rtrs-clt.c

   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * RDMA Transport Layer
   4 *
   5 * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved.
   6 * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved.
   7 * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved.
   8 */
   9
  10#undef pr_fmt
  11#define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
  12
  13#include <linux/module.h>
  14#include <linux/rculist.h>
  15#include <linux/random.h>
  16
  17#include "rtrs-clt.h"
  18#include "rtrs-log.h"
  19#include "rtrs-clt-trace.h"
  20
  21#define RTRS_CONNECT_TIMEOUT_MS 30000
  22/*
  23 * Wait a bit before trying to reconnect after a failure
  24 * in order to give server time to finish clean up which
  25 * leads to "false positives" failed reconnect attempts
  26 */
  27#define RTRS_RECONNECT_BACKOFF 1000
  28/*
  29 * Wait for additional random time between 0 and 8 seconds
  30 * before starting to reconnect to avoid clients reconnecting
  31 * all at once in case of a major network outage
  32 */
  33#define RTRS_RECONNECT_SEED 8
  34
  35#define FIRST_CONN 0x01
  36/* limit to 128 * 4k = 512k max IO */
  37#define RTRS_MAX_SEGMENTS          128
  38
  39MODULE_DESCRIPTION("RDMA Transport Client");
  40MODULE_LICENSE("GPL");
  41
  42static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
  43static struct rtrs_rdma_dev_pd dev_pd = {
  44	.ops = &dev_pd_ops
  45};
  46
  47static struct workqueue_struct *rtrs_wq;
  48static const struct class rtrs_clt_dev_class = {
  49	.name = "rtrs-client",
  50};
  51
  52static inline bool rtrs_clt_is_connected(const struct rtrs_clt_sess *clt)
  53{
  54	struct rtrs_clt_path *clt_path;
  55	bool connected = false;
  56
  57	rcu_read_lock();
  58	list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry)
  59		if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED) {
  60			connected = true;
  61			break;
  62		}
  63	rcu_read_unlock();
  64
  65	return connected;
  66}
  67
  68static struct rtrs_permit *
  69__rtrs_get_permit(struct rtrs_clt_sess *clt, enum rtrs_clt_con_type con_type)
  70{
  71	size_t max_depth = clt->queue_depth;
  72	struct rtrs_permit *permit;
  73	int bit;
  74
  75	/*
  76	 * Adapted from null_blk get_tag(). Callers from different cpus may
  77	 * grab the same bit, since find_first_zero_bit is not atomic.
  78	 * But then the test_and_set_bit_lock will fail for all the
  79	 * callers but one, so that they will loop again.
  80	 * This way an explicit spinlock is not required.
  81	 */
  82	do {
  83		bit = find_first_zero_bit(clt->permits_map, max_depth);
  84		if (bit >= max_depth)
  85			return NULL;
  86	} while (test_and_set_bit_lock(bit, clt->permits_map));
  87
  88	permit = get_permit(clt, bit);
  89	WARN_ON(permit->mem_id != bit);
  90	permit->cpu_id = raw_smp_processor_id();
  91	permit->con_type = con_type;
  92
  93	return permit;
  94}
  95
  96static inline void __rtrs_put_permit(struct rtrs_clt_sess *clt,
  97				      struct rtrs_permit *permit)
  98{
  99	clear_bit_unlock(permit->mem_id, clt->permits_map);
 100}
 101
 102/**
 103 * rtrs_clt_get_permit() - allocates permit for future RDMA operation
 104 * @clt:	Current session
 105 * @con_type:	Type of connection to use with the permit
 106 * @can_wait:	Wait type
 107 *
 108 * Description:
 109 *    Allocates permit for the following RDMA operation.  Permit is used
 110 *    to preallocate all resources and to propagate memory pressure
 111 *    up earlier.
 112 *
 113 * Context:
 114 *    Can sleep if @wait == RTRS_PERMIT_WAIT
 115 */
 116struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt_sess *clt,
 117					  enum rtrs_clt_con_type con_type,
 118					  enum wait_type can_wait)
 119{
 120	struct rtrs_permit *permit;
 121	DEFINE_WAIT(wait);
 122
 123	permit = __rtrs_get_permit(clt, con_type);
 124	if (permit || !can_wait)
 125		return permit;
 126
 127	do {
 128		prepare_to_wait(&clt->permits_wait, &wait,
 129				TASK_UNINTERRUPTIBLE);
 130		permit = __rtrs_get_permit(clt, con_type);
 131		if (permit)
 132			break;
 133
 134		io_schedule();
 135	} while (1);
 136
 137	finish_wait(&clt->permits_wait, &wait);
 138
 139	return permit;
 140}
 141EXPORT_SYMBOL(rtrs_clt_get_permit);
 142
 143/**
 144 * rtrs_clt_put_permit() - puts allocated permit
 145 * @clt:	Current session
 146 * @permit:	Permit to be freed
 147 *
 148 * Context:
 149 *    Does not matter
 150 */
 151void rtrs_clt_put_permit(struct rtrs_clt_sess *clt,
 152			 struct rtrs_permit *permit)
 153{
 154	if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
 155		return;
 156
 157	__rtrs_put_permit(clt, permit);
 158
 159	/*
 160	 * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
 161	 * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
 162	 * it must have added itself to &clt->permits_wait before
 163	 * __rtrs_put_permit() finished.
 164	 * Hence it is safe to guard wake_up() with a waitqueue_active() test.
 165	 */
 166	if (waitqueue_active(&clt->permits_wait))
 167		wake_up(&clt->permits_wait);
 168}
 169EXPORT_SYMBOL(rtrs_clt_put_permit);
 170
 171/**
 172 * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
 173 * @clt_path: client path pointer
 174 * @permit: permit for the allocation of the RDMA buffer
 175 * Note:
 176 *     IO connection starts from 1.
 177 *     0 connection is for user messages.
 178 */
 179static
 180struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_path *clt_path,
 181					    struct rtrs_permit *permit)
 182{
 183	int id = 0;
 184
 185	if (permit->con_type == RTRS_IO_CON)
 186		id = (permit->cpu_id % (clt_path->s.irq_con_num - 1)) + 1;
 187
 188	return to_clt_con(clt_path->s.con[id]);
 189}
 190
 191/**
 192 * rtrs_clt_change_state() - change the session state through session state
 193 * machine.
 194 *
 195 * @clt_path: client path to change the state of.
 196 * @new_state: state to change to.
 197 *
 198 * returns true if sess's state is changed to new state, otherwise return false.
 199 *
 200 * Locks:
 201 * state_wq lock must be hold.
 202 */
 203static bool rtrs_clt_change_state(struct rtrs_clt_path *clt_path,
 204				     enum rtrs_clt_state new_state)
 205{
 206	enum rtrs_clt_state old_state;
 207	bool changed = false;
 208
 209	lockdep_assert_held(&clt_path->state_wq.lock);
 210
 211	old_state = clt_path->state;
 212	switch (new_state) {
 213	case RTRS_CLT_CONNECTING:
 214		switch (old_state) {
 215		case RTRS_CLT_RECONNECTING:
 216			changed = true;
 217			fallthrough;
 218		default:
 219			break;
 220		}
 221		break;
 222	case RTRS_CLT_RECONNECTING:
 223		switch (old_state) {
 224		case RTRS_CLT_CONNECTED:
 225		case RTRS_CLT_CONNECTING_ERR:
 226		case RTRS_CLT_CLOSED:
 227			changed = true;
 228			fallthrough;
 229		default:
 230			break;
 231		}
 232		break;
 233	case RTRS_CLT_CONNECTED:
 234		switch (old_state) {
 235		case RTRS_CLT_CONNECTING:
 236			changed = true;
 237			fallthrough;
 238		default:
 239			break;
 240		}
 241		break;
 242	case RTRS_CLT_CONNECTING_ERR:
 243		switch (old_state) {
 244		case RTRS_CLT_CONNECTING:
 245			changed = true;
 246			fallthrough;
 247		default:
 248			break;
 249		}
 250		break;
 251	case RTRS_CLT_CLOSING:
 252		switch (old_state) {
 253		case RTRS_CLT_CONNECTING:
 254		case RTRS_CLT_CONNECTING_ERR:
 255		case RTRS_CLT_RECONNECTING:
 256		case RTRS_CLT_CONNECTED:
 257			changed = true;
 258			fallthrough;
 259		default:
 260			break;
 261		}
 262		break;
 263	case RTRS_CLT_CLOSED:
 264		switch (old_state) {
 265		case RTRS_CLT_CLOSING:
 266			changed = true;
 267			fallthrough;
 268		default:
 269			break;
 270		}
 271		break;
 272	case RTRS_CLT_DEAD:
 273		switch (old_state) {
 274		case RTRS_CLT_CLOSED:
 275			changed = true;
 276			fallthrough;
 277		default:
 278			break;
 279		}
 280		break;
 281	default:
 282		break;
 283	}
 284	if (changed) {
 285		clt_path->state = new_state;
 286		wake_up_locked(&clt_path->state_wq);
 287	}
 288
 289	return changed;
 290}
 291
 292static bool rtrs_clt_change_state_from_to(struct rtrs_clt_path *clt_path,
 293					   enum rtrs_clt_state old_state,
 294					   enum rtrs_clt_state new_state)
 295{
 296	bool changed = false;
 297
 298	spin_lock_irq(&clt_path->state_wq.lock);
 299	if (clt_path->state == old_state)
 300		changed = rtrs_clt_change_state(clt_path, new_state);
 301	spin_unlock_irq(&clt_path->state_wq.lock);
 302
 303	return changed;
 304}
 305
 306static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path);
 307static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
 308{
 309	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 310
 311	trace_rtrs_rdma_error_recovery(clt_path);
 312
 313	if (rtrs_clt_change_state_from_to(clt_path,
 314					   RTRS_CLT_CONNECTED,
 315					   RTRS_CLT_RECONNECTING)) {
 316		queue_work(rtrs_wq, &clt_path->err_recovery_work);
 317	} else {
 318		/*
 319		 * Error can happen just on establishing new connection,
 320		 * so notify waiter with error state, waiter is responsible
 321		 * for cleaning the rest and reconnect if needed.
 322		 */
 323		rtrs_clt_change_state_from_to(clt_path,
 324					       RTRS_CLT_CONNECTING,
 325					       RTRS_CLT_CONNECTING_ERR);
 326	}
 327}
 328
 329static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
 330{
 331	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 332
 333	if (wc->status != IB_WC_SUCCESS) {
 334		rtrs_err_rl(con->c.path, "Failed IB_WR_REG_MR: %s\n",
 335			  ib_wc_status_msg(wc->status));
 336		rtrs_rdma_error_recovery(con);
 337	}
 338}
 339
 340static struct ib_cqe fast_reg_cqe = {
 341	.done = rtrs_clt_fast_reg_done
 342};
 343
 344static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 345			      bool notify, bool can_wait);
 346
 347static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 348{
 349	struct rtrs_clt_io_req *req =
 350		container_of(wc->wr_cqe, typeof(*req), inv_cqe);
 351	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 352
 353	if (wc->status != IB_WC_SUCCESS) {
 354		rtrs_err_rl(con->c.path, "Failed IB_WR_LOCAL_INV: %s\n",
 355			  ib_wc_status_msg(wc->status));
 356		rtrs_rdma_error_recovery(con);
 357	}
 358	req->mr->need_inval = false;
 359	if (req->need_inv_comp)
 360		complete(&req->inv_comp);
 361	else
 362		/* Complete request from INV callback */
 363		complete_rdma_req(req, req->inv_errno, true, false);
 364}
 365
 366static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
 367{
 368	struct rtrs_clt_con *con = req->con;
 369	struct ib_send_wr wr = {
 370		.opcode		    = IB_WR_LOCAL_INV,
 371		.wr_cqe		    = &req->inv_cqe,
 372		.send_flags	    = IB_SEND_SIGNALED,
 373		.ex.invalidate_rkey = req->mr->rkey,
 374	};
 375	req->inv_cqe.done = rtrs_clt_inv_rkey_done;
 376
 377	return ib_post_send(con->c.qp, &wr, NULL);
 378}
 379
 380static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
 381			      bool notify, bool can_wait)
 382{
 383	struct rtrs_clt_con *con = req->con;
 384	struct rtrs_clt_path *clt_path;
 385	int err;
 386
 387	if (!req->in_use)
 388		return;
 389	if (WARN_ON(!req->con))
 390		return;
 391	clt_path = to_clt_path(con->c.path);
 392
 393	if (req->sg_cnt) {
 394		if (req->mr->need_inval) {
 395			/*
 396			 * We are here to invalidate read/write requests
 397			 * ourselves.  In normal scenario server should
 398			 * send INV for all read requests, we do local
 399			 * invalidate for write requests ourselves, but
 400			 * we are here, thus three things could happen:
 401			 *
 402			 *    1.  this is failover, when errno != 0
 403			 *        and can_wait == 1,
 404			 *
 405			 *    2.  something totally bad happened and
 406			 *        server forgot to send INV, so we
 407			 *        should do that ourselves.
 408			 *
 409			 *    3.  write request finishes, we need to do local
 410			 *        invalidate
 411			 */
 412
 413			if (can_wait) {
 414				req->need_inv_comp = true;
 415			} else {
 416				/* This should be IO path, so always notify */
 417				WARN_ON(!notify);
 418				/* Save errno for INV callback */
 419				req->inv_errno = errno;