = cm_id;
1909         con->cm_err = 0;
1910         /* allow the port to be reused */
1911         err = rdma_set_reuseaddr(cm_id, 1);
1912         if (err != 0) {
1913                 rtrs_err(s, "Set address reuse failed, err: %d\n", err);
1914                 goto destroy_cm;
1915         }
1916         err = rdma_resolve_addr(cm_id, (struct sockaddr *)&sess->s.src_addr,
1917                                 (struct sockaddr *)&sess->s.dst_addr,
1918                                 RTRS_CONNECT_TIMEOUT_MS);
1919         if (err) {
1920                 rtrs_err(s, "Failed to resolve address, err: %d\n", err);
1921                 goto destroy_cm;
1922         }
1923         /*
1924          * Combine connection status and session events. This is needed
1925          * for waiting two possible cases: cm_err has something meaningful
1926          * or session state was really changed to error by device removal.
1927          */
1928         err = wait_event_interruptible_timeout(
1929                         sess->state_wq,
1930                         con->cm_err || sess->state != RTRS_CLT_CONNECTING,
1931                         msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
1932         if (err == 0 || err == -ERESTARTSYS) {
1933                 if (err == 0)
1934                         err = -ETIMEDOUT;
1935                 /* Timedout or interrupted */
1936                 goto errr;
1937         }
1938         if (con->cm_err < 0) {
1939                 err = con->cm_err;
1940                 goto errr;
1941         }
1942         if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTING) {
1943                 /* Device removal */
1944                 err = -ECONNABORTED;
1945                 goto errr;
1946         }
1947
1948         return 0;
1949
1950 errr:
1951         stop_cm(con);
1952         /* Is safe to call destroy if cq_qp is not inited */
1953         destroy_con_cq_qp(con);
1954 destroy_cm:
1955         destroy_cm(con);
1956
1957         return err;
1958 }
1959
1960 static void rtrs_clt_sess_up(struct rtrs_clt_sess *sess)
1961 {
1962         struct rtrs_clt *clt = sess->clt;
1963         int up;
1964
1965         /*
1966          * We can fire RECONNECTED event only when all paths were
1967          * connected on rtrs_clt_open(), then each was disconnected
1968          * and the first one connected again.  That's why this nasty
1969          * game with counter value.
1970          */
1971
1972         mutex_lock(&clt->paths_ev_mutex);
1973         up = ++clt->paths_up;
1974         /*
1975          * Here it is safe to access paths num directly since up counter
1976          * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
1977          * in progress, thus paths removals are impossible.
1978          */
1979         if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
1980                 clt->paths_up = clt->paths_num;
1981         else if (up == 1)
1982                 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
1983         mutex_unlock(&clt->paths_ev_mutex);
1984
1985         /* Mark session as established */
1986         sess->established = true;
1987         sess->reconnect_attempts = 0;
1988         sess->stats->reconnects.successful_cnt++;
1989 }
1990
1991 static void rtrs_clt_sess_down(struct rtrs_clt_sess *sess)
1992 {
1993         struct rtrs_clt *clt = sess->clt;
1994
1995         if (!sess->established)
1996                 return;
1997
1998         sess->established = false;
1999         mutex_lock(&clt->paths_ev_mutex);
2000         WARN_ON(!clt->paths_up);
2001         if (--clt->paths_up == 0)
2002                 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2003         mutex_unlock(&clt->paths_ev_mutex);
2004 }
2005
2006 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_sess *sess)
2007 {
2008         struct rtrs_clt_con *con;
2009         unsigned int cid;
2010
2011         WARN_ON(READ_ONCE(sess->state) == RTRS_CLT_CONNECTED);
2012
2013         /*
2014          * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2015          * exactly in between.  Start destroying after it finishes.
2016          */
2017         mutex_lock(&sess->init_mutex);
2018         mutex_unlock(&sess->init_mutex);
2019
2020         /*
2021          * All IO paths must observe !CONNECTED state before we
2022          * free everything.
2023          */
2024         synchronize_rcu();
2025
2026         rtrs_clt_stop_hb(sess);
2027
2028         /*
2029          * The order it utterly crucial: firstly disconnect and complete all
2030          * rdma requests with error (thus set in_use=false for requests),
2031          * then fail outstanding requests checking in_use for each, and
2032          * eventually notify upper layer about session disconnection.
2033          */
2034
2035         for (cid = 0; cid < sess->s.con_num; cid++) {
2036                 if (!sess->s.con[cid])
2037                         break;
2038                 con = to_clt_con(sess->s.con[cid]);
2039                 stop_cm(con);
2040         }
2041         fail_all_outstanding_reqs(sess);
2042         free_sess_reqs(sess);
2043         rtrs_clt_sess_down(sess);
2044
2045         /*
2046          * Wait for graceful shutdown, namely when peer side invokes
2047          * rdma_disconnect(). 'connected_cnt' is decremented only on
2048          * CM events, thus if other side had crashed and hb has detected
2049          * something is wrong, here we will stuck for exactly timeout ms,
2050          * since CM does not fire anything.  That is fine, we are not in
2051          * hurry.
2052          */
2053         wait_event_timeout(sess->state_wq, !atomic_read(&sess->connected_cnt),
2054                            msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2055
2056         for (cid = 0; cid < sess->s.con_num; cid++) {
2057                 if (!sess->s.con[cid])
2058                         break;
2059                 con = to_clt_con(sess->s.con[cid]);
2060                 destroy_con_cq_qp(con);
2061                 destroy_cm(con);
2062                 destroy_con(con);
2063         }
2064 }
2065
2066 static inline bool xchg_sessions(struct rtrs_clt_sess __rcu **rcu_ppcpu_path,
2067                                  struct rtrs_clt_sess *sess,
2068                                  struct rtrs_clt_sess *next)
2069 {
2070         struct rtrs_clt_sess **ppcpu_path;
2071
2072         /* Call cmpxchg() without sparse warnings */
2073         ppcpu_path = (typeof(ppcpu_path))rcu_ppcpu_path;
2074         return sess == cmpxchg(ppcpu_path, sess, next);
2075 }
2076
2077 static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_sess *sess)
2078 {
2079         struct rtrs_clt *clt = sess->clt;
2080         struct rtrs_clt_sess *next;
2081         bool wait_for_grace = false;
2082         int cpu;
2083
2084         mutex_lock(&clt->paths_mutex);
2085         list_del_rcu(&sess->s.entry);
2086
2087         /* Make sure everybody observes path removal. */
2088         synchronize_rcu();
2089
2090         /*
2091          * At this point nobody sees @sess in the list, but still we have
2092          * dangling pointer @pcpu_path which _can_ point to @sess.  Since
2093          * nobody can observe @sess in the list, we guarantee that IO path
2094          * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2095          * to @sess, but can never again become @sess.
2096          */
2097
2098         /*
2099          * Decrement paths number only after grace period, because
2100          * caller of do_each_path() must firstly observe list without
2101          * path and only then decremented paths number.
2102          *
2103          * Otherwise there can be the following situation:
2104          *    o Two paths exist and IO is coming.
2105          *    o One path is removed:
2106          *      CPU#0                          CPU#1
2107          *      do_each_path():                rtrs_clt_remove_path_from_arr():
2108          *          path = get_next_path()
2109          *          ^^^                            list_del_rcu(path)
2110          *          [!CONNECTED path]              clt->paths_num--
2111          *                                              ^^^^^^^^^
2112          *          load clt->paths_num                 from 2 to 1
2113          *                    ^^^^^^^^^
2114          *                    sees 1
2115          *
2116          *      path is observed as !CONNECTED, but do_each_path() loop
2117          *      ends, because expression i < clt->paths_num is false.
2118          */
2119         clt->paths_num--;
2120
2121         /*
2122          * Get @next connection from current @sess which is going to be
2123          * removed.  If @sess is the last element, then @next is NULL.
2124          */
2125         rcu_read_lock();
2126         next = list_next_or_null_rr_rcu(&clt->paths_list, &sess->s.entry,
2127                                         typeof(*next), s.entry);
2128         rcu_read_unlock();
2129
2130         /*
2131          * @pcpu paths can still point to the path which is going to be
2132          * removed, so change the pointer manually.
2133          */
2134         for_each_possible_cpu(cpu) {
2135                 struct rtrs_clt_sess __rcu **ppcpu_path;
2136
2137                 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2138                 if (rcu_dereference_protected(*ppcpu_path,
2139                         lockdep_is_held(&clt->paths_mutex)) != sess)
2140                         /*
2141                          * synchronize_rcu() was called just after deleting
2142                          * entry from the list, thus IO code path cannot
2143                          * change pointer back to the pointer which is going
2144                          * to be removed, we are safe here.
2145                          */
2146                         continue;
2147
2148                 /*
2149                  * We race with IO code path, which also changes pointer,
2150                  * thus we have to be careful not to overwrite it.
2151                  */
2152                 if (xchg_sessions(ppcpu_path, sess, next))
2153                         /*
2154                          * @ppcpu_path was successfully replaced with @next,
2155                          * that means that someone could also pick up the
2156                          * @sess and dereferencing it right now, so wait for
2157                          * a grace period is required.
2158                          */
2159                         wait_for_grace = true;
2160         }
2161         if (wait_for_grace)
2162                 synchronize_rcu();
2163
2164         mutex_unlock(&clt->paths_mutex);
2165 }
2166
2167 static void rtrs_clt_add_path_to_arr(struct rtrs_clt_sess *sess,
2168                                       struct rtrs_addr *addr)
2169 {
2170         struct rtrs_clt *clt = sess->clt;
2171
2172         mutex_lock(&clt->paths_mutex);
2173         clt->paths_num++;
2174
2175         list_add_tail_rcu(&sess->s.entry, &clt->paths_list);
2176         mutex_unlock(&clt->paths_mutex);
2177 }
2178
2179 static void rtrs_clt_close_work(struct work_struct *work)
2180 {
2181         struct rtrs_clt_sess *sess;
2182
2183         sess = container_of(work, struct rtrs_clt_sess, close_work);
2184
2185         cancel_delayed_work_sync(&sess->reconnect_dwork);
2186         rtrs_clt_stop_and_destroy_conns(sess);
2187         rtrs_clt_change_state(sess, RTRS_CLT_CLOSED);
2188 }
2189
2190 static int init_conns(struct rtrs_clt_sess *sess)
2191 {
2192         unsigned int cid;
2193         int err;
2194
2195         /*
2196          * On every new session connections increase reconnect counter
2197          * to avoid clashes with previous sessions not yet closed
2198          * sessions on a server side.
