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

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * RDMA Transport Layer
   4  *
   5  * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved.
   6  * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved.
   7  * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved.
   8  */
   9
  10 #undef pr_fmt
  11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
  12
  13 #include <linux/module.h>
  14 #include <linux/rculist.h>
  15 #include <linux/random.h>
  16
  17 #include "rtrs-clt.h"
  18 #include "rtrs-log.h"
  19
  20 #define RTRS_CONNECT_TIMEOUT_MS 30000
  21 /*
  22  * Wait a bit before trying to reconnect after a failure
  23  * in order to give server time to finish clean up which
  24  * leads to "false positives" failed reconnect attempts
  25  */
  26 #define RTRS_RECONNECT_BACKOFF 1000
  27 /*
  28  * Wait for additional random time between 0 and 8 seconds
  29  * before starting to reconnect to avoid clients reconnecting
  30  * all at once in case of a major network outage
  31  */
  32 #define RTRS_RECONNECT_SEED 8
  33
  34 #define FIRST_CONN 0x01
  35
  36 MODULE_DESCRIPTION("RDMA Transport Client");
  37 MODULE_LICENSE("GPL");
  38
  39 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
  40 static struct rtrs_rdma_dev_pd dev_pd = {
  41         .ops = &dev_pd_ops
  42 };
  43
  44 static struct workqueue_struct *rtrs_wq;
  45 static struct class *rtrs_clt_dev_class;
  46
  47 static inline bool rtrs_clt_is_connected(const struct rtrs_clt *clt)
  48 {
  49         struct rtrs_clt_sess *sess;
  50         bool connected = false;
  51
  52         rcu_read_lock();
  53         list_for_each_entry_rcu(sess, &clt->paths_list, s.entry)
  54                 connected |= READ_ONCE(sess->state) == RTRS_CLT_CONNECTED;
  55         rcu_read_unlock();
  56
  57         return connected;
  58 }
  59
  60 static struct rtrs_permit *
  61 __rtrs_get_permit(struct rtrs_clt *clt, enum rtrs_clt_con_type con_type)
  62 {
  63         size_t max_depth = clt->queue_depth;
  64         struct rtrs_permit *permit;
  65         int bit;
  66
  67         /*
  68          * Adapted from null_blk get_tag(). Callers from different cpus may
  69          * grab the same bit, since find_first_zero_bit is not atomic.
  70          * But then the test_and_set_bit_lock will fail for all the
  71          * callers but one, so that they will loop again.
  72          * This way an explicit spinlock is not required.
  73          */
  74         do {
  75                 bit = find_first_zero_bit(clt->permits_map, max_depth);
  76                 if (unlikely(bit >= max_depth))
  77                         return NULL;
  78         } while (unlikely(test_and_set_bit_lock(bit, clt->permits_map)));
  79
  80         permit = get_permit(clt, bit);
  81         WARN_ON(permit->mem_id != bit);
  82         permit->cpu_id = raw_smp_processor_id();
  83         permit->con_type = con_type;
  84
  85         return permit;
  86 }
  87
  88 static inline void __rtrs_put_permit(struct rtrs_clt *clt,
  89                                       struct rtrs_permit *permit)
  90 {
  91         clear_bit_unlock(permit->mem_id, clt->permits_map);
  92 }
  93
  94 /**
  95  * rtrs_clt_get_permit() - allocates permit for future RDMA operation
  96  * @clt:        Current session
  97  * @con_type:   Type of connection to use with the permit
  98  * @can_wait:   Wait type
  99  *
 100  * Description:
 101  *    Allocates permit for the following RDMA operation.  Permit is used
 102  *    to preallocate all resources and to propagate memory pressure
 103  *    up earlier.
 104  *
 105  * Context:
 106  *    Can sleep if @wait == RTRS_TAG_WAIT
 107  */
 108 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt *clt,
 109                                           enum rtrs_clt_con_type con_type,
 110                                           int can_wait)
 111 {
 112         struct rtrs_permit *permit;
 113         DEFINE_WAIT(wait);
 114
 115         permit = __rtrs_get_permit(clt, con_type);
 116         if (likely(permit) || !can_wait)
 117                 return permit;
 118
 119         do {
 120                 prepare_to_wait(&clt->permits_wait, &wait,
 121                                 TASK_UNINTERRUPTIBLE);
 122                 permit = __rtrs_get_permit(clt, con_type);
 123                 if (likely(permit))
 124                         break;
 125
 126                 io_schedule();
 127         } while (1);
 128
 129         finish_wait(&clt->permits_wait, &wait);
 130
 131         return permit;
 132 }
 133 EXPORT_SYMBOL(rtrs_clt_get_permit);
 134
 135 /**
 136  * rtrs_clt_put_permit() - puts allocated permit
 137  * @clt:        Current session
 138  * @permit:     Permit to be freed
 139  *
 140  * Context:
 141  *    Does not matter
 142  */
 143 void rtrs_clt_put_permit(struct rtrs_clt *clt, struct rtrs_permit *permit)
 144 {
 145         if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
 146                 return;
 147
 148         __rtrs_put_permit(clt, permit);
 149
 150         /*
 151          * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
 152          * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
 153          * it must have added itself to &clt->permits_wait before
 154          * __rtrs_put_permit() finished.
 155          * Hence it is safe to guard wake_up() with a waitqueue_active() test.
 156          */
 157         if (waitqueue_active(&clt->permits_wait))
 158                 wake_up(&clt->permits_wait);
 159 }
 160 EXPORT_SYMBOL(rtrs_clt_put_permit);
 161
 162 void *rtrs_permit_to_pdu(struct rtrs_permit *permit)
 163 {
 164         return permit + 1;
 165 }
 166 EXPORT_SYMBOL(rtrs_permit_to_pdu);
 167
 168 /**
 169  * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
 170  * @sess: client session pointer
 171  * @permit: permit for the allocation of the RDMA buffer
 172  * Note:
 173  *     IO connection starts from 1.
 174  *     0 connection is for user messages.
 175  */
 176 static
 177 struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_sess *sess,
 178                                             struct rtrs_permit *permit)
 179 {
 180         int id = 0;
 181
 182         if (likely(permit->con_type == RTRS_IO_CON))
 183                 id = (permit->cpu_id % (sess->s.con_num - 1)) + 1;
 184
 185         return to_clt_con(sess->s.con[id]);
 186 }
 187
 188 /**
 189  * __rtrs_clt_change_state() - change the session state through session state
 190  * machine.
 191  *
 192  * @sess: client session to change the state of.
 193  * @new_state: state to change to.
 194  *
 195  * returns true if successful, false if the requested state can not be set.
 196  *
 197  * Locks:
 198  * state_wq lock must be hold.
 199  */
 200 static bool __rtrs_clt_change_state(struct rtrs_clt_sess *sess,
 201                                      enum rtrs_clt_state new_state)
 202 {
 203         enum rtrs_clt_state old_state;
 204         bool changed = false;
 205
 206         lockdep_assert_held(&sess->state_wq.lock);
 207
 208         old_state = sess->state;
 209         switch (new_state) {
 210         case RTRS_CLT_CONNECTING:
 211                 switch (old_state) {
 212                 case RTRS_CLT_RECONNECTING:
 213                         changed = true;
 214                         fallthrough;
 215                 default:
 216                         break;
 217                 }
 218                 break;
 219         case RTRS_CLT_RECONNECTING:
 220                 switch (old_state) {
 221                 case RTRS_CLT_CONNECTED:
 222                 case RTRS_CLT_CONNECTING_ERR:
 223                 case RTRS_CLT_CLOSED:
 224                         changed = true;
 225                         fallthrough;
 226                 default:
 227                         break;
 228                 }
 229                 break;
 230         case RTRS_CLT_CONNECTED:
 231                 switch (old_state) {
 232                 case RTRS_CLT_CONNECTING:
 233                         changed = true;
 234                         fallthrough;
 235                 default:
 236                         break;
 237                 }
 238                 break;
 239         case RTRS_CLT_CONNECTING_ERR:
 240                 switch (old_state) {
 241                 case RTRS_CLT_CONNECTING:
 242                         changed = true;
 243                         fallthrough;
 244                 default:
 245                         break;
 246                 }
 247                 break;
 248         case RTRS_CLT_CLOSING:
 249                 switch (old_state) {
 250                 case RTRS_CLT_CONNECTING:
 251                 case RTRS_CLT_CONNECTING_ERR:
 252                 case RTRS_CLT_RECONNECTING:
 253                 case RTRS_CLT_CONNECTED:
 254                         changed = true;
 255                         fallthrough;
 256                 default:
 257                         break;
 258                 }
 259                 break;
 260         case RTRS_CLT_CLOSED:
 261                 switch (old_state) {
 262                 case RTRS_CLT_CLOSING:
 263                         changed = true;
 264                         fallthrough;
 265                 default:
 266                         break;
 267                 }
 268                 break;
 269         case RTRS_CLT_DEAD:
 270                 switch (old_state) {
 271                 case RTRS_CLT_CLOSED:
 272                         changed = true;
 273                         fallthrough;
 274                 default:
 275                         break;
 276                 }
 277                 break;
 278         default:
 279                 break;
 280         }
 281         if (changed) {
 282                 sess->state = new_state;
 283                 wake_up_locked(&sess->state_wq);
 284         }
 285
 286         return changed;
 287 }
 288
 289 static bool rtrs_clt_change_state_from_to(struct rtrs_clt_sess *sess,
 290                                            enum rtrs_clt_state old_state,
 291                                            enum rtrs_clt_state new_state)
 292 {
 293         bool changed = false;
 294
 295         spin_lock_irq(&sess->state_wq.lock);
 296         if (sess->state == old_state)
 297                 changed = __rtrs_clt_change_state(sess, new_state);
 298         spin_unlock_irq(&sess->state_wq.lock);
 299
 300         return changed;
 301 }
 302
 303 static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
 304 {
 305         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 306
 307         if (rtrs_clt_change_state_from_to(sess,
 308                                            RTRS_CLT_CONNECTED,
 309                                            RTRS_CLT_RECONNECTING)) {
 310                 struct rtrs_clt *clt = sess->clt;
 311                 unsigned int delay_ms;
 312
 313                 /*
 314                  * Normal scenario, reconnect if we were successfully connected
 315                  */
 316                 delay_ms = clt->reconnect_delay_sec * 1000;
 317                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork,
 318                                    msecs_to_jiffies(delay_ms +
 319                                                     prandom_u32() % RTRS_RECONNECT_SEED));
 320         } else {
 321                 /*
 322                  * Error can happen just on establishing new connection,
 323                  * so notify waiter with error state, waiter is responsible
 324                  * for cleaning the rest and reconnect if needed.
 325                  */
 326                 rtrs_clt_change_state_from_to(sess,
 327                                                RTRS_CLT_CONNECTING,
 328                                                RTRS_CLT_CONNECTING_ERR);
 329         }
 330 }
 331
 332 static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
 333 {
 334         struct rtrs_clt_con *con = cq->cq_context;
 335
 336         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 337                 rtrs_err(con->c.sess, "Failed IB_WR_REG_MR: %s\n",
 338                           ib_wc_status_msg(wc->status));
 339                 rtrs_rdma_error_recovery(con);
 340         }
 341 }
 342
 343 static struct ib_cqe fast_reg_cqe = {
 344         .done = rtrs_clt_fast_reg_done
 345 };
 346
 347 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 348                               bool notify, bool can_wait);
 349
 350 static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 351 {
 352         struct rtrs_clt_io_req *req =
 353                 container_of(wc->wr_cqe, typeof(*req), inv_cqe);
 354         struct rtrs_clt_con *con = cq->cq_context;
 355
 356         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 357                 rtrs_err(con->c.sess, "Failed IB_WR_LOCAL_INV: %s\n",
 358                           ib_wc_status_msg(wc->status));
 359                 rtrs_rdma_error_recovery(con);
 360         }
 361         req->need_inv = false;
 362         if (likely(req->need_inv_comp))
 363                 complete(&req->inv_comp);
 364         else
 365                 /* Complete request from INV callback */
 366                 complete_rdma_req(req, req->inv_errno, true, false);
 367 }
 368
 369 static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
 370 {
 371         struct rtrs_clt_con *con = req->con;
 372         struct ib_send_wr wr = {
 373                 .opcode             = IB_WR_LOCAL_INV,
 374                 .wr_cqe             = &req->inv_cqe,
 375                 .send_flags         = IB_SEND_SIGNALED,
 376                 .ex.invalidate_rkey = req->mr->rkey,
 377         };
 378         req->inv_cqe.done = rtrs_clt_inv_rkey_done;
 379
 380         return ib_post_send(con->c.qp, &wr, NULL);
 381 }
 382
 383 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 384                               bool notify, bool can_wait)
 385 {
 386         struct rtrs_clt_con *con = req->con;
 387         struct rtrs_clt_sess *sess;
 388         int err;
 389
 390         if (WARN_ON(!req->in_use))
 391                 return;
 392         if (WARN_ON(!req->con))
 393                 return;
 394         sess = to_clt_sess(con->c.sess);
 395
 396         if (req->sg_cnt) {
 397                 if (unlikely(req->dir == DMA_FROM_DEVICE && req->need_inv)) {
 398                         /*
 399                          * We are here to invalidate read requests
 400                          * ourselves.  In normal scenario server should
 401                          * send INV for all read requests, but
 402                          * we are here, thus two things could happen:
 403                          *
 404                          *    1.  this is failover, when errno != 0
 405                          *        and can_wait == 1,
 406                          *
 407                          *    2.  something totally bad happened and
 408                          *        server forgot to send INV, so we
 409                          *        should do that ourselves.
 410                          */
 411
 412                         if (likely(can_wait)) {
 413                                 req->need_inv_comp = true;
 414                         } else {
 415                                 /* This should be IO path, so always notify */
 416                                 WARN_ON(!notify);
 417                                 /* Save errno for INV callback */
 418                                 req->inv_errno = errno;
 419                         }
 420
 421                         err = rtrs_inv_rkey(req);
 422                         if (unlikely(err)) {
 423                                 rtrs_err(con->c.sess, "Send INV WR key=%#x: %d\n",
 424                                           req->mr->rkey, err);
 425                         } else if (likely(can_wait)) {
 426                                 wait_for_completion(&req->inv_comp);
 427                         } else {
 428                                 /*
 429                                  * Something went wrong, so request will be
 430                                  * completed from INV callback.
 431                                  */
 432                                 WARN_ON_ONCE(1);
 433
 434                                 return;
 435                         }
 436                 }
 437                 ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
 438                                 req->sg_cnt, req->dir);
 439         }
 440         if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
 441                 atomic_dec(&sess->stats->inflight);
 442
 443         req->in_use = false;
 444         req->con = NULL;
 445
 446         if (notify)
 447                 req->conf(req->priv, errno);
 448 }
 449
 450 static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
 451                                 struct rtrs_clt_io_req *req,
 452                                 struct rtrs_rbuf *rbuf, u32 off,
 453                                 u32 imm, struct ib_send_wr *wr)
 454 {
 455         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 456         enum ib_send_flags flags;
 457         struct ib_sge sge;
 458
 459         if (unlikely(!req->sg_size)) {
 460                 rtrs_wrn(con->c.sess,
 461                          "Doing RDMA Write failed, no data supplied\n");
 462                 return -EINVAL;
 463         }
 464
 465         /* user data and user message in the first list element */
 466         sge.addr   = req->iu->dma_addr;
 467         sge.length = req->sg_size;
 468         sge.lkey   = sess->s.dev->ib_pd->local_dma_lkey;
 469
 470         /*
 471          * From time to time we have to post signalled sends,
 472          * or send queue will fill up and only QP reset can help.
 473          */
 474         flags = atomic_inc_return(&con->io_cnt) % sess->queue_depth ?
 475                         0 : IB_SEND_SIGNALED;
 476
 477         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, req->iu->dma_addr,
 478                                       req->sg_size, DMA_TO_DEVICE);
 479
 480         return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
 481                                             rbuf->rkey, rbuf->addr + off,
 482                                             imm, flags, wr);
 483 }
 484
 485 static void process_io_rsp(struct rtrs_clt_sess *sess, u32 msg_id,
 486                            s16 errno, bool w_inval)
 487 {
 488         struct rtrs_clt_io_req *req;
 489
 490         if (WARN_ON(msg_id >= sess->queue_depth))
 491                 return;
 492
 493         req = &sess->reqs[msg_id];
 494         /* Drop need_inv if server responded with send with invalidation */
 495         req->need_inv &= !w_inval;
 496         complete_rdma_req(req, errno, true, false);
 497 }
 498
 499 static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
 500 {
 501         struct rtrs_iu *iu;
 502         int err;
 503         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 504
 505         WARN_ON((sess->flags & RTRS_MSG_NEW_RKEY_F) == 0);
 506         iu = container_of(wc->wr_cqe, struct rtrs_iu,
 507                           cqe);
 508         err = rtrs_iu_post_recv(&con->c, iu);
 509         if (unlikely(err)) {
 510                 rtrs_err(con->c.sess, "post iu failed %d\n", err);
 511                 rtrs_rdma_error_recovery(con);
 512         }
 513 }
 514
 515 static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
 516 {
 517         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 518         struct rtrs_msg_rkey_rsp *msg;
 519         u32 imm_type, imm_payload;
 520         bool w_inval = false;
 521         struct rtrs_iu *iu;
 522         u32 buf_id;
 523         int err;
 524
 525         WARN_ON((sess->flags & RTRS_MSG_NEW_RKEY_F) == 0);
 526
 527         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
 528
 529         if (unlikely(wc->byte_len < sizeof(*msg))) {
 530                 rtrs_err(con->c.sess, "rkey response is malformed: size %d\n",
 531                           wc->byte_len);
 532                 goto out;
 533         }
 534         ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, iu->dma_addr,
 535                                    iu->size, DMA_FROM_DEVICE);
 536         msg = iu->buf;
 537         if (unlikely(le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP)) {
 538                 rtrs_err(sess->clt, "rkey response is malformed: type %d\n",
 539                           le16_to_cpu(msg->type));
 540                 goto out;
 541         }
 542         buf_id = le16_to_cpu(msg->buf_id);
 543         if (WARN_ON(buf_id >= sess->queue_depth))
 544                 goto out;
 545
 546         rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
 547         if (likely(imm_type == RTRS_IO_RSP_IMM ||
 548                    imm_type == RTRS_IO_RSP_W_INV_IMM)) {
 549                 u32 msg_id;
 550
 551                 w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 552                 rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 553
 554                 if (WARN_ON(buf_id != msg_id))
 555                         goto out;
 556                 sess->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
 557                 process_io_rsp(sess, msg_id, err, w_inval);
 558         }
 559         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, iu->dma_addr,
 560                                       iu->size, DMA_FROM_DEVICE);
 561         return rtrs_clt_recv_done(con, wc);
 562 out:
 563         rtrs_rdma_error_recovery(con);
 564 }
 565
 566 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
 567
 568 static struct ib_cqe io_comp_cqe = {
 569         .done = rtrs_clt_rdma_done
 570 };
 571
 572 /*
 573  * Post x2 empty WRs: first is for this RDMA with IMM,
 574  * second is for RECV with INV, which happened earlier.
 575  */
 576 static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
 577 {
 578         struct ib_recv_wr wr_arr[2], *wr;
 579         int i;
 580
 581         memset(wr_arr, 0, sizeof(wr_arr));
 582         for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
 583                 wr = &wr_arr[i];
 584                 wr->wr_cqe  = cqe;
 585                 if (i)
 586                         /* Chain backwards */
 587                         wr->next = &wr_arr[i - 1];
 588         }
 589
 590         return ib_post_recv(con->qp, wr, NULL);
 591 }
 592
 593 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
 594 {
 595         struct rtrs_clt_con *con = cq->cq_context;
 596         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 597         u32 imm_type, imm_payload;
 598         bool w_inval = false;
 599         int err;
 600
 601         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 602                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 603                         rtrs_err(sess->clt, "RDMA failed: %s\n",
 604                                   ib_wc_status_msg(wc->status));
 605                         rtrs_rdma_error_recovery(con);
 606                 }
 607                 return;
 608         }
 609         rtrs_clt_update_wc_stats(con);
 610
 611         switch (wc->opcode) {
 612         case IB_WC_RECV_RDMA_WITH_IMM:
 613                 /*
 614                  * post_recv() RDMA write completions of IO reqs (read/write)
 615                  * and hb
 616                  */
 617                 if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
 618                         return;
 619                 rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
 620                                &imm_type, &imm_payload);
 621                 if (likely(imm_type == RTRS_IO_RSP_IMM ||
 622                            imm_type == RTRS_IO_RSP_W_INV_IMM)) {
 623                         u32 msg_id;
 624
 625                         w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 626                         rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 627
 628                         process_io_rsp(sess, msg_id, err, w_inval);
 629                 } else if (imm_type == RTRS_HB_MSG_IMM) {
 630                         WARN_ON(con->c.cid);
 631                         rtrs_send_hb_ack(&sess->s);
 632                         if (sess->flags & RTRS_MSG_NEW_RKEY_F)
 633                                 return  rtrs_clt_recv_done(con, wc);
 634                 } else if (imm_type == RTRS_HB_ACK_IMM) {
 635                         WARN_ON(con->c.cid);
 636                         sess->s.hb_missed_cnt = 0;
 637                         if (sess->flags & RTRS_MSG_NEW_RKEY_F)
 638                                 return  rtrs_clt_recv_done(con, wc);
 639                 } else {
 640                         rtrs_wrn(con->c.sess, "Unknown IMM type %u\n",
 641                                   imm_type);
 642                 }
 643                 if (w_inval)
 644                         /*
 645                          * Post x2 empty WRs: first is for this RDMA with IMM,
 646                          * second is for RECV with INV, which happened earlier.
 647                          */
 648                         err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
 649                 else
 650                         err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 651                 if (unlikely(err)) {
 652                         rtrs_err(con->c.sess, "rtrs_post_recv_empty(): %d\n",
 653                                   err);
 654                         rtrs_rdma_error_recovery(con);
 655                         break;
 656                 }
 657                 break;
 658         case IB_WC_RECV:
 659                 /*
 660                  * Key invalidations from server side
 661                  */
 662                 WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
 663                           wc->wc_flags & IB_WC_WITH_IMM));
 664                 WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
 665                 if (sess->flags & RTRS_MSG_NEW_RKEY_F) {
 666                         if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
 667                                 return  rtrs_clt_recv_done(con, wc);
 668
 669                         return  rtrs_clt_rkey_rsp_done(con, wc);
 670                 }
 671                 break;
 672         case IB_WC_RDMA_WRITE:
 673                 /*
 674                  * post_send() RDMA write completions of IO reqs (read/write)
 675                  */
 676                 break;
 677
 678         default:
 679                 rtrs_wrn(sess->clt, "Unexpected WC type: %d\n", wc->opcode);
 680                 return;
 681         }
 682 }
 683
 684 static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
 685 {
 686         int err, i;
 687         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 688
 689         for (i = 0; i < q_size; i++) {
 690                 if (sess->flags & RTRS_MSG_NEW_RKEY_F) {
 691                         struct rtrs_iu *iu = &con->rsp_ius[i];
 692
 693                         err = rtrs_iu_post_recv(&con->c, iu);
 694                 } else {
 695                         err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 696                 }
 697                 if (unlikely(err))
 698                         return err;
 699         }
 700
 701         return 0;
 702 }
 703
 704 static int post_recv_sess(struct rtrs_clt_sess *sess)
 705 {
 706         size_t q_size = 0;
 707         int err, cid;
 708
 709         for (cid = 0; cid < sess->s.con_num; cid++) {
 710                 if (cid == 0)
 711                         q_size = SERVICE_CON_QUEUE_DEPTH;
 712                 else
 713                         q_size = sess->queue_depth;
 714
 715                 /*
 716                  * x2 for RDMA read responses + FR key invalidations,
 717                  * RDMA writes do not require any FR registrations.
 718                  */
 719                 q_size *= 2;
 720
 721                 err = post_recv_io(to_clt_con(sess->s.con[cid]), q_size);
 722                 if (unlikely(err)) {
 723                         rtrs_err(sess->clt, "post_recv_io(), err: %d\n", err);
 724                         return err;
 725                 }
 726         }
 727
 728         return 0;
 729 }
 730
 731 struct path_it {
 732         int i;
 733         struct list_head skip_list;
 734         struct rtrs_clt *clt;
 735         struct rtrs_clt_sess *(*next_path)(struct path_it *it);
 736 };
 737
 738 /**
 739  * list_next_or_null_rr_rcu - get next list element in round-robin fashion.
 740  * @head:       the head for the list.
 741  * @ptr:        the list head to take the next element from.
 742  * @type:       the type of the struct this is embedded in.
 743  * @memb:       the name of the list_head within the struct.
 744  *
 745  * Next element returned in round-robin fashion, i.e. head will be skipped,
 746  * but if list is observed as empty, NULL will be returned.
 747  *
 748  * This primitive may safely run concurrently with the _rcu list-mutation
 749  * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
 750  */
 751 #define list_next_or_null_rr_rcu(head, ptr, type, memb) \
 752 ({ \
 753         list_next_or_null_rcu(head, ptr, type, memb) ?: \
 754                 list_next_or_null_rcu(head, READ_ONCE((ptr)->next), \
 755                                       type, memb); \
 756 })
 757
 758 /**
 759  * get_next_path_rr() - Returns path in round-robin fashion.
 760  * @it: the path pointer
 761  *
 762  * Related to @MP_POLICY_RR
 763  *
 764  * Locks:
 765  *    rcu_read_lock() must be hold.
 766  */
 767 static struct rtrs_clt_sess *get_next_path_rr(struct path_it *it)
 768 {
 769         struct rtrs_clt_sess __rcu **ppcpu_path;
 770         struct rtrs_clt_sess *path;
 771         struct rtrs_clt *clt;
 772
 773         clt = it->clt;
 774
 775         /*
 776          * Here we use two RCU objects: @paths_list and @pcpu_path
 777          * pointer.  See rtrs_clt_remove_path_from_arr() for details
 778          * how that is handled.
 779          */
 780
 781         ppcpu_path = this_cpu_ptr(clt->pcpu_path);
 782         path = rcu_dereference(*ppcpu_path);
 783         if (unlikely(!path))
 784                 path = list_first_or_null_rcu(&clt->paths_list,
 785                                               typeof(*path), s.entry);
 786         else
 787                 path = list_next_or_null_rr_rcu(&clt->paths_list,
 788                                                 &path->s.entry,
 789                                                 typeof(*path),
 790                                                 s.entry);
 791         rcu_assign_pointer(*ppcpu_path, path);
 792
 793         return path;
 794 }
 795
 796 /**
 797  * get_next_path_min_inflight() - Returns path with minimal inflight count.
 798  * @it: the path pointer
 799  *
 800  * Related to @MP_POLICY_MIN_INFLIGHT
 801  *
 802  * Locks:
 803  *    rcu_read_lock() must be hold.
 804  */
 805 static struct rtrs_clt_sess *get_next_path_min_inflight(struct path_it *it)
 806 {
 807         struct rtrs_clt_sess *min_path = NULL;
 808         struct rtrs_clt *clt = it->clt;
 809         struct rtrs_clt_sess *sess;
 810         int min_inflight = INT_MAX;
 811         int inflight;
 812
 813         list_for_each_entry_rcu(sess, &clt->paths_list, s.entry) {
 814                 if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
 815                         continue;
 816
 817                 if (unlikely(!list_empty(raw_cpu_ptr(sess->mp_skip_entry))))
 818                         continue;
 819
 820                 inflight = atomic_read(&sess->stats->inflight);
 821
 822                 if (inflight < min_inflight) {
 823                         min_inflight = inflight;
 824                         min_path = sess;
 825                 }
 826         }
 827
 828         /*
 829          * add the path to the skip list, so that next time we can get
 830          * a different one
 831          */
 832         if (min_path)
 833                 list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
 834
 835         return min_path;
 836 }
 837
 838 static inline void path_it_init(struct path_it *it, struct rtrs_clt *clt)
 839 {
 840         INIT_LIST_HEAD(&it->skip_list);
 841         it->clt = clt;
 842         it->i = 0;
 843
 844         if (clt->mp_policy == MP_POLICY_RR)
 845                 it->next_path = get_next_path_rr;
 846         else
 847                 it->next_path = get_next_path_min_inflight;
 848 }
 849
 850 static inline void path_it_deinit(struct path_it *it)
 851 {
 852         struct list_head *skip, *tmp;
 853         /*
 854          * The skip_list is used only for the MIN_INFLIGHT policy.
 855          * We need to remove paths from it, so that next IO can insert
 856          * paths (->mp_skip_entry) into a skip_list again.
 857          */
 858         list_for_each_safe(skip, tmp, &it->skip_list)
 859                 list_del_init(skip);
 860 }
 861
 862 /**
 863  * rtrs_clt_init_req() Initialize an rtrs_clt_io_req holding information
 864  * about an inflight IO.
 865  * The user buffer holding user control message (not data) is copied into
 866  * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
 867  * also hold the control message of rtrs.
 868  * @req: an io request holding information about IO.
 869  * @sess: client session
 870  * @conf: conformation callback function to notify upper layer.
 871  * @permit: permit for allocation of RDMA remote buffer
 872  * @priv: private pointer
 873  * @vec: kernel vector containing control message
 874  * @usr_len: length of the user message
 875  * @sg: scater list for IO data
 876  * @sg_cnt: number of scater list entries
 877  * @data_len: length of the IO data
 878  * @dir: direction of the IO.
 879  */
 880 static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
 881                               struct rtrs_clt_sess *sess,
 882                               void (*conf)(void *priv, int errno),
 883                               struct rtrs_permit *permit, void *priv,
 884                               const struct kvec *vec, size_t usr_len,
 885                               struct scatterlist *sg, size_t sg_cnt,
 886                               size_t data_len, int dir)
 887 {
 888         struct iov_iter iter;
 889         size_t len;
 890
 891         req->permit = permit;
 892         req->in_use = true;
 893         req->usr_len = usr_len;
 894         req->data_len = data_len;
 895         req->sglist = sg;
 896         req->sg_cnt = sg_cnt;
 897         req->priv = priv;
 898         req->dir = dir;
 899         req->con = rtrs_permit_to_clt_con(sess, permit);
 900         req->conf = conf;
 901         req->need_inv = false;
 902         req->need_inv_comp = false;
 903         req->inv_errno = 0;
 904
 905         iov_iter_kvec(&iter, READ, vec, 1, usr_len);
 906         len = _copy_from_iter(req->iu->buf, usr_len, &iter);
 907         WARN_ON(len != usr_len);
 908
 909         reinit_completion(&req->inv_comp);
 910 }
 911
 912 static struct rtrs_clt_io_req *
 913 rtrs_clt_get_req(struct rtrs_clt_sess *sess,
 914                  void (*conf)(void *priv, int errno),
 915                  struct rtrs_permit *permit, void *priv,
 916                  const struct kvec *vec, size_t usr_len,
 917                  struct scatterlist *sg, size_t sg_cnt,
 918                  size_t data_len, int dir)
 919 {
 920         struct rtrs_clt_io_req *req;
 921
 922         req = &sess->reqs[permit->mem_id];
 923         rtrs_clt_init_req(req, sess, conf, permit, priv, vec, usr_len,
 924                            sg, sg_cnt, data_len, dir);
 925         return req;
 926 }
 927
 928 static struct rtrs_clt_io_req *
 929 rtrs_clt_get_copy_req(struct rtrs_clt_sess *alive_sess,
 930                        struct rtrs_clt_io_req *fail_req)
 931 {
 932         struct rtrs_clt_io_req *req;
 933         struct kvec vec = {
 934                 .iov_base = fail_req->iu->buf,
 935                 .iov_len  = fail_req->usr_len
 936         };
 937
 938         req = &alive_sess->reqs[fail_req->permit->mem_id];
 939         rtrs_clt_init_req(req, alive_sess, fail_req->conf, fail_req->permit,
 940                            fail_req->priv, &vec, fail_req->usr_len,
 941                            fail_req->sglist, fail_req->sg_cnt,
 942                            fail_req->data_len, fail_req->dir);
 943         return req;
 944 }
 945
 946 static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
 947                                     struct rtrs_clt_io_req *req,
 948                                     struct rtrs_rbuf *rbuf,
 949                                     u32 size, u32 imm)
 950 {
 951         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
 952         struct ib_sge *sge = req->sge;
 953         enum ib_send_flags flags;
 954         struct scatterlist *sg;
 955         size_t num_sge;
 956         int i;
 957
 958         for_each_sg(req->sglist, sg, req->sg_cnt, i) {
 959                 sge[i].addr   = sg_dma_address(sg);
 960                 sge[i].length = sg_dma_len(sg);
 961                 sge[i].lkey   = sess->s.dev->ib_pd->local_dma_lkey;
 962         }
 963         sge[i].addr   = req->iu->dma_addr;
 964         sge[i].length = size;
 965         sge[i].lkey   = sess->s.dev->ib_pd->local_dma_lkey;
 966
 967         num_sge = 1 + req->sg_cnt;
 968
 969         /*
 970          * From time to time we have to post signalled sends,
 971          * or send queue will fill up and only QP reset can help.
 972          */
 973         flags = atomic_inc_return(&con->io_cnt) % sess->queue_depth ?
 974                         0 : IB_SEND_SIGNALED;
 975
 976         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, req->iu->dma_addr,
 977                                       size, DMA_TO_DEVICE);
 978
 979         return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
 980                                             rbuf->rkey, rbuf->addr, imm,
 981                                             flags, NULL);
 982 }
 983
 984 static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
 985 {
 986         struct rtrs_clt_con *con = req->con;
 987         struct rtrs_sess *s = con->c.sess;
 988         struct rtrs_clt_sess *sess = to_clt_sess(s);
 989         struct rtrs_msg_rdma_write *msg;
 990
 991         struct rtrs_rbuf *rbuf;
 992         int ret, count = 0;
 993         u32 imm, buf_id;
 994
 995         const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
 996
 997         if (unlikely(tsize > sess->chunk_size)) {
 998                 rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
 999                           tsize, sess->chunk_size);
1000                 return -EMSGSIZE;
1001         }
1002         if (req->sg_cnt) {
1003                 count = ib_dma_map_sg(sess->s.dev->ib_dev, req->sglist,
1004                                       req->sg_cnt, req->dir);
1005                 if (unlikely(!count)) {
1006                         rtrs_wrn(s, "Write request failed, map failed\n");
1007                         return -EINVAL;
1008                 }
1009         }
1010         /* put rtrs msg after sg and user message */
1011         msg = req->iu->buf + req->usr_len;
1012         msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1013         msg->usr_len = cpu_to_le16(req->usr_len);
1014
1015         /* rtrs message on server side will be after user data and message */
1016         imm = req->permit->mem_off + req->data_len + req->usr_len;
1017         imm = rtrs_to_io_req_imm(imm);
1018         buf_id = req->permit->mem_id;
1019         req->sg_size = tsize;
1020         rbuf = &sess->rbufs[buf_id];
1021
1022         /*
1023          * Update stats now, after request is successfully sent it is not
1024          * safe anymore to touch it.
1025          */
1026         rtrs_clt_update_all_stats(req, WRITE);
1027
1028         ret = rtrs_post_rdma_write_sg(req->con, req, rbuf,
1029                                        req->usr_len + sizeof(*msg),
1030                                        imm);
1031         if (unlikely(ret)) {
1032                 rtrs_err(s, "Write request failed: %d\n", ret);
1033                 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
1034                         atomic_dec(&sess->stats->inflight);
1035                 if (req->sg_cnt)
1036                         ib_dma_unmap_sg(sess->s.dev->ib_dev, req->sglist,
1037                                         req->sg_cnt, req->dir);
1038         }
1039
1040         return ret;
1041 }
1042
1043 static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1044 {
1045         int nr;
1046
1047         /* Align the MR to a 4K page size to match the block virt boundary */
1048         nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1049         if (nr < 0)
1050                 return nr;
1051         if (unlikely(nr < req->sg_cnt))
1052                 return -EINVAL;
1053         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1054
1055         return nr;
1056 }
1057
1058 static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1059 {
1060         struct rtrs_clt_con *con = req->con;
1061         struct rtrs_sess *s = con->c.sess;
1062         struct rtrs_clt_sess *sess = to_clt_sess(s);
1063         struct rtrs_msg_rdma_read *msg;
1064         struct rtrs_ib_dev *dev;
1065
1066         struct ib_reg_wr rwr;
1067         struct ib_send_wr *wr = NULL;
1068
1069         int ret, count = 0;
1070         u32 imm, buf_id;
1071
1072         const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1073
1074         s = &sess->s;
1075         dev = sess->s.dev;
1076
1077         if (unlikely(tsize > sess->chunk_size)) {
1078                 rtrs_wrn(s,
1079                           "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1080                           tsize, sess->chunk_size);
1081                 return -EMSGSIZE;
1082         }
1083
1084         if (req->sg_cnt) {
1085                 count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1086                                       req->dir);
1087                 if (unlikely(!count)) {
1088                         rtrs_wrn(s,
1089                                   "Read request failed, dma map failed\n");
1090                         return -EINVAL;
1091                 }
1092         }
1093         /* put our message into req->buf after user message*/
1094         msg = req->iu->buf + req->usr_len;
1095         msg->type = cpu_to_le16(RTRS_MSG_READ);
1096         msg->usr_len = cpu_to_le16(req->usr_len);
1097
1098         if (count) {
1099                 ret = rtrs_map_sg_fr(req, count);
1100                 if (ret < 0) {
1101                         rtrs_err_rl(s,
1102                                      "Read request failed, failed to map  fast reg. data, err: %d\n",
1103                                      ret);
1104                         ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1105                                         req->dir);
1106                         return ret;
1107                 }
1108                 rwr = (struct ib_reg_wr) {
1109                         .wr.opcode = IB_WR_REG_MR,
1110                         .wr.wr_cqe = &fast_reg_cqe,
1111                         .mr = req->mr,
1112                         .key = req->mr->rkey,
1113                         .access = (IB_ACCESS_LOCAL_WRITE |
1114                                    IB_ACCESS_REMOTE_WRITE),
1115                 };
1116                 wr = &rwr.wr;
1117
1118                 msg->sg_cnt = cpu_to_le16(1);
1119                 msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1120
1121                 msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1122                 msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1123                 msg->desc[0].len = cpu_to_le32(req->mr->length);
1124
1125                 /* Further invalidation is required */
1126                 req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1127
1128         } else {
1129                 msg->sg_cnt = 0;
1130                 msg->flags = 0;
1131         }
1132         /*
1133          * rtrs message will be after the space reserved for disk data and
1134          * user message
1135          */
1136         imm = req->permit->mem_off + req->data_len + req->usr_len;
1137         imm = rtrs_to_io_req_imm(imm);
1138         buf_id = req->permit->mem_id;
1139
1140         req->sg_size  = sizeof(*msg);
1141         req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1142         req->sg_size += req->usr_len;
1143
1144         /*
1145          * Update stats now, after request is successfully sent it is not
1146          * safe anymore to touch it.
1147          */
1148         rtrs_clt_update_all_stats(req, READ);
1149
1150         ret = rtrs_post_send_rdma(req->con, req, &sess->rbufs[buf_id],
1151                                    req->data_len, imm, wr);
1152         if (unlikely(ret)) {
1153                 rtrs_err(s, "Read request failed: %d\n", ret);
1154                 if (sess->clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
1155                         atomic_dec(&sess->stats->inflight);
1156                 req->need_inv = false;
1157                 if (req->sg_cnt)
1158                         ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1159                                         req->sg_cnt, req->dir);
1160         }
1161
1162         return ret;
1163 }
1164
1165 /**
1166  * rtrs_clt_failover_req() Try to find an active path for a failed request
1167  * @clt: clt context
1168  * @fail_req: a failed io request.
1169  */
1170 static int rtrs_clt_failover_req(struct rtrs_clt *clt,
1171                                  struct rtrs_clt_io_req *fail_req)
1172 {
1173         struct rtrs_clt_sess *alive_sess;
1174         struct rtrs_clt_io_req *req;
1175         int err = -ECONNABORTED;
1176         struct path_it it;
1177
1178         rcu_read_lock();
1179         for (path_it_init(&it, clt);
1180              (alive_sess = it.next_path(&it)) && it.i < it.clt->paths_num;
1181              it.i++) {
1182                 if (unlikely(READ_ONCE(alive_sess->state) !=
1183                              RTRS_CLT_CONNECTED))
1184                         continue;
1185                 req = rtrs_clt_get_copy_req(alive_sess, fail_req);
1186                 if (req->dir == DMA_TO_DEVICE)
1187                         err = rtrs_clt_write_req(req);
1188                 else
1189                         err = rtrs_clt_read_req(req);
1190                 if (unlikely(err)) {
1191                         req->in_use = false;
1192                         continue;
1193                 }
1194                 /* Success path */
1195                 rtrs_clt_inc_failover_cnt(alive_sess->stats);
1196                 break;
1197         }
1198         path_it_deinit(&it);
1199         rcu_read_unlock();
1200
1201         return err;
1202 }
1203
1204 static void fail_all_outstanding_reqs(struct rtrs_clt_sess *sess)
1205 {
1206         struct rtrs_clt *clt = sess->clt;
1207         struct rtrs_clt_io_req *req;
1208         int i, err;
1209
1210         if (!sess->reqs)
1211                 return;
1212         for (i = 0; i < sess->queue_depth; ++i) {
1213                 req = &sess->reqs[i];
1214                 if (!req->in_use)
1215                         continue;
1216
1217                 /*
1218                  * Safely (without notification) complete failed request.
1219                  * After completion this request is still useble and can
1220                  * be failovered to another path.
1221                  */
1222                 complete_rdma_req(req, -ECONNABORTED, false, true);
1223
1224                 err = rtrs_clt_failover_req(clt, req);
1225                 if (unlikely(err))
1226                         /* Failover failed, notify anyway */
1227                         req->conf(req->priv, err);
1228         }
1229 }
1230
1231 static void free_sess_reqs(struct rtrs_clt_sess *sess)
1232 {
1233         struct rtrs_clt_io_req *req;
1234         int i;
1235
1236         if (!sess->reqs)
1237                 return;
1238         for (i = 0; i < sess->queue_depth; ++i) {
1239                 req = &sess->reqs[i];
1240                 if (req->mr)
1241                         ib_dereg_mr(req->mr);
1242                 kfree(req->sge);
1243                 rtrs_iu_free(req->iu, sess->s.dev->ib_dev, 1);
1244         }
1245         kfree(sess->reqs);
1246         sess->reqs = NULL;
1247 }
1248
1249 static int alloc_sess_reqs(struct rtrs_clt_sess *sess)
1250 {
1251         struct rtrs_clt_io_req *req;
1252         struct rtrs_clt *clt = sess->clt;
1253         int i, err = -ENOMEM;
1254
1255         sess->reqs = kcalloc(sess->queue_depth, sizeof(*sess->reqs),
1256                              GFP_KERNEL);
1257         if (!sess->reqs)
1258                 return -ENOMEM;
1259
1260         for (i = 0; i < sess->queue_depth; ++i) {
1261                 req = &sess->reqs[i];
1262                 req->iu = rtrs_iu_alloc(1, sess->max_hdr_size, GFP_KERNEL,
1263                                          sess->s.dev->ib_dev,
1264                                          DMA_TO_DEVICE,
1265                                          rtrs_clt_rdma_done);
1266                 if (!req->iu)
1267                         goto out;
1268
1269                 req->sge = kmalloc_array(clt->max_segments + 1,
1270                                          sizeof(*req->sge), GFP_KERNEL);
1271                 if (!req->sge)
1272                         goto out;
1273
1274                 req->mr = ib_alloc_mr(sess->s.dev->ib_pd, IB_MR_TYPE_MEM_REG,
1275                                       sess->max_pages_per_mr);
1276                 if (IS_ERR(req->mr)) {
1277                         err = PTR_ERR(req->mr);
1278                         req->mr = NULL;
1279                         pr_err("Failed to alloc sess->max_pages_per_mr %d\n",
1280                                sess->max_pages_per_mr);
1281                         goto out;
1282                 }
1283
1284                 init_completion(&req->inv_comp);
1285         }
1286
1287         return 0;
1288
1289 out:
1290         free_sess_reqs(sess);
1291
1292         return err;
1293 }
1294
1295 static int alloc_permits(struct rtrs_clt *clt)
1296 {
1297         unsigned int chunk_bits;
1298         int err, i;
1299
1300         clt->permits_map = kcalloc(BITS_TO_LONGS(clt->queue_depth),
1301                                    sizeof(long), GFP_KERNEL);
1302         if (!clt->permits_map) {
1303                 err = -ENOMEM;
1304                 goto out_err;
1305         }
1306         clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1307         if (!clt->permits) {
1308                 err = -ENOMEM;
1309                 goto err_map;
1310         }
1311         chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1312         for (i = 0; i < clt->queue_depth; i++) {
1313                 struct rtrs_permit *permit;
1314
1315                 permit = get_permit(clt, i);
1316                 permit->mem_id = i;
1317                 permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1318         }
1319
1320         return 0;
1321
1322 err_map:
1323         kfree(clt->permits_map);
1324         clt->permits_map = NULL;
1325 out_err:
1326         return err;
1327 }
1328
1329 static void free_permits(struct rtrs_clt *clt)
1330 {
1331         if (clt->permits_map) {
1332                 size_t sz = clt->queue_depth;
1333
1334                 wait_event(clt->permits_wait,
1335                            find_first_bit(clt->permits_map, sz) >= sz);
1336         }
1337         kfree(clt->permits_map);
1338         clt->permits_map = NULL;
1339         kfree(clt->permits);
1340         clt->permits = NULL;
1341 }
1342
1343 static void query_fast_reg_mode(struct rtrs_clt_sess *sess)
1344 {
1345         struct ib_device *ib_dev;
1346         u64 max_pages_per_mr;
1347         int mr_page_shift;
1348
1349         ib_dev = sess->s.dev->ib_dev;
1350
1351         /*
1352          * Use the smallest page size supported by the HCA, down to a
1353          * minimum of 4096 bytes. We're unlikely to build large sglists
1354          * out of smaller entries.
1355          */
1356         mr_page_shift      = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1357         max_pages_per_mr   = ib_dev->attrs.max_mr_size;
1358         do_div(max_pages_per_mr, (1ull << mr_page_shift));
1359         sess->max_pages_per_mr =
1360                 min3(sess->max_pages_per_mr, (u32)max_pages_per_mr,
1361                      ib_dev->attrs.max_fast_reg_page_list_len);
1362         sess->max_send_sge = ib_dev->attrs.max_send_sge;
1363 }
1364
1365 static bool rtrs_clt_change_state_get_old(struct rtrs_clt_sess *sess,
1366                                            enum rtrs_clt_state new_state,
1367                                            enum rtrs_clt_state *old_state)
1368 {
1369         bool changed;
1370
1371         spin_lock_irq(&sess->state_wq.lock);
1372         *old_state = sess->state;
1373         changed = __rtrs_clt_change_state(sess, new_state);
1374         spin_unlock_irq(&sess->state_wq.lock);
1375
1376         return changed;
1377 }
1378
1379 static bool rtrs_clt_change_state(struct rtrs_clt_sess *sess,
1380                                    enum rtrs_clt_state new_state)
1381 {
1382         enum rtrs_clt_state old_state;
1383
1384         return rtrs_clt_change_state_get_old(sess, new_state, &old_state);
1385 }
1386
1387 static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1388 {
1389         struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1390
1391         rtrs_rdma_error_recovery(con);
1392 }
1393
1394 static void rtrs_clt_init_hb(struct rtrs_clt_sess *sess)
1395 {
1396         rtrs_init_hb(&sess->s, &io_comp_cqe,
1397                       RTRS_HB_INTERVAL_MS,
1398                       RTRS_HB_MISSED_MAX,
1399                       rtrs_clt_hb_err_handler,
1400                       rtrs_wq);
1401 }
1402
1403 static void rtrs_clt_start_hb(struct rtrs_clt_sess *sess)
1404 {
1405         rtrs_start_hb(&sess->s);
1406 }
1407
1408 static void rtrs_clt_stop_hb(struct rtrs_clt_sess *sess)
1409 {
1410         rtrs_stop_hb(&sess->s);
1411 }
1412
1413 static void rtrs_clt_reconnect_work(struct work_struct *work);
1414 static void rtrs_clt_close_work(struct work_struct *work);
1415
1416 static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
1417                                          const struct rtrs_addr *path,
1418                                          size_t con_num, u16 max_segments,
1419                                          size_t max_segment_size)
1420 {
1421         struct rtrs_clt_sess *sess;
1422         int err = -ENOMEM;
1423         int cpu;
1424
1425         sess = kzalloc(sizeof(*sess), GFP_KERNEL);
1426         if (!sess)
1427                 goto err;
1428
1429         /* Extra connection for user messages */
1430         con_num += 1;
1431
1432         sess->s.con = kcalloc(con_num, sizeof(*sess->s.con), GFP_KERNEL);
1433         if (!sess->s.con)
1434                 goto err_free_sess;
1435
1436         sess->stats = kzalloc(sizeof(*sess->stats), GFP_KERNEL);
1437         if (!sess->stats)
1438                 goto err_free_con;
1439
1440         mutex_init(&sess->init_mutex);
1441         uuid_gen(&sess->s.uuid);
1442         memcpy(&sess->s.dst_addr, path->dst,
1443                rdma_addr_size((struct sockaddr *)path->dst));
1444
1445         /*
1446          * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1447          * checks the sa_family to be non-zero. If user passed src_addr=NULL
1448          * the sess->src_addr will contain only zeros, which is then fine.
1449          */
1450         if (path->src)
1451                 memcpy(&sess->s.src_addr, path->src,
1452                        rdma_addr_size((struct sockaddr *)path->src));
1453         strlcpy(sess->s.sessname, clt->sessname, sizeof(sess->s.sessname));
1454         sess->s.con_num = con_num;
1455         sess->clt = clt;
1456         sess->max_pages_per_mr = max_segments * max_segment_size >> 12;
1457         init_waitqueue_head(&sess->state_wq);
1458         sess->state = RTRS_CLT_CONNECTING;
1459         atomic_set(&sess->connected_cnt, 0);
1460         INIT_WORK(&sess->close_work, rtrs_clt_close_work);
1461         INIT_DELAYED_WORK(&sess->reconnect_dwork, rtrs_clt_reconnect_work);
1462         rtrs_clt_init_hb(sess);
1463
1464         sess->mp_skip_entry = alloc_percpu(typeof(*sess->mp_skip_entry));
1465         if (!sess->mp_skip_entry)
1466                 goto err_free_stats;
1467
1468         for_each_possible_cpu(cpu)
1469                 INIT_LIST_HEAD(per_cpu_ptr(sess->mp_skip_entry, cpu));
1470
1471         err = rtrs_clt_init_stats(sess->stats);
1472         if (err)
1473                 goto err_free_percpu;
1474
1475         return sess;
1476
1477 err_free_percpu:
1478         free_percpu(sess->mp_skip_entry);
1479 err_free_stats:
1480         kfree(sess->stats);
1481 err_free_con:
1482         kfree(sess->s.con);
1483 err_free_sess:
1484         kfree(sess);
1485 err:
1486         return ERR_PTR(err);
1487 }
1488
1489 void free_sess(struct rtrs_clt_sess *sess)
1490 {
1491         free_percpu(sess->mp_skip_entry);
1492         mutex_destroy(&sess->init_mutex);
1493         kfree(sess->s.con);
1494         kfree(sess->rbufs);
1495         kfree(sess);
1496 }
1497
1498 static int create_con(struct rtrs_clt_sess *sess, unsigned int cid)
1499 {
1500         struct rtrs_clt_con *con;
1501
1502         con = kzalloc(sizeof(*con), GFP_KERNEL);
1503         if (!con)
1504                 return -ENOMEM;
1505
1506         /* Map first two connections to the first CPU */
1507         con->cpu  = (cid ? cid - 1 : 0) % nr_cpu_ids;
1508         con->c.cid = cid;
1509         con->c.sess = &sess->s;
1510         atomic_set(&con->io_cnt, 0);
1511
1512         sess->s.con[cid] = &con->c;
1513
1514         return 0;
1515 }
1516
1517 static void destroy_con(struct rtrs_clt_con *con)
1518 {
1519         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1520
1521         sess->s.con[con->c.cid] = NULL;
1522         kfree(con);
1523 }
1524
1525 static int create_con_cq_qp(struct rtrs_clt_con *con)
1526 {
1527         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1528         u32 max_send_wr, max_recv_wr, cq_size;
1529         int err, cq_vector;
1530         struct rtrs_msg_rkey_rsp *rsp;
1531
1532         /*
1533          * This function can fail, but still destroy_con_cq_qp() should
1534          * be called, this is because create_con_cq_qp() is called on cm
1535          * event path, thus caller/waiter never knows: have we failed before
1536          * create_con_cq_qp() or after.  To solve this dilemma without
1537          * creating any additional flags just allow destroy_con_cq_qp() be
1538          * called many times.
1539          */
1540
1541         if (con->c.cid == 0) {
1542                 /*
1543                  * One completion for each receive and two for each send
1544                  * (send request + registration)
1545                  * + 2 for drain and heartbeat
1546                  * in case qp gets into error state
1547                  */
1548                 max_send_wr = SERVICE_CON_QUEUE_DEPTH * 2 + 2;
1549                 max_recv_wr = SERVICE_CON_QUEUE_DEPTH * 2 + 2;
1550                 /* We must be the first here */
1551                 if (WARN_ON(sess->s.dev))
1552                         return -EINVAL;
1553
1554                 /*
1555                  * The whole session uses device from user connection.
1556                  * Be careful not to close user connection before ib dev
1557                  * is gracefully put.
1558                  */
1559                 sess->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1560                                                        &dev_pd);
1561                 if (!sess->s.dev) {
1562                         rtrs_wrn(sess->clt,
1563                                   "rtrs_ib_dev_find_get_or_add(): no memory\n");
1564                         return -ENOMEM;
1565                 }
1566                 sess->s.dev_ref = 1;
1567                 query_fast_reg_mode(sess);
1568         } else {
1569                 /*
1570                  * Here we assume that session members are correctly set.
1571                  * This is always true if user connection (cid == 0) is
1572                  * established first.
1573                  */
1574                 if (WARN_ON(!sess->s.dev))
1575                         return -EINVAL;
1576                 if (WARN_ON(!sess->queue_depth))
1577                         return -EINVAL;
1578
1579                 /* Shared between connections */
1580                 sess->s.dev_ref++;
1581                 max_send_wr =
1582                         min_t(int, sess->s.dev->ib_dev->attrs.max_qp_wr,
1583                               /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1584                               sess->queue_depth * 3 + 1);
1585                 max_recv_wr =
1586                         min_t(int, sess->s.dev->ib_dev->attrs.max_qp_wr,
1587                               sess->queue_depth * 3 + 1);
1588         }
1589         /* alloc iu to recv new rkey reply when server reports flags set */
1590         if (sess->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1591                 con->rsp_ius = rtrs_iu_alloc(max_recv_wr, sizeof(*rsp),
1592                                               GFP_KERNEL, sess->s.dev->ib_dev,
1593                                               DMA_FROM_DEVICE,
1594                                               rtrs_clt_rdma_done);
1595                 if (!con->rsp_ius)
1596                         return -ENOMEM;
1597                 con->queue_size = max_recv_wr;
1598         }
1599         cq_size = max_send_wr + max_recv_wr;
1600         cq_vector = con->cpu % sess->s.dev->ib_dev->num_comp_vectors;
1601         err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
1602                                  cq_vector, cq_size, max_send_wr,
1603                                  max_recv_wr, IB_POLL_SOFTIRQ);
1604         /*
1605          * In case of error we do not bother to clean previous allocations,
1606          * since destroy_con_cq_qp() must be called.
1607          */
1608         return err;
1609 }
1610
1611 static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1612 {
1613         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1614
1615         /*
1616          * Be careful here: destroy_con_cq_qp() can be called even
1617          * create_con_cq_qp() failed, see comments there.
1618          */
1619
1620         rtrs_cq_qp_destroy(&con->c);
1621         if (con->rsp_ius) {
1622                 rtrs_iu_free(con->rsp_ius, sess->s.dev->ib_dev, con->queue_size);
1623                 con->rsp_ius = NULL;
1624                 con->queue_size = 0;
1625         }
1626         if (sess->s.dev_ref && !--sess->s.dev_ref) {
1627                 rtrs_ib_dev_put(sess->s.dev);
1628                 sess->s.dev = NULL;
1629         }
1630 }
1631
1632 static void stop_cm(struct rtrs_clt_con *con)
1633 {
1634         rdma_disconnect(con->c.cm_id);
1635         if (con->c.qp)
1636                 ib_drain_qp(con->c.qp);
1637 }
1638
1639 static void destroy_cm(struct rtrs_clt_con *con)
1640 {
1641         rdma_destroy_id(con->c.cm_id);
1642         con->c.cm_id = NULL;
1643 }
1644
1645 static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1646 {
1647         struct rtrs_sess *s = con->c.sess;
1648         int err;
1649
1650         err = create_con_cq_qp(con);
1651         if (err) {
1652                 rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1653                 return err;
1654         }
1655         err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1656         if (err)
1657                 rtrs_err(s, "Resolving route failed, err: %d\n", err);
1658
1659         return err;
1660 }
1661
1662 static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1663 {
1664         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1665         struct rtrs_clt *clt = sess->clt;
1666         struct rtrs_msg_conn_req msg;
1667         struct rdma_conn_param param;
1668
1669         int err;
1670
1671         param = (struct rdma_conn_param) {
1672                 .retry_count = 7,
1673                 .rnr_retry_count = 7,
1674                 .private_data = &msg,
1675                 .private_data_len = sizeof(msg),
1676         };
1677
1678         msg = (struct rtrs_msg_conn_req) {
1679                 .magic = cpu_to_le16(RTRS_MAGIC),
1680                 .version = cpu_to_le16(RTRS_PROTO_VER),
1681                 .cid = cpu_to_le16(con->c.cid),
1682                 .cid_num = cpu_to_le16(sess->s.con_num),
1683                 .recon_cnt = cpu_to_le16(sess->s.recon_cnt),
1684         };
1685         msg.first_conn = sess->for_new_clt ? FIRST_CONN : 0;
1686         uuid_copy(&msg.sess_uuid, &sess->s.uuid);
1687         uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1688
1689         err = rdma_connect_locked(con->c.cm_id, &param);
1690         if (err)
1691                 rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1692
1693         return err;
1694 }
1695
1696 static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1697                                        struct rdma_cm_event *ev)
1698 {
1699         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1700         struct rtrs_clt *clt = sess->clt;
1701         const struct rtrs_msg_conn_rsp *msg;
1702         u16 version, queue_depth;
1703         int errno;
1704         u8 len;
1705
1706         msg = ev->param.conn.private_data;
1707         len = ev->param.conn.private_data_len;
1708         if (len < sizeof(*msg)) {
1709                 rtrs_err(clt, "Invalid RTRS connection response\n");
1710                 return -ECONNRESET;
1711         }
1712         if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1713                 rtrs_err(clt, "Invalid RTRS magic\n");
1714                 return -ECONNRESET;
1715         }
1716         version = le16_to_cpu(msg->version);
1717         if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1718                 rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1719                           version >> 8, RTRS_PROTO_VER_MAJOR);
1720                 return -ECONNRESET;
1721         }
1722         errno = le16_to_cpu(msg->errno);
1723         if (errno) {
1724                 rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1725                           errno);
1726                 return -ECONNRESET;
1727         }
1728         if (con->c.cid == 0) {
1729                 queue_depth = le16_to_cpu(msg->queue_depth);
1730
1731                 if (sess->queue_depth > 0 && queue_depth != sess->queue_depth) {
1732                         rtrs_err(clt, "Error: queue depth changed\n");
1733
1734                         /*
1735                          * Stop any more reconnection attempts
1736                          */
1737                         sess->reconnect_attempts = -1;
1738                         rtrs_err(clt,
1739                                 "Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n");
1740                         return -ECONNRESET;
1741                 }
1742
1743                 if (!sess->rbufs) {
1744                         kfree(sess->rbufs);
1745                         sess->rbufs = kcalloc(queue_depth, sizeof(*sess->rbufs),
1746                                               GFP_KERNEL);
1747                         if (!sess->rbufs)
1748                                 return -ENOMEM;
1749                 }
1750                 sess->queue_depth = queue_depth;
1751                 sess->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1752                 sess->max_io_size = le32_to_cpu(msg->max_io_size);
1753                 sess->flags = le32_to_cpu(msg->flags);
1754                 sess->chunk_size = sess->max_io_size + sess->max_hdr_size;
1755
1756                 /*
1757                  * Global IO size is always a minimum.
1758                  * If while a reconnection server sends us a value a bit
1759                  * higher - client does not care and uses cached minimum.
1760                  *
1761                  * Since we can have several sessions (paths) restablishing
1762                  * connections in parallel, use lock.
1763                  */
1764                 mutex_lock(&clt->paths_mutex);
1765                 clt->queue_depth = sess->queue_depth;
1766                 clt->max_io_size = min_not_zero(sess->max_io_size,
1767                                                 clt->max_io_size);
1768                 mutex_unlock(&clt->paths_mutex);
1769
1770                 /*
1771                  * Cache the hca_port and hca_name for sysfs
1772                  */
1773                 sess->hca_port = con->c.cm_id->port_num;
1774                 scnprintf(sess->hca_name, sizeof(sess->hca_name),
1775                           sess->s.dev->ib_dev->name);
1776                 sess->s.src_addr = con->c.cm_id->route.addr.src_addr;
1777                 /* set for_new_clt, to allow future reconnect on any path */
1778                 sess->for_new_clt = 1;
1779         }
1780
1781         return 0;
1782 }
1783
1784 static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1785 {
1786         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
1787
1788         atomic_inc(&sess->connected_cnt);
1789         con->cm_err = 1;
1790 }
1791
1792 static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1793                                     struct rdma_cm_event *ev)
1794 {
1795         struct rtrs_sess *s = con->c.sess;
1796         const struct rtrs_msg_conn_rsp *msg;
1797         const char *rej_msg;
1798         int status, errno;
1799         u8 data_len;
1800
1801         status = ev->status;
1802         rej_msg = rdma_reject_msg(con->c.cm_id, status);
1803         msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1804
1805         if (msg && data_len >= sizeof(*msg)) {
1806                 errno = (int16_t)le16_to_cpu(msg->errno);
1807                 if (errno == -EBUSY)
1808                         rtrs_err(s,
1809                                   "Previous session is still exists on the server, please reconnect later\n");
1810                 else
1811                         rtrs_err(s,
1812                                   "Connect rejected: status %d (%s), rtrs errno %d\n",
1813                                   status, rej_msg, errno);
1814         } else {
1815                 rtrs_err(s,
1816                           "Connect rejected but with malformed message: status %d (%s)\n",
1817                           status, rej_msg);
1818         }
1819
1820         return -ECONNRESET;
1821 }
1822
1823 static void rtrs_clt_close_conns(struct rtrs_clt_sess *sess, bool wait)
1824 {
1825         if (rtrs_clt_change_state(sess, RTRS_CLT_CLOSING))
1826                 queue_work(rtrs_wq, &sess->close_work);
1827         if (wait)
1828                 flush_work(&sess->close_work);
1829 }
1830
1831 static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1832 {
1833         if (con->cm_err == 1) {
1834                 struct rtrs_clt_sess *sess;
1835
1836                 sess = to_clt_sess(con->c.sess);
1837                 if (atomic_dec_and_test(&sess->connected_cnt))
1838
1839                         wake_up(&sess->state_wq);
1840         }
1841         con->cm_err = cm_err;
1842 }
1843
1844 static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1845                                      struct rdma_cm_event *ev)
1846 {
1847         struct rtrs_clt_con *con = cm_id->context;
1848         struct rtrs_sess *s = con->c.sess;
1849         struct rtrs_clt_sess *sess = to_clt_sess(s);
1850         int cm_err = 0;
1851
1852         switch (ev->event) {
1853         case RDMA_CM_EVENT_ADDR_RESOLVED:
1854                 cm_err = rtrs_rdma_addr_resolved(con);
1855                 break;
1856         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1857                 cm_err = rtrs_rdma_route_resolved(con);
1858                 break;
1859         case RDMA_CM_EVENT_ESTABLISHED:
1860                 cm_err = rtrs_rdma_conn_established(con, ev);
1861                 if (likely(!cm_err)) {
1862                         /*
1863                          * Report success and wake up. Here we abuse state_wq,
1864                          * i.e. wake up without state change, but we set cm_err.
1865                          */
1866                         flag_success_on_conn(con);
1867                         wake_up(&sess->state_wq);
1868                         return 0;
1869                 }
1870                 break;
1871         case RDMA_CM_EVENT_REJECTED:
1872                 cm_err = rtrs_rdma_conn_rejected(con, ev);
1873                 break;
1874         case RDMA_CM_EVENT_CONNECT_ERROR:
1875         case RDMA_CM_EVENT_UNREACHABLE:
1876                 rtrs_wrn(s, "CM error event %d\n", ev->event);
1877                 cm_err = -ECONNRESET;
1878                 break;
1879         case RDMA_CM_EVENT_ADDR_ERROR:
1880         case RDMA_CM_EVENT_ROUTE_ERROR:
1881                 cm_err = -EHOSTUNREACH;
1882                 break;
1883         case RDMA_CM_EVENT_DISCONNECTED:
1884         case RDMA_CM_EVENT_ADDR_CHANGE:
1885         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1886                 cm_err = -ECONNRESET;
1887                 break;
1888         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1889                 /*
1890                  * Device removal is a special case.  Queue close and return 0.
1891                  */
1892                 rtrs_clt_close_conns(sess, false);
1893                 return 0;
1894         default:
1895                 rtrs_err(s, "Unexpected RDMA CM event (%d)\n", ev->event);
1896                 cm_err = -ECONNRESET;
1897                 break;
1898         }
1899
1900         if (cm_err) {
1901                 /*
1902                  * cm error makes sense only on connection establishing,
1903                  * in other cases we rely on normal procedure of reconnecting.
1904                  */
1905                 flag_error_on_conn(con, cm_err);
1906                 rtrs_rdma_error_recovery(con);
1907         }
1908
1909         return 0;
1910 }
1911
1912 static int create_cm(struct rtrs_clt_con *con)
1913 {
1914         struct rtrs_sess *s = con->c.sess;
1915         struct rtrs_clt_sess *sess = to_clt_sess(s);
1916         struct rdma_cm_id *cm_id;
1917         int err;
1918
1919         cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
1920                                sess->s.dst_addr.ss_family == AF_IB ?
1921                                RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
1922         if (IS_ERR(cm_id)) {
1923                 err = PTR_ERR(cm_id);
1924                 rtrs_err(s, "Failed to create CM ID, err: %d\n", err);
1925
1926                 return err;
1927         }
1928         con->c.cm_id = cm_id;
1929         con->cm_err = 0;
1930         /* allow the port to be reused */
1931         err = rdma_set_reuseaddr(cm_id, 1);
1932         if (err != 0) {
1933                 rtrs_err(s, "Set address reuse failed, err: %d\n", err);
1934                 goto destroy_cm;
1935         }
1936         err = rdma_resolve_addr(cm_id, (struct sockaddr *)&sess->s.src_addr,
1937                                 (struct sockaddr *)&sess->s.dst_addr,
1938                                 RTRS_CONNECT_TIMEOUT_MS);
1939         if (err) {
1940                 rtrs_err(s, "Failed to resolve address, err: %d\n", err);
1941                 goto destroy_cm;
1942         }
1943         /*
1944          * Combine connection status and session events. This is needed
1945          * for waiting two possible cases: cm_err has something meaningful
1946          * or session state was really changed to error by device removal.
1947          */
1948         err = wait_event_interruptible_timeout(
1949                         sess->state_wq,
1950                         con->cm_err || sess->state != RTRS_CLT_CONNECTING,
1951                         msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
1952         if (err == 0 || err == -ERESTARTSYS) {
1953                 if (err == 0)
1954                         err = -ETIMEDOUT;
1955                 /* Timedout or interrupted */
1956                 goto errr;
1957         }
1958         if (con->cm_err < 0) {
1959                 err = con->cm_err;
1960                 goto errr;
1961         }
1962         if (READ_ONCE(sess->state) != RTRS_CLT_CONNECTING) {
1963                 /* Device removal */
1964                 err = -ECONNABORTED;
1965                 goto errr;
1966         }
1967
1968         return 0;
1969
1970 errr:
1971         stop_cm(con);
1972         /* Is safe to call destroy if cq_qp is not inited */
1973         destroy_con_cq_qp(con);
1974 destroy_cm:
1975         destroy_cm(con);
1976
1977         return err;
1978 }
1979
1980 static void rtrs_clt_sess_up(struct rtrs_clt_sess *sess)
1981 {
1982         struct rtrs_clt *clt = sess->clt;
1983         int up;
1984
1985         /*
1986          * We can fire RECONNECTED event only when all paths were
1987          * connected on rtrs_clt_open(), then each was disconnected
1988          * and the first one connected again.  That's why this nasty
1989          * game with counter value.
1990          */
1991
1992         mutex_lock(&clt->paths_ev_mutex);
1993         up = ++clt->paths_up;
1994         /*
1995          * Here it is safe to access paths num directly since up counter
1996          * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
1997          * in progress, thus paths removals are impossible.
1998          */
1999         if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
2000                 clt->paths_up = clt->paths_num;
2001         else if (up == 1)
2002                 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
2003         mutex_unlock(&clt->paths_ev_mutex);
2004
2005         /* Mark session as established */
2006         sess->established = true;
2007         sess->reconnect_attempts = 0;
2008         sess->stats->reconnects.successful_cnt++;
2009 }
2010
2011 static void rtrs_clt_sess_down(struct rtrs_clt_sess *sess)
2012 {
2013         struct rtrs_clt *clt = sess->clt;
2014
2015         if (!sess->established)
2016                 return;
2017
2018         sess->established = false;
2019         mutex_lock(&clt->paths_ev_mutex);
2020         WARN_ON(!clt->paths_up);
2021         if (--clt->paths_up == 0)
2022                 clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2023         mutex_unlock(&clt->paths_ev_mutex);
2024 }
2025
2026 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_sess *sess)
2027 {
2028         struct rtrs_clt_con *con;
2029         unsigned int cid;
2030
2031         WARN_ON(READ_ONCE(sess->state) == RTRS_CLT_CONNECTED);
2032
2033         /*
2034          * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2035          * exactly in between.  Start destroying after it finishes.
2036          */
2037         mutex_lock(&sess->init_mutex);
2038         mutex_unlock(&sess->init_mutex);
2039
2040         /*
2041          * All IO paths must observe !CONNECTED state before we
2042          * free everything.
2043          */
2044         synchronize_rcu();
2045
2046         rtrs_clt_stop_hb(sess);
2047
2048         /*
2049          * The order it utterly crucial: firstly disconnect and complete all
2050          * rdma requests with error (thus set in_use=false for requests),
2051          * then fail outstanding requests checking in_use for each, and
2052          * eventually notify upper layer about session disconnection.
2053          */
2054
2055         for (cid = 0; cid < sess->s.con_num; cid++) {
2056                 if (!sess->s.con[cid])
2057                         break;
2058                 con = to_clt_con(sess->s.con[cid]);
2059                 stop_cm(con);
2060         }
2061         fail_all_outstanding_reqs(sess);
2062         free_sess_reqs(sess);
2063         rtrs_clt_sess_down(sess);
2064
2065         /*
2066          * Wait for graceful shutdown, namely when peer side invokes
2067          * rdma_disconnect(). 'connected_cnt' is decremented only on
2068          * CM events, thus if other side had crashed and hb has detected
2069          * something is wrong, here we will stuck for exactly timeout ms,
2070          * since CM does not fire anything.  That is fine, we are not in
2071          * hurry.
2072          */
2073         wait_event_timeout(sess->state_wq, !atomic_read(&sess->connected_cnt),
2074                            msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2075
2076         for (cid = 0; cid < sess->s.con_num; cid++) {
2077                 if (!sess->s.con[cid])
2078                         break;
2079                 con = to_clt_con(sess->s.con[cid]);
2080                 destroy_con_cq_qp(con);
2081                 destroy_cm(con);
2082                 destroy_con(con);
2083         }
2084 }
2085
2086 static inline bool xchg_sessions(struct rtrs_clt_sess __rcu **rcu_ppcpu_path,
2087                                  struct rtrs_clt_sess *sess,
2088                                  struct rtrs_clt_sess *next)
2089 {
2090         struct rtrs_clt_sess **ppcpu_path;
2091
2092         /* Call cmpxchg() without sparse warnings */
2093         ppcpu_path = (typeof(ppcpu_path))rcu_ppcpu_path;
2094         return sess == cmpxchg(ppcpu_path, sess, next);
2095 }
2096
2097 static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_sess *sess)
2098 {
2099         struct rtrs_clt *clt = sess->clt;
2100         struct rtrs_clt_sess *next;
2101         bool wait_for_grace = false;
2102         int cpu;
2103
2104         mutex_lock(&clt->paths_mutex);
2105         list_del_rcu(&sess->s.entry);
2106
2107         /* Make sure everybody observes path removal. */
2108         synchronize_rcu();
2109
2110         /*
2111          * At this point nobody sees @sess in the list, but still we have
2112          * dangling pointer @pcpu_path which _can_ point to @sess.  Since
2113          * nobody can observe @sess in the list, we guarantee that IO path
2114          * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2115          * to @sess, but can never again become @sess.
2116          */
2117
2118         /*
2119          * Decrement paths number only after grace period, because
2120          * caller of do_each_path() must firstly observe list without
2121          * path and only then decremented paths number.
2122          *
2123          * Otherwise there can be the following situation:
2124          *    o Two paths exist and IO is coming.
2125          *    o One path is removed:
2126          *      CPU#0                          CPU#1
2127          *      do_each_path():                rtrs_clt_remove_path_from_arr():
2128          *          path = get_next_path()
2129          *          ^^^                            list_del_rcu(path)
2130          *          [!CONNECTED path]              clt->paths_num--
2131          *                                              ^^^^^^^^^
2132          *          load clt->paths_num                 from 2 to 1
2133          *                    ^^^^^^^^^
2134          *                    sees 1
2135          *
2136          *      path is observed as !CONNECTED, but do_each_path() loop
2137          *      ends, because expression i < clt->paths_num is false.
2138          */
2139         clt->paths_num--;
2140
2141         /*
2142          * Get @next connection from current @sess which is going to be
2143          * removed.  If @sess is the last element, then @next is NULL.
2144          */
2145         rcu_read_lock();
2146         next = list_next_or_null_rr_rcu(&clt->paths_list, &sess->s.entry,
2147                                         typeof(*next), s.entry);
2148         rcu_read_unlock();
2149
2150         /*
2151          * @pcpu paths can still point to the path which is going to be
2152          * removed, so change the pointer manually.
2153          */
2154         for_each_possible_cpu(cpu) {
2155                 struct rtrs_clt_sess __rcu **ppcpu_path;
2156
2157                 ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2158                 if (rcu_dereference_protected(*ppcpu_path,
2159                         lockdep_is_held(&clt->paths_mutex)) != sess)
2160                         /*
2161                          * synchronize_rcu() was called just after deleting
2162                          * entry from the list, thus IO code path cannot
2163                          * change pointer back to the pointer which is going
2164                          * to be removed, we are safe here.
2165                          */
2166                         continue;
2167
2168                 /*
2169                  * We race with IO code path, which also changes pointer,
2170                  * thus we have to be careful not to overwrite it.
2171                  */
2172                 if (xchg_sessions(ppcpu_path, sess, next))
2173                         /*
2174                          * @ppcpu_path was successfully replaced with @next,
2175                          * that means that someone could also pick up the
2176                          * @sess and dereferencing it right now, so wait for
2177                          * a grace period is required.
2178                          */
2179                         wait_for_grace = true;
2180         }
2181         if (wait_for_grace)
2182                 synchronize_rcu();
2183
2184         mutex_unlock(&clt->paths_mutex);
2185 }
2186
2187 static void rtrs_clt_add_path_to_arr(struct rtrs_clt_sess *sess,
2188                                       struct rtrs_addr *addr)
2189 {
2190         struct rtrs_clt *clt = sess->clt;
2191
2192         mutex_lock(&clt->paths_mutex);
2193         clt->paths_num++;
2194
2195         list_add_tail_rcu(&sess->s.entry, &clt->paths_list);
2196         mutex_unlock(&clt->paths_mutex);
2197 }
2198
2199 static void rtrs_clt_close_work(struct work_struct *work)
2200 {
2201         struct rtrs_clt_sess *sess;
2202
2203         sess = container_of(work, struct rtrs_clt_sess, close_work);
2204
2205         cancel_delayed_work_sync(&sess->reconnect_dwork);
2206         rtrs_clt_stop_and_destroy_conns(sess);
2207         rtrs_clt_change_state(sess, RTRS_CLT_CLOSED);
2208 }
2209
2210 static int init_conns(struct rtrs_clt_sess *sess)
2211 {
2212         unsigned int cid;
2213         int err;
2214
2215         /*
2216          * On every new session connections increase reconnect counter
2217          * to avoid clashes with previous sessions not yet closed
2218          * sessions on a server side.
2219          */
2220         sess->s.recon_cnt++;
2221
2222         /* Establish all RDMA connections  */
2223         for (cid = 0; cid < sess->s.con_num; cid++) {
2224                 err = create_con(sess, cid);
2225                 if (err)
2226                         goto destroy;
2227
2228                 err = create_cm(to_clt_con(sess->s.con[cid]));
2229                 if (err) {
2230                         destroy_con(to_clt_con(sess->s.con[cid]));
2231                         goto destroy;
2232                 }
2233         }
2234         err = alloc_sess_reqs(sess);
2235         if (err)
2236                 goto destroy;
2237
2238         rtrs_clt_start_hb(sess);
2239
2240         return 0;
2241
2242 destroy:
2243         while (cid--) {
2244                 struct rtrs_clt_con *con = to_clt_con(sess->s.con[cid]);
2245
2246                 stop_cm(con);
2247                 destroy_con_cq_qp(con);
2248                 destroy_cm(con);
2249                 destroy_con(con);
2250         }
2251         /*
2252          * If we've never taken async path and got an error, say,
2253          * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2254          * manually to keep reconnecting.
2255          */
2256         rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2257
2258         return err;
2259 }
2260
2261 static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2262 {
2263         struct rtrs_clt_con *con = cq->cq_context;
2264         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
2265         struct rtrs_iu *iu;
2266
2267         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2268         rtrs_iu_free(iu, sess->s.dev->ib_dev, 1);
2269
2270         if (unlikely(wc->status != IB_WC_SUCCESS)) {
2271                 rtrs_err(sess->clt, "Sess info request send failed: %s\n",
2272                           ib_wc_status_msg(wc->status));
2273                 rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2274                 return;
2275         }
2276
2277         rtrs_clt_update_wc_stats(con);
2278 }
2279
2280 static int process_info_rsp(struct rtrs_clt_sess *sess,
2281                             const struct rtrs_msg_info_rsp *msg)
2282 {
2283         unsigned int sg_cnt, total_len;
2284         int i, sgi;
2285
2286         sg_cnt = le16_to_cpu(msg->sg_cnt);
2287         if (unlikely(!sg_cnt))
2288                 return -EINVAL;
2289         /*
2290          * Check if IB immediate data size is enough to hold the mem_id and
2291          * the offset inside the memory chunk.
2292          */
2293         if (unlikely((ilog2(sg_cnt - 1) + 1) +
2294                      (ilog2(sess->chunk_size - 1) + 1) >
2295                      MAX_IMM_PAYL_BITS)) {
2296                 rtrs_err(sess->clt,
2297                           "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2298                           MAX_IMM_PAYL_BITS, sg_cnt, sess->chunk_size);
2299                 return -EINVAL;
2300         }
2301         if (unlikely(!sg_cnt || (sess->queue_depth % sg_cnt))) {
2302                 rtrs_err(sess->clt, "Incorrect sg_cnt %d, is not multiple\n",
2303                           sg_cnt);
2304                 return -EINVAL;
2305         }
2306         total_len = 0;
2307         for (sgi = 0, i = 0; sgi < sg_cnt && i < sess->queue_depth; sgi++) {
2308                 const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2309                 u32 len, rkey;
2310                 u64 addr;
2311
2312                 addr = le64_to_cpu(desc->addr);
2313                 rkey = le32_to_cpu(desc->key);
2314                 len  = le32_to_cpu(desc->len);
2315
2316                 total_len += len;
2317
2318                 if (unlikely(!len || (len % sess->chunk_size))) {
2319                         rtrs_err(sess->clt, "Incorrect [%d].len %d\n", sgi,
2320                                   len);
2321                         return -EINVAL;
2322                 }
2323                 for ( ; len && i < sess->queue_depth; i++) {
2324                         sess->rbufs[i].addr = addr;
2325                         sess->rbufs[i].rkey = rkey;
2326
2327                         len  -= sess->chunk_size;
2328                         addr += sess->chunk_size;
2329                 }
2330         }
2331         /* Sanity check */
2332         if (unlikely(sgi != sg_cnt || i != sess->queue_depth)) {
2333                 rtrs_err(sess->clt, "Incorrect sg vector, not fully mapped\n");
2334                 return -EINVAL;
2335         }
2336         if (unlikely(total_len != sess->chunk_size * sess->queue_depth)) {
2337                 rtrs_err(sess->clt, "Incorrect total_len %d\n", total_len);
2338                 return -EINVAL;
2339         }
2340
2341         return 0;
2342 }
2343
2344 static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2345 {
2346         struct rtrs_clt_con *con = cq->cq_context;
2347         struct rtrs_clt_sess *sess = to_clt_sess(con->c.sess);
2348         struct rtrs_msg_info_rsp *msg;
2349         enum rtrs_clt_state state;
2350         struct rtrs_iu *iu;
2351         size_t rx_sz;
2352         int err;
2353
2354         state = RTRS_CLT_CONNECTING_ERR;
2355
2356         WARN_ON(con->c.cid);
2357         iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2358         if (unlikely(wc->status != IB_WC_SUCCESS)) {
2359                 rtrs_err(sess->clt, "Sess info response recv failed: %s\n",
2360                           ib_wc_status_msg(wc->status));
2361                 goto out;
2362         }
2363         WARN_ON(wc->opcode != IB_WC_RECV);
2364
2365         if (unlikely(wc->byte_len < sizeof(*msg))) {
2366                 rtrs_err(sess->clt, "Sess info response is malformed: size %d\n",
2367                           wc->byte_len);
2368                 goto out;
2369         }
2370         ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, iu->dma_addr,
2371                                    iu->size, DMA_FROM_DEVICE);
2372         msg = iu->buf;
2373         if (unlikely(le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP)) {
2374                 rtrs_err(sess->clt, "Sess info response is malformed: type %d\n",
2375                           le16_to_cpu(msg->type));
2376                 goto out;
2377         }
2378         rx_sz  = sizeof(*msg);
2379         rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2380         if (unlikely(wc->byte_len < rx_sz)) {
2381                 rtrs_err(sess->clt, "Sess info response is malformed: size %d\n",
2382                           wc->byte_len);
2383                 goto out;
2384         }
2385         err = process_info_rsp(sess, msg);
2386         if (unlikely(err))
2387                 goto out;
2388
2389         err = post_recv_sess(sess);
2390         if (unlikely(err))
2391                 goto out;
2392
2393         state = RTRS_CLT_CONNECTED;
2394
2395 out:
2396         rtrs_clt_update_wc_stats(con);
2397         rtrs_iu_free(iu, sess->s.dev->ib_dev, 1);
2398         rtrs_clt_change_state(sess, state);
2399 }
2400
2401 static int rtrs_send_sess_info(struct rtrs_clt_sess *sess)
2402 {
2403         struct rtrs_clt_con *usr_con = to_clt_con(sess->s.con[0]);
2404         struct rtrs_msg_info_req *msg;
2405         struct rtrs_iu *tx_iu, *rx_iu;
2406         size_t rx_sz;
2407         int err;
2408
2409         rx_sz  = sizeof(struct rtrs_msg_info_rsp);
2410         rx_sz += sizeof(u64) * MAX_SESS_QUEUE_DEPTH;
2411
2412         tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2413                                sess->s.dev->ib_dev, DMA_TO_DEVICE,
2414                                rtrs_clt_info_req_done);
2415         rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, sess->s.dev->ib_dev,
2416                                DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2417         if (unlikely(!tx_iu || !rx_iu)) {
2418                 err = -ENOMEM;
2419                 goto out;
2420         }
2421         /* Prepare for getting info response */
2422         err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2423         if (unlikely(err)) {
2424                 rtrs_err(sess->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2425                 goto out;
2426         }
2427         rx_iu = NULL;
2428
2429         msg = tx_iu->buf;
2430         msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2431         memcpy(msg->sessname, sess->s.sessname, sizeof(msg->sessname));
2432
2433         ib_dma_sync_single_for_device(sess->s.dev->ib_dev, tx_iu->dma_addr,
2434                                       tx_iu->size, DMA_TO_DEVICE);
2435
2436         /* Send info request */
2437         err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2438         if (unlikely(err)) {
2439                 rtrs_err(sess->clt, "rtrs_iu_post_send(), err: %d\n", err);
2440                 goto out;
2441         }
2442         tx_iu = NULL;
2443
2444         /* Wait for state change */
2445         wait_event_interruptible_timeout(sess->state_wq,
2446                                          sess->state != RTRS_CLT_CONNECTING,
2447                                          msecs_to_jiffies(
2448                                                  RTRS_CONNECT_TIMEOUT_MS));
2449         if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED)) {
2450                 if (READ_ONCE(sess->state) == RTRS_CLT_CONNECTING_ERR)
2451                         err = -ECONNRESET;
2452                 else
2453                         err = -ETIMEDOUT;
2454                 goto out;
2455         }
2456
2457 out:
2458         if (tx_iu)
2459                 rtrs_iu_free(tx_iu, sess->s.dev->ib_dev, 1);
2460         if (rx_iu)
2461                 rtrs_iu_free(rx_iu, sess->s.dev->ib_dev, 1);
2462         if (unlikely(err))
2463                 /* If we've never taken async path because of malloc problems */
2464                 rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING_ERR);
2465
2466         return err;
2467 }
2468
2469 /**
2470  * init_sess() - establishes all session connections and does handshake
2471  * @sess: client session.
2472  * In case of error full close or reconnect procedure should be taken,
2473  * because reconnect or close async works can be started.
2474  */
2475 static int init_sess(struct rtrs_clt_sess *sess)
2476 {
2477         int err;
2478
2479         mutex_lock(&sess->init_mutex);
2480         err = init_conns(sess);
2481         if (err) {
2482                 rtrs_err(sess->clt, "init_conns(), err: %d\n", err);
2483                 goto out;
2484         }
2485         err = rtrs_send_sess_info(sess);
2486         if (err) {
2487                 rtrs_err(sess->clt, "rtrs_send_sess_info(), err: %d\n", err);
2488                 goto out;
2489         }
2490         rtrs_clt_sess_up(sess);
2491 out:
2492         mutex_unlock(&sess->init_mutex);
2493
2494         return err;
2495 }
2496
2497 static void rtrs_clt_reconnect_work(struct work_struct *work)
2498 {
2499         struct rtrs_clt_sess *sess;
2500         struct rtrs_clt *clt;
2501         unsigned int delay_ms;
2502         int err;
2503
2504         sess = container_of(to_delayed_work(work), struct rtrs_clt_sess,
2505                             reconnect_dwork);
2506         clt = sess->clt;
2507
2508         if (READ_ONCE(sess->state) != RTRS_CLT_RECONNECTING)
2509                 return;
2510
2511         if (sess->reconnect_attempts >= clt->max_reconnect_attempts) {
2512                 /* Close a session completely if max attempts is reached */
2513                 rtrs_clt_close_conns(sess, false);
2514                 return;
2515         }
2516         sess->reconnect_attempts++;
2517
2518         /* Stop everything */
2519         rtrs_clt_stop_and_destroy_conns(sess);
2520         msleep(RTRS_RECONNECT_BACKOFF);
2521         if (rtrs_clt_change_state(sess, RTRS_CLT_CONNECTING)) {
2522                 err = init_sess(sess);
2523                 if (err)
2524                         goto reconnect_again;
2525         }
2526
2527         return;
2528
2529 reconnect_again:
2530         if (rtrs_clt_change_state(sess, RTRS_CLT_RECONNECTING)) {
2531                 sess->stats->reconnects.fail_cnt++;
2532                 delay_ms = clt->reconnect_delay_sec * 1000;
2533                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork,
2534                                    msecs_to_jiffies(delay_ms +
2535                                                     prandom_u32() %
2536                                                     RTRS_RECONNECT_SEED));
2537         }
2538 }
2539
2540 static void rtrs_clt_dev_release(struct device *dev)
2541 {
2542         struct rtrs_clt *clt = container_of(dev, struct rtrs_clt, dev);
2543
2544         mutex_destroy(&clt->paths_ev_mutex);
2545         mutex_destroy(&clt->paths_mutex);
2546         kfree(clt);
2547 }
2548
2549 static struct rtrs_clt *alloc_clt(const char *sessname, size_t paths_num,
2550                                   u16 port, size_t pdu_sz, void *priv,
2551                                   void  (*link_ev)(void *priv,
2552                                                    enum rtrs_clt_link_ev ev),
2553                                   unsigned int max_segments,
2554                                   size_t max_segment_size,
2555                                   unsigned int reconnect_delay_sec,
2556                                   unsigned int max_reconnect_attempts)
2557 {
2558         struct rtrs_clt *clt;
2559         int err;
2560
2561         if (!paths_num || paths_num > MAX_PATHS_NUM)
2562                 return ERR_PTR(-EINVAL);
2563
2564         if (strlen(sessname) >= sizeof(clt->sessname))
2565                 return ERR_PTR(-EINVAL);
2566
2567         clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2568         if (!clt)
2569                 return ERR_PTR(-ENOMEM);
2570
2571         clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2572         if (!clt->pcpu_path) {
2573                 kfree(clt);
2574                 return ERR_PTR(-ENOMEM);
2575         }
2576
2577         clt->dev.class = rtrs_clt_dev_class;
2578         clt->dev.release = rtrs_clt_dev_release;
2579         uuid_gen(&clt->paths_uuid);
2580         INIT_LIST_HEAD_RCU(&clt->paths_list);
2581         clt->paths_num = paths_num;
2582         clt->paths_up = MAX_PATHS_NUM;
2583         clt->port = port;
2584         clt->pdu_sz = pdu_sz;
2585         clt->max_segments = max_segments;
2586         clt->max_segment_size = max_segment_size;
2587         clt->reconnect_delay_sec = reconnect_delay_sec;
2588         clt->max_reconnect_attempts = max_reconnect_attempts;
2589         clt->priv = priv;
2590         clt->link_ev = link_ev;
2591         clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2592         strlcpy(clt->sessname, sessname, sizeof(clt->sessname));
2593         init_waitqueue_head(&clt->permits_wait);
2594         mutex_init(&clt->paths_ev_mutex);
2595         mutex_init(&clt->paths_mutex);
2596         device_initialize(&clt->dev);
2597
2598         err = dev_set_name(&clt->dev, "%s", sessname);
2599         if (err)
2600                 goto err_put;
2601
2602         /*
2603          * Suppress user space notification until
2604          * sysfs files are created
2605          */
2606         dev_set_uevent_suppress(&clt->dev, true);
2607         err = device_add(&clt->dev);
2608         if (err)
2609                 goto err_put;
2610
2611         clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2612         if (!clt->kobj_paths) {
2613                 err = -ENOMEM;
2614                 goto err_del;
2615         }
2616         err = rtrs_clt_create_sysfs_root_files(clt);
2617         if (err) {
2618                 kobject_del(clt->kobj_paths);
2619                 kobject_put(clt->kobj_paths);
2620                 goto err_del;
2621         }
2622         dev_set_uevent_suppress(&clt->dev, false);
2623         kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2624
2625         return clt;
2626 err_del:
2627         device_del(&clt->dev);
2628 err_put:
2629         free_percpu(clt->pcpu_path);
2630         put_device(&clt->dev);
2631         return ERR_PTR(err);
2632 }
2633
2634 static void free_clt(struct rtrs_clt *clt)
2635 {
2636         free_percpu(clt->pcpu_path);
2637
2638         /*
2639          * release callback will free clt and destroy mutexes in last put
2640          */
2641         device_unregister(&clt->dev);
2642 }
2643
2644 /**
2645  * rtrs_clt_open() - Open a session to an RTRS server
2646  * @ops: holds the link event callback and the private pointer.
2647  * @sessname: name of the session
2648  * @paths: Paths to be established defined by their src and dst addresses
2649  * @paths_num: Number of elements in the @paths array
2650  * @port: port to be used by the RTRS session
2651  * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2652  * @reconnect_delay_sec: time between reconnect tries
2653  * @max_segments: Max. number of segments per IO request
2654  * @max_segment_size: Max. size of one segment
2655  * @max_reconnect_attempts: Number of times to reconnect on error before giving
2656  *                          up, 0 for * disabled, -1 for forever
2657  *
2658  * Starts session establishment with the rtrs_server. The function can block
2659  * up to ~2000ms before it returns.
2660  *
2661  * Return a valid pointer on success otherwise PTR_ERR.
2662  */
2663 struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
2664                                  const char *sessname,
2665                                  const struct rtrs_addr *paths,
2666                                  size_t paths_num, u16 port,
2667                                  size_t pdu_sz, u8 reconnect_delay_sec,
2668                                  u16 max_segments,
2669                                  size_t max_segment_size,
2670                                  s16 max_reconnect_attempts)
2671 {
2672         struct rtrs_clt_sess *sess, *tmp;
2673         struct rtrs_clt *clt;
2674         int err, i;
2675
2676         clt = alloc_clt(sessname, paths_num, port, pdu_sz, ops->priv,
2677                         ops->link_ev,
2678                         max_segments, max_segment_size, reconnect_delay_sec,
2679                         max_reconnect_attempts);
2680         if (IS_ERR(clt)) {
2681                 err = PTR_ERR(clt);
2682                 goto out;
2683         }
2684         for (i = 0; i < paths_num; i++) {
2685                 struct rtrs_clt_sess *sess;
2686
2687                 sess = alloc_sess(clt, &paths[i], nr_cpu_ids,
2688                                   max_segments, max_segment_size);
2689                 if (IS_ERR(sess)) {
2690                         err = PTR_ERR(sess);
2691                         goto close_all_sess;
2692                 }
2693                 if (!i)
2694                         sess->for_new_clt = 1;
2695                 list_add_tail_rcu(&sess->s.entry, &clt->paths_list);
2696
2697                 err = init_sess(sess);
2698                 if (err) {
2699                         list_del_rcu(&sess->s.entry);
2700                         rtrs_clt_close_conns(sess, true);
2701                         free_percpu(sess->stats->pcpu_stats);
2702                         kfree(sess->stats);
2703                         free_sess(sess);
2704                         goto close_all_sess;
2705                 }
2706
2707                 err = rtrs_clt_create_sess_files(sess);
2708                 if (err) {
2709                         list_del_rcu(&sess->s.entry);
2710                         rtrs_clt_close_conns(sess, true);
2711                         free_percpu(sess->stats->pcpu_stats);
2712                         kfree(sess->stats);
2713                         free_sess(sess);
2714                         goto close_all_sess;
2715                 }
2716         }
2717         err = alloc_permits(clt);
2718         if (err)
2719                 goto close_all_sess;
2720
2721         return clt;
2722
2723 close_all_sess:
2724         list_for_each_entry_safe(sess, tmp, &clt->paths_list, s.entry) {
2725                 rtrs_clt_destroy_sess_files(sess, NULL);
2726                 rtrs_clt_close_conns(sess, true);
2727                 kobject_put(&sess->kobj);
2728         }
2729         rtrs_clt_destroy_sysfs_root_files(clt);
2730         rtrs_clt_destroy_sysfs_root_folders(clt);
2731         free_clt(clt);
2732
2733 out:
2734         return ERR_PTR(err);
2735 }
2736 EXPORT_SYMBOL(rtrs_clt_open);
2737
2738 /**
2739  * rtrs_clt_close() - Close a session
2740  * @clt: Session handle. Session is freed upon return.
2741  */
2742 void rtrs_clt_close(struct rtrs_clt *clt)
2743 {
2744         struct rtrs_clt_sess *sess, *tmp;
2745
2746         /* Firstly forbid sysfs access */
2747         rtrs_clt_destroy_sysfs_root_files(clt);
2748         rtrs_clt_destroy_sysfs_root_folders(clt);
2749
2750         /* Now it is safe to iterate over all paths without locks */
2751         list_for_each_entry_safe(sess, tmp, &clt->paths_list, s.entry) {
2752                 rtrs_clt_close_conns(sess, true);
2753                 rtrs_clt_destroy_sess_files(sess, NULL);
2754                 kobject_put(&sess->kobj);
2755         }
2756         free_permits(clt);
2757         free_clt(clt);
2758 }
2759 EXPORT_SYMBOL(rtrs_clt_close);
2760
2761 int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_sess *sess)
2762 {
2763         enum rtrs_clt_state old_state;
2764         int err = -EBUSY;
2765         bool changed;
2766
2767         changed = rtrs_clt_change_state_get_old(sess, RTRS_CLT_RECONNECTING,
2768                                                  &old_state);
2769         if (changed) {
2770                 sess->reconnect_attempts = 0;
2771                 queue_delayed_work(rtrs_wq, &sess->reconnect_dwork, 0);
2772         }
2773         if (changed || old_state == RTRS_CLT_RECONNECTING) {
2774                 /*
2775                  * flush_delayed_work() queues pending work for immediate
2776                  * execution, so do the flush if we have queued something
2777                  * right now or work is pending.
2778                  */
2779                 flush_delayed_work(&sess->reconnect_dwork);
2780                 err = (READ_ONCE(sess->state) ==
2781                        RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2782         }
2783
2784         return err;
2785 }
2786
2787 int rtrs_clt_disconnect_from_sysfs(struct rtrs_clt_sess *sess)
2788 {
2789         rtrs_clt_close_conns(sess, true);
2790
2791         return 0;
2792 }
2793
2794 int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_sess *sess,
2795                                      const struct attribute *sysfs_self)
2796 {
2797         enum rtrs_clt_state old_state;
2798         bool changed;
2799
2800         /*
2801          * Continue stopping path till state was changed to DEAD or
2802          * state was observed as DEAD:
2803          * 1. State was changed to DEAD - we were fast and nobody
2804          *    invoked rtrs_clt_reconnect(), which can again start
2805          *    reconnecting.
2806          * 2. State was observed as DEAD - we have someone in parallel
2807          *    removing the path.
2808          */
2809         do {
2810                 rtrs_clt_close_conns(sess, true);
2811                 changed = rtrs_clt_change_state_get_old(sess,
2812                                                         RTRS_CLT_DEAD,
2813                                                         &old_state);
2814         } while (!changed && old_state != RTRS_CLT_DEAD);
2815
2816         if (likely(changed)) {
2817                 rtrs_clt_remove_path_from_arr(sess);
2818                 rtrs_clt_destroy_sess_files(sess, sysfs_self);
2819                 kobject_put(&sess->kobj);
2820         }
2821
2822         return 0;
2823 }
2824
2825 void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt *clt, int value)
2826 {
2827         clt->max_reconnect_attempts = (unsigned int)value;
2828 }
2829
2830 int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt *clt)
2831 {
2832         return (int)clt->max_reconnect_attempts;
2833 }
2834
2835 /**
2836  * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2837  *
2838  * @dir:        READ/WRITE
2839  * @ops:        callback function to be called as confirmation, and the pointer.
2840  * @clt:        Session
2841  * @permit:     Preallocated permit
2842  * @vec:        Message that is sent to server together with the request.
2843  *              Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2844  *              Since the msg is copied internally it can be allocated on stack.
2845  * @nr:         Number of elements in @vec.
2846  * @data_len:   length of data sent to/from server
2847  * @sg:         Pages to be sent/received to/from server.
2848  * @sg_cnt:     Number of elements in the @sg
2849  *
2850  * Return:
2851  * 0:           Success
2852  * <0:          Error
2853  *
2854  * On dir=READ rtrs client will request a data transfer from Server to client.
2855  * The data that the server will respond with will be stored in @sg when
2856  * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2857  * On dir=WRITE rtrs client will rdma write data in sg to server side.
2858  */
2859 int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2860                      struct rtrs_clt *clt, struct rtrs_permit *permit,
2861                       const struct kvec *vec, size_t nr, size_t data_len,
2862                       struct scatterlist *sg, unsigned int sg_cnt)
2863 {
2864         struct rtrs_clt_io_req *req;
2865         struct rtrs_clt_sess *sess;
2866
2867         enum dma_data_direction dma_dir;
2868         int err = -ECONNABORTED, i;
2869         size_t usr_len, hdr_len;
2870         struct path_it it;
2871
2872         /* Get kvec length */
2873         for (i = 0, usr_len = 0; i < nr; i++)
2874                 usr_len += vec[i].iov_len;
2875
2876         if (dir == READ) {
2877                 hdr_len = sizeof(struct rtrs_msg_rdma_read) +
2878                           sg_cnt * sizeof(struct rtrs_sg_desc);
2879                 dma_dir = DMA_FROM_DEVICE;
2880         } else {
2881                 hdr_len = sizeof(struct rtrs_msg_rdma_write);
2882                 dma_dir = DMA_TO_DEVICE;
2883         }
2884
2885         rcu_read_lock();
2886         for (path_it_init(&it, clt);
2887              (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
2888                 if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
2889                         continue;
2890
2891                 if (unlikely(usr_len + hdr_len > sess->max_hdr_size)) {
2892                         rtrs_wrn_rl(sess->clt,
2893                                      "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
2894                                      dir == READ ? "Read" : "Write",
2895                                      usr_len, hdr_len, sess->max_hdr_size);
2896                         err = -EMSGSIZE;
2897                         break;
2898                 }
2899                 req = rtrs_clt_get_req(sess, ops->conf_fn, permit, ops->priv,
2900                                        vec, usr_len, sg, sg_cnt, data_len,
2901                                        dma_dir);
2902                 if (dir == READ)
2903                         err = rtrs_clt_read_req(req);
2904                 else
2905                         err = rtrs_clt_write_req(req);
2906                 if (unlikely(err)) {
2907                         req->in_use = false;
2908                         continue;
2909                 }
2910                 /* Success path */
2911                 break;
2912         }
2913         path_it_deinit(&it);
2914         rcu_read_unlock();
2915
2916         return err;
2917 }
2918 EXPORT_SYMBOL(rtrs_clt_request);
2919
2920 /**
2921  * rtrs_clt_query() - queries RTRS session attributes
2922  *@clt: session pointer
2923  *@attr: query results for session attributes.
2924  * Returns:
2925  *    0 on success
2926  *    -ECOMM            no connection to the server
2927  */
2928 int rtrs_clt_query(struct rtrs_clt *clt, struct rtrs_attrs *attr)
2929 {
2930         if (!rtrs_clt_is_connected(clt))
2931                 return -ECOMM;
2932
2933         attr->queue_depth      = clt->queue_depth;
2934         attr->max_io_size      = clt->max_io_size;
2935         attr->sess_kobj        = &clt->dev.kobj;
2936         strlcpy(attr->sessname, clt->sessname, sizeof(attr->sessname));
2937
2938         return 0;
2939 }
2940 EXPORT_SYMBOL(rtrs_clt_query);
2941
2942 int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
2943                                      struct rtrs_addr *addr)
2944 {
2945         struct rtrs_clt_sess *sess;
2946         int err;
2947
2948         sess = alloc_sess(clt, addr, nr_cpu_ids, clt->max_segments,
2949                           clt->max_segment_size);
2950         if (IS_ERR(sess))
2951                 return PTR_ERR(sess);
2952
2953         /*
2954          * It is totally safe to add path in CONNECTING state: coming
2955          * IO will never grab it.  Also it is very important to add
2956          * path before init, since init fires LINK_CONNECTED event.
2957          */
2958         rtrs_clt_add_path_to_arr(sess, addr);
2959
2960         err = init_sess(sess);
2961         if (err)
2962                 goto close_sess;
2963
2964         err = rtrs_clt_create_sess_files(sess);
2965         if (err)
2966                 goto close_sess;
2967
2968         return 0;
2969
2970 close_sess:
2971         rtrs_clt_remove_path_from_arr(sess);
2972         rtrs_clt_close_conns(sess, true);
2973         free_percpu(sess->stats->pcpu_stats);
2974         kfree(sess->stats);
2975         free_sess(sess);
2976
2977         return err;
2978 }
2979
2980 static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
2981 {
2982         if (!(dev->ib_dev->attrs.device_cap_flags &
2983               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
2984                 pr_err("Memory registrations not supported.\n");
2985                 return -ENOTSUPP;
2986         }
2987
2988         return 0;
2989 }
2990
2991 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
2992         .init = rtrs_clt_ib_dev_init
2993 };
2994
2995 static int __init rtrs_client_init(void)
2996 {
2997         rtrs_rdma_dev_pd_init(0, &dev_pd);
2998
2999         rtrs_clt_dev_class = class_create(THIS_MODULE, "rtrs-client");
3000         if (IS_ERR(rtrs_clt_dev_class)) {
3001                 pr_err("Failed to create rtrs-client dev class\n");
3002                 return PTR_ERR(rtrs_clt_dev_class);
3003         }
3004         rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
3005         if (!rtrs_wq) {
3006                 class_destroy(rtrs_clt_dev_class);
3007                 return -ENOMEM;
3008         }
3009
3010         return 0;
3011 }
3012
3013 static void __exit rtrs_client_exit(void)
3014 {
3015         destroy_workqueue(rtrs_wq);
3016         class_destroy(rtrs_clt_dev_class);
3017         rtrs_rdma_dev_pd_deinit(&dev_pd);
3018 }
3019
3020 module_init(rtrs_client_init);
3021 module_exit(rtrs_client_exit);