2199          */
2200         sess->s.recon_cnt++;
2201
2202         /* Establish all RDMA connections  */
2203         for (cid = 0; cid < sess->s.con_num; cid++) {
2204                 err = create_con(sess, cid);
2205                 if (err)
2206                         goto destroy;
2207
2208                 err = create_cm(to_clt_con(sess->s.con[cid]));
2209                 if (err) {
2210                         destroy_con(to_clt_con(sess->s.con[cid]));
2211                         goto destroy;
2212                 }
2213         }
2214         err = alloc_sess_reqs(sess);
2215         if (err)
2216                 goto destroy;
2217
2218         rtrs_clt_start_hb(sess);
2219
2220         return 0;
2221
2222 destroy:
2223         while (cid--) {
2224                 struct rtrs_clt_con *con = to_clt_con(sess->s.con[cid]);
2225
2226                 stop_cm(con);
2227                 destroy_con_cq_qp(con);
2228                 destroy_cm(con);
2229                 destroy_con(con);
2230         }
2231         /*
2232          * If we've never taken async path and got an error, say,
2233          * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2234          * manually to keep reconnecting.
2235          */
2236         rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2237
2238         return err;
2239 }
2240
2241 static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2242 {
2243         struct rtrs_clt_con *con = cq->cq_context;
2244         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
2245         struct rtrs_iu *iu;
2246
2247         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2248         rtrs_iu_free(iu, DMA_TO_DEVICE, sess->s.dev->ib_dev, 1);
2249
2250         if (unlikely(wc->status != IB_WC_SUCCESS)) {
2251                 rtrs_err(sess->clt, "Sess info request send failed: %s\n",
2252                           ib_wc_status_msg(wc->status));
2253                 rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2254                 return;
2255         }
2256
2257         rtrs_clt_update_wc_stats(con);
2258 }
2259
2260 static int process_info_rsp(struct rtrs_clt_sess *sess,
2261                             const struct rtrs_msg_info_rsp *msg)
2262 {
2263         unsigned int sg_cnt, total_len;
2264         int i, sgi;
2265
2266         sg_cnt = le16_to_cpu(msg->sg_cnt);
2267         if (unlikely(!sg_cnt))
2268                 return -EINVAL;
2269         /*
2270          * Check if IB immediate data size is enough to hold the mem_id and
2271          * the offset inside the memory chunk.
2272          */
2273         if (unlikely((ilog2(sg_cnt - 1) + 1) +
2274                      (ilog2(sess->chunk_size - 1) + 1) >
2275                      MAX_IMM_PAYL_BITS)) {
2276                 rtrs_err(sess->clt,
2277                           "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2278                           MAX_IMM_PAYL_BITS, sg_cnt, sess->chunk_size);
2279                 return -EINVAL;
2280         }
2281         if (unlikely(!sg_cnt || (sess->queue_depth % sg_cnt))) {
2282                 rtrs_err(sess->clt, "Incorrect sg_cnt %d, is not multiple\n",
2283                           sg_cnt);
2284                 return -EINVAL;
2285         }
2286         total_len = 0;
2287         for (sgi = 0, i = 0; sgi < sg_cnt && i < sess->queue_depth; sgi++) {
2288                 const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2289                 u32 len, rkey;
2290                 u64 addr;
2291
2292                 addr = le64_to_cpu(desc->addr);
2293                 rkey = le32_to_cpu(desc->key);
2294                 len  = le32_to_cpu(desc->len);
2295
2296                 total_len += len;
2297
2298                 if (unlikely(!len || (len % sess->chunk_size))) {
2299                         rtrs_err(sess->clt, "Incorrect [%d].len %d\n", sgi,
2300                                   len);
2301                         return -EINVAL;
2302                 }
2303                 for ( ; len && i < sess->queue_depth; i++) {
2304                         sess->rbufs[i].addr = addr;
2305                         sess->rbufs[i].rkey = rkey;
2306
2307                         len  -= sess->chunk_size;
2308                         addr += sess->chunk_size;
2309                 }
2310         }
2311         /* Sanity check */
2312         if (unlikely(sgi != sg_cnt || i != sess->queue_depth)) {
2313                 rtrs_err(sess->clt, "Incorrect sg vector, not fully mapped\n");
2314                 return -EINVAL;
2315         }
2316         if (unlikely(total_len != sess->chunk_size * sess->queue_depth)) {
2317                 rtrs_err(sess->clt, "Incorrect total_len %d\n", total_len);
2318                 return -EINVAL;
2319         }
2320
2321         return 0;
2322 }
2323
2324 static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2325 {
2326         struct rtrs_clt_con *con = cq->cq_context;
2327         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
2328         struct rtrs_msg_info_rsp *msg;
2329         enum rtrs_clt_state state;
2330         struct rtrs_iu *iu;
2331         size_t rx_sz;
2332         int err;
2333
2334         state = RTRS_CLT_CONNECTING_ERR;
2335
2336         WARN_ON(con->c.cid);
2337         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2338         if (unlikely(wc->status != IB_WC_SUCCESS)) {
2339                 rtrs_err(sess->clt, "Sess info response recv failed: %s\n",
2340                           ib_wc_status_msg(wc->status));
2341                 goto out;
2342         }
2343         WARN_ON(wc->opcode != IB_WC_RECV);
2344
2345         if (unlikely(wc->byte_len < sizeof(*msg))) {
2346                 rtrs_err(sess->clt, "Sess info response is malformed: size %d\n",
2347                           wc->byte_len);
2348                 goto out;
2349         }
2350         ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, iu->dma_addr,
2351                                    iu->size, DMA_FROM_DEVICE);
2352         msg = iu->buf;
2353         if (unlikely(le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP)) {
2354                 rtrs_err(sess->clt, "Sess info response is malformed: type %d\n",
2355                           le16_to_cpu(msg->type));
2356                 goto out;
2357         }
2358         rx_sz  = sizeof(*msg);
2359         rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2360         if (unlikely(wc->byte_len < rx_sz)) {
2361                 rtrs_err(sess->clt, "Sess info response is malformed: size %d\n",
2362                           wc->byte_len);
2363                 goto out;
2364         }
2365         err = process_info_rsp(sess, msg);
2366         if (unlikely(err))
2367                 goto out;
2368
2369         err = post_recv_sess(sess);
2370         if (unlikely(err))
2371                 goto out;
2372
2373         state = RTRS_CLT_CONNECTED;
2374
2375 out:
2376         rtrs_clt_update_wc_stats(con);
2377         rtrs_iu_free(iu, DMA_FROM_DEVICE, sess->s.dev->ib_dev, 1);
2378         rtrs_clt_change_state(sess, state);
2379 }
2380
2381 static int rtrs_send_sess_info(struct rtrs_clt_sess *sess)
2382 {
2383         struct rtrs_clt_con *usr_con = to_clt_con(sess->s.con[0]);
2384         struct rtrs_msg_info_req *msg;
2385         struct rtrs_iu *tx_iu, *rx_iu;
2386         size_t rx_sz;
2387         int err;
2388
2389         rx_sz  = sizeof(struct rtrs_msg_info_rsp);
2390         rx_sz += sizeof(u64) * MAX_SESS_QUEUE_DEPTH;
2391
2392         tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2393                                sess->s.dev->ib_dev, DMA_TO_DEVICE,
2394                                rtrs_clt_info_req_done);
2395         rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, sess->s.dev->ib_dev,
2396                                DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2397         if (unlikely(!tx_iu || !rx_iu)) {
2398                 err = -ENOMEM;
2399                 goto out;
2400         }
2401         /* Prepare for getting info response */
2402         err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2403         if (unlikely(err)) {
2404                 rtrs_err(sess->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2405                 goto out;
2406         }
2407         rx_iu = NULL;
2408
2409         msg = tx_iu->buf;
2410         msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2411         memcpy(msg->sessname, sess->s.sessname, sizeof(msg->sessname));
2412
2413         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, tx_iu->dma_addr,
2414                                       tx_iu->size, DMA_TO_DEVICE);
2415
2416         /* Send info request */
2417         err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2418         if (unlikely(err)) {
2419                 rtrs_err(sess->clt, "rtrs_iu_post_send(), err: %d\n", err);
2420                 goto out;
2421         }
2422         tx_iu = NULL;
2423
2424         /* Wait for state change */
2425         wait_event_interruptible_timeout(sess->state_wq,
2426                                          sess->state != RTRS_CLT_CONNECTING,
2427                                          msecs_to_jiffies(
2428                                                  RTRS_CONNECT_TIMEOUT_MS));
2429         if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED)) {
2430                 if (READ_ONCE(sess->state) == RTRS_CLT_CONNECTING_ERR)
2431                         err = -ECONNRESET;
2432                 else
2433                         err = -ETIMEDOUT;
2434                 goto out;
2435         }
2436
2437 out:
2438         if (tx_iu)
2439                 rtrs_iu_free(tx_iu, DMA_TO_DEVICE, sess->s.dev->ib_dev, 1);
2440         if (rx_iu)
2441                 rtrs_iu_free(rx_iu, DMA_FROM_DEVICE, sess->s.dev->ib_dev, 1);
2442         if (unlikely(err))
2443                 /* If we've never taken async path because of malloc problems */
2444                 rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2445
2446         return err;
2447 }
2448
2449 /**
2450  * init_sess() - establishes all session connections and does handshake
2451  * @sess: client session.
2452  * In case of error full close or reconnect procedure should be taken,
2453  * because reconnect or close async works can be started.
2454  */
2455 static int init_sess(struct rtrs_clt_sess *sess)
2456 {
2457         int err;
2458
2459         mutex_lock(&sess->init_mutex);
2460         err = init_conns(sess);
2461         if (err) {
2462                 rtrs_err(sess->clt, "init_conns(), err: %d\n", err);
2463                 goto out;
2464         }
2465         err = rtrs_send_sess_info(sess);
2466         if (err) {
2467                 rtrs_err(sess->clt, "rtrs_send_sess_info(), err: %d\n", err);
2468                 goto out;
2469         }
2470         rtrs_clt_sess_up(sess);
2471 out:
2472         mutex_unlock(&sess->init_mutex);
2473
2474         return err;
2475 }
2476
2477 static void rtrs_clt_reconnect_work(struct work_struct *work)
2478 {
2479         struct rtrs_clt_sess *sess;
2480         struct rtrs_clt *clt;
2481         unsigned int delay_ms;
2482         int err;
2483
2484         sess = container_of(to_delayed_work(work), struct rtrs_clt_sess,
2485                             reconnect_dwork);
2486         clt = sess->clt;
2487
2488         if (READ_ONCE(sess->state) != RTRS_CLT_RECONNECTING)
2489                 return;
2490
2491         if (sess->reconnect_attempts >= clt->max_reconnect_attempts) {
2492                 /* Close a session completely if max attempts is reached */
2493                 rtrs_clt_close_conns(sess, false);
2494                 return;
2495         }
2496         sess->reconnect_attempts++;
2497
2498         /* Stop everything */
2499         rtrs_clt_stop_and_destroy_conns(sess);
2500         msleep(RTRS_RECONNECT_BACKOFF);
2501         if (rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING)) {
2502                 err = init_sess(sess);
2503                 if (err)
2504                         goto reconnect_again;
2505         }
2506
2507         return;
2508
2509 reconnect_again:
2510         if (rtrs_clt_change_state(sess, RTRS_CLT_RECONNECTING)) {
2511                 sess->stats->reconnects.fail_cnt++;
2512                 delay_ms = clt->reconnect_delay_sec * 1000;
2513                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork,
2514                                    msecs_to_jiffies(delay_ms +
2515                                                     prandom_u32() %
2516                                                     RTRS_RECONNECT_SEED));
2517         }
2518 }
2519
2520 static void rtrs_clt_dev_release(struct device *dev)
2521 {
2522         struct rtrs_clt *clt = container_of(dev, struct rtrs_clt, dev);
2523
2524         kfree(clt);
2525 }
2526
2527 static struct rtrs_clt *alloc_clt(const char *sessname, size_t paths_num,
2528                                   u16 port, size_t pdu_sz, void *priv,
2529                                   void  (*link_ev)(void *priv,
2530                                                    enum rtrs_clt_link_ev ev),
2531                                   unsigned int max_segments,
2532                                   size_t max_segment_size,
2533                                   unsigned int reconnect_delay_sec,
2534                                   unsigned int max_reconnect_attempts)
2535 {
2536         struct rtrs_clt *clt;
2537         int err;
2538
2539         if (!paths_num || paths_num > MAX_PATHS_NUM)
2540                 return ERR_PTR(-EINVAL);
2541
2542         if (strlen(sessname) >= sizeof(clt->sessname))
2543                 return ERR_PTR(-EINVAL);
2544
2545         clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2546         if (!clt)
2547                 return ERR_PTR(-ENOMEM);
2548
2549         clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2550         if (!clt->pcpu_path) {
2551                 kfree(clt);
2552                 return ERR_PTR(-ENOMEM);
2553         }
2554
2555         uuid_gen(&clt->paths_uuid);
2556         INIT_LIST_HEAD_RCU(&clt->paths_list);
2557         clt->paths_num = paths_num;
2558         clt->paths_up = MAX_PATHS_NUM;
2559         clt->port = port;
2560         clt->pdu_sz = pdu_sz;
2561         clt->max_segments = max_segments;
2562         clt->max_segment_size = max_segment_size;
2563         clt->reconnect_delay_sec = reconnect_delay_sec;
2564         clt->max_reconnect_attempts = max_reconnect_attempts;
2565         clt->priv = priv;
2566         clt->link_ev = link_ev;
2567         clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2568         strlcpy(clt->sessname, sessname, sizeof(clt->sessname));
2569         init_waitqueue_head(&clt->permits_wait);
2570         mutex_init(&clt->paths_ev_mutex);
2571         mutex_init(&clt->paths_mutex);
2572
2573         clt->dev.class = rtrs_clt_dev_class;
2574         clt->dev.release = rtrs_clt_dev_release;
2575         err = dev_set_name(&clt->dev, "%s", sessname);
2576         if (err) {
2577                 free_percpu(clt->pcpu_path);
2578                 kfree(clt);
2579                 return ERR_PTR(err);
2580         }
2581         /*
2582          * Suppress user space notification until
2583          * sysfs files are created
2584          */
2585         dev_set_uevent_suppress(&clt->dev, true);
2586         err = device_register(&clt->dev);
2587         if (err) {
2588                 free_percpu(clt->pcpu_path);
2589                 put_device(&clt->dev);
2590                 return ERR_PTR(err);
2591         }
2592
2593         clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2594         if (!clt->kobj_paths) {
2595                 free_percpu(clt->pcpu_path);
2596                 device_unregister(&clt->dev);
2597                 return NULL;
2598         }
2599         err = rtrs_clt_create_sysfs_root_files(clt);
2600         if (err) {
2601                 free_percpu(clt->pcpu_path);
2602                 kobject_del(clt->kobj_paths);
2603                 kobject_put(clt->kobj_paths);
2604                 device_unregister(&clt->dev);
2605                 return ERR_PTR(err);
2606         }
2607         dev_set_uevent_suppress(&clt->dev, false);
2608         kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2609
2610         return clt;
2611 }
2612
2613 static void wait_for_inflight_permits(struct rtrs_clt *clt)
2614 {
2615         if (clt->permits_map) {
2616                 size_t sz = clt->queue_depth;
2617
2618                 wait_event(clt->permits_wait,
2619                            find_first_bit(clt->permits_map, sz) >= sz);
2620         }
2621 }
2622
2623 static void free_clt(struct rtrs_clt *clt)
2624 {
2625         wait_for_inflight_permits(clt);
2626         free_permits(clt);
2627         free_percpu(clt->pcpu_path);
2628         mutex_destroy(&clt->paths_ev_mutex);
2629         mutex_destroy(&clt->paths_mutex);
2630         /* release callback will free clt in last put */
2631         device_unregister(&clt->dev);
2632 }
2633
2634 /**
2635  * rtrs_clt_open() - Open a session to an RTRS server
2636  * @ops: holds the link event callback and the private pointer.
2637  * @sessname: name of the session
2638  * @paths: Paths to be established defined by their src and dst addresses
2639  * @paths_num: Number of elements in the @paths array
2640  * @port: port to be used by the RTRS session
2641  * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2642  * @reconnect_delay_sec: time between reconnect tries
2643  * @max_segments: Max. number of segments per IO request
2644  * @max_segment_size: Max. size of one segment
2645  * @max_reconnect_attempts: Number of times to reconnect on error before giving
2646  *                          up, 0 for * disabled, -1 for forever
2647  *
2648  * Starts session establishment with the rtrs_server. The function can block
2649  * up to ~2000ms before it returns.
2650  *
2651  * Return a valid pointer on success otherwise PTR_ERR.
2652  */
2653 struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
2654                                  const char *sessname,
2655                                  const struct rtrs_addr *paths,
2656                                  size_t paths_num, u16 port,
2657                                  size_t pdu_sz, u8 reconnect_delay_sec,
2658                                  u16 max_segments,
2659                                  size_t max_segment_size,
2660                                  s16 max_reconnect_attempts)
2661 {
2662         struct rtrs_clt_sess *sess, *tmp;
2663         struct rtrs_clt *clt;
2664         int err, i;
2665
2666         clt = alloc_clt(sessname, paths_num, port, pdu_sz, ops->priv,
2667                         ops->link_ev,
2668                         max_segments, max_segment_size, reconnect_delay_sec,
2669                         max_reconnect_attempts);
2670         if (IS_ERR(clt)) {
2671                 err = PTR_ERR(clt);
2672                 goto out;
2673         }
2674         for (i = 0; i < paths_num; i++) {
2675                 struct rtrs_clt_sess *sess;
2676
2677                 sess = alloc_sess(clt, &paths[i], nr_cpu_ids,
2678                                   max_segments, max_segment_size);
2679                 if (IS_ERR(sess)) {
2680                         err = PTR_ERR(sess);
2681                         goto close_all_sess;
2682                 }
2683                 list_add_tail_rcu(&sess->s.entry, &clt->paths_list);
2684
2685                 err = init_sess(sess);
2686                 if (err) {
2687                         list_del_rcu(&sess->s.entry);
2688                         rtrs_clt_close_conns(sess, true);
2689                         free_sess(sess);
2690                         goto close_all_sess;
2691                 }
2692
2693                 err = rtrs_clt_create_sess_files(sess);
2694                 if (err) {
2695                         list_del_rcu(&sess->s.entry);
2696                         rtrs_clt_close_conns(sess, true);
2697                         free_sess(sess);
2698                         goto close_all_sess;
2699                 }
2700         }
2701         err = alloc_permits(clt);
2702         if (err)
2703                 goto close_all_sess;
2704
2705         return clt;
2706
2707 close_all_sess:
2708         list_for_each_entry_safe(sess, tmp, &clt->paths_list, s.entry) {
2709                 rtrs_clt_destroy_sess_files(sess, NULL);
2710                 rtrs_clt_close_conns(sess, true);
2711                 kobject_put(&sess->kobj);
2712         }
2713         rtrs_clt_destroy_sysfs_root_files(clt);
2714         rtrs_clt_destroy_sysfs_root_folders(clt);
2715         free_clt(clt);
2716
2717 out:
2718         return ERR_PTR(err);
2719 }
2720 EXPORT_SYMBOL(rtrs_clt_open);
2721
2722 /**
2723  * rtrs_clt_close() - Close a session
2724  * @clt: Session handle. Session is freed upon return.
2725  */
2726 void rtrs_clt_close(struct rtrs_clt *clt)
2727 {
2728         struct rtrs_clt_sess *sess, *tmp;
2729
2730         /* Firstly forbid sysfs access */
2731         rtrs_clt_destroy_sysfs_root_files(clt);
2732         rtrs_clt_destroy_sysfs_root_folders(clt);
2733
2734         /* Now it is safe to iterate over all paths without locks */
2735         list_for_each_entry_safe(sess, tmp, &clt->paths_list, s.entry) {
2736                 rtrs_clt_destroy_sess_files(sess, NULL);
2737                 rtrs_clt_close_conns(sess, true);
2738                 kobject_put(&sess->kobj);
2739         }
2740         free_clt(clt);
2741 }
2742 EXPORT_SYMBOL(rtrs_clt_close);
2743
2744 int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_sess *sess)
2745 {
2746         enum rtrs_clt_state old_state;
2747         int err = -EBUSY;
2748         bool changed;
2749
2750         changed = rtrs_clt_change_state_get_old(sess, RTRS_CLT_RECONNECTING,
2751                                                  &old_state);
2752         if (changed) {
2753                 sess->reconnect_attempts = 0;
2754                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork, 0);
2755         }
2756         if (changed || old_state == RTRS_CLT_RECONNECTING) {
2757                 /*
2758                  * flush_delayed_work() queues pending work for immediate
2759                  * execution, so do the flush if we have queued something
2760                  * right now or work is pending.
2761                  */
2762                 flush_delayed_work(&sess->reconnect_dwork);
2763                 err = (READ_ONCE(sess->state) ==
2764                        RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2765         }
2766
2767         return err;
2768 }
2769
2770 int rtrs_clt_disconnect_from_sysfs(struct rtrs_clt_sess *sess)
2771 {
2772         rtrs_clt_close_conns(sess, true);
2773
2774         return 0;
2775 }
2776
2777 int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_sess *sess,
2778                                      const struct attribute *sysfs_self)
2779 {
2780         enum rtrs_clt_state old_state;
2781         bool changed;
2782
2783         /*
2784          * Continue stopping path till state was changed to DEAD or
2785          * state was observed as DEAD:
2786          * 1. State was changed to DEAD - we were fast and nobody
2787          *    invoked rtrs_clt_reconnect(), which can again start
2788          *    reconnecting.
2789          * 2. State was observed as DEAD - we have someone in parallel
2790          *    removing the path.
2791          */
2792         do {
2793                 rtrs_clt_close_conns(sess, true);
2794                 changed = rtrs_clt_change_state_get_old(sess,
2795                                                         RTRS_CLT_DEAD,
2796                                                         &old_state);
2797         } while (!changed && old_state != RTRS_CLT_DEAD);
2798
2799         if (likely(changed)) {
2800                 rtrs_clt_destroy_sess_files(sess, sysfs_self);
2801                 rtrs_clt_remove_path_from_arr(sess);
2802                 kobject_put(&sess->kobj);
2803         }
2804
2805         return 0;
2806 }
2807
2808 void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt *clt, int value)
2809 {
2810         clt->max_reconnect_attempts = (unsigned int)value;
2811 }
2812
2813 int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt *clt)
2814 {
2815         return (int)clt->max_reconnect_attempts;
2816 }
2817
2818 /**
2819  * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2820  *
2821  * @dir:        READ/WRITE
2822  * @ops:        callback function to be called as confirmation, and the pointer.
2823  * @clt:        Session
2824  * @permit:     Preallocated permit
2825  * @vec:        Message that is sent to server together with the request.
2826  *              Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2827  *              Since the msg is copied internally it can be allocated on stack.
2828  * @nr:         Number of elements in @vec.
2829  * @data_len:   length of data sent to/from server
2830  * @sg:         Pages to be sent/received to/from server.
2831  * @sg_cnt:     Number of elements in the @sg
2832  *
2833  * Return:
2834  * 0:           Success
2835  * <0:          Error
2836  *
2837  * On dir=READ rtrs client will request a data transfer from Server to client.
2838  * The data that the server will respond with will be stored in @sg when
2839  * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2840  * On dir=WRITE rtrs client will rdma write data in sg to server side.
2841  */
2842 int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2843                      struct rtrs_clt *clt, struct rtrs_permit *permit,
2844                       const struct kvec *vec, size_t nr, size_t data_len,
2845                       struct scatterlist *sg, unsigned int sg_cnt)
2846 {
2847         struct rtrs_clt_io_req *req;
2848         struct rtrs_clt_sess *sess;
2849
2850         enum dma_data_direction dma_dir;
2851         int err = -ECONNABORTED, i;
2852         size_t usr_len, hdr_len;
2853         struct path_it it;
2854
2855         /* Get kvec length */
2856         for (i = 0, usr_len = 0; i < nr; i++)
2857                 usr_len += vec[i].iov_len;
2858
2859         if (dir == READ) {
2860                 hdr_len = sizeof(struct rtrs_msg_rdma_read) +
2861                           sg_cnt * sizeof(struct rtrs_sg_desc);
2862                 dma_dir = DMA_FROM_DEVICE;
2863         } else {
2864                 hdr_len = sizeof(struct rtrs_msg_rdma_write);
2865                 dma_dir = DMA_TO_DEVICE;
2866         }
2867
2868         rcu_read_lock();
2869         for (path_it_init(&it, clt);
2870              (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
2871                 if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
2872                         continue;
2873
2874                 if (unlikely(usr_len + hdr_len > sess->max_hdr_size)) {
2875                         rtrs_wrn_rl(sess->clt,
2876                                      "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
2877                                      dir == READ ? "Read" : "Write",
2878                                      usr_len, hdr_len, sess->max_hdr_size);
2879                         err = -EMSGSIZE;
2880                         break;
2881                 }
2882                 req = rtrs_clt_get_req(sess, ops->conf_fn, permit, ops->priv,
2883                                        vec, usr_len, sg, sg_cnt, data_len,
2884                                        dma_dir);
2885                 if (dir == READ)
2886                         err = rtrs_clt_read_req(req);
2887                 else
2888                         err = rtrs_clt_write_req(req);
2889                 if (unlikely(err)) {
2890                         req->in_use = false;
2891                         continue;
2892                 }
2893                 /* Success path */
2894                 break;
2895         }
2896         path_it_deinit(&it);
2897         rcu_read_unlock();
2898
2899         return err;
2900 }
2901 EXPORT_SYMBOL(rtrs_clt_request);
2902
2903 /**
2904  * rtrs_clt_query() - queries RTRS session attributes
2905  *@clt: session pointer
2906  *@attr: query results for session attributes.
2907  * Returns:
2908  *    0 on success
2909  *    -ECOMM            no connection to the server
2910  */
2911 int rtrs_clt_query(struct rtrs_clt *clt, struct rtrs_attrs *attr)
2912 {
2913         if (!rtrs_clt_is_connected(clt))
2914                 return -ECOMM;
2915
2916         attr->queue_depth      = clt->queue_depth;
2917         attr->max_io_size      = clt->max_io_size;
2918         attr->sess_kobj        = &clt->dev.kobj;
2919         strlcpy(attr->sessname, clt->sessname, sizeof(attr->sessname));
2920
2921         return 0;
2922 }
2923 EXPORT_SYMBOL(rtrs_clt_query);
2924
2925 int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
2926                                      struct rtrs_addr *addr)
2927 {
2928         struct rtrs_clt_sess *sess;
2929         int err;
2930
2931         sess = alloc_sess(clt, addr, nr_cpu_ids, clt->max_segments,
2932                           clt->max_segment_size);
2933         if (IS_ERR(sess))
2934                 return PTR_ERR(sess);
2935
2936         /*
2937          * It is totally safe to add path in CONNECTING state: coming
2938          * IO will never grab it.  Also it is very important to add
2939          * path before init, since init fires LINK_CONNECTED event.
2940          */
2941         rtrs_clt_add_path_to_arr(sess, addr);
2942
2943         err = init_sess(sess);
2944         if (err)
2945                 goto close_sess;
2946
2947         err = rtrs_clt_create_sess_files(sess);
2948         if (err)
2949                 goto close_sess;
2950
2951         return 0;
2952
2953 close_sess:
2954         rtrs_clt_remove_path_from_arr(sess);
2955         rtrs_clt_close_conns(sess, true);
2956         free_sess(sess);
2957
2958         return err;
2959 }
2960
2961 static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
2962 {
2963         if (!(dev->ib_dev->attrs.device_cap_flags &
2964               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
2965                 pr_err("Memory registrations not supported.\n");
2966                 return -ENOTSUPP;
2967         }
2968
2969         return 0;
2970 }
2971
2972 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
2973         .init = rtrs_clt_ib_dev_init
2974 };
2975
2976 static int __init rtrs_client_init(void)
2977 {
2978         rtrs_rdma_dev_pd_init(0, &dev_pd);
2979
2980         rtrs_clt_dev_class = class_create(THIS_MODULE, "rtrs-client");
2981         if (IS_ERR(rtrs_clt_dev_class)) {
2982                 pr_err("Failed to create rtrs-client dev class\n");
2983                 return PTR_ERR(rtrs_clt_dev_class);
2984         }
2985         rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
2986         if (!rtrs_wq) {
2987                 class_destroy(rtrs_clt_dev_class);
2988                 return -ENOMEM;
2989         }
2990
2991         return 0;
2992 }
2993
2994 static void __exit rtrs_client_exit(void)
2995 {
2996         destroy_workqueue(rtrs_wq);
2997         class_destroy(rtrs_clt_dev_class);
2998         rtrs_rdma_dev_pd_deinit(&dev_pd);
2999 }
3000
3001 module_init(rtrs_client_init);
3002 module_exit(rtrs_client_exit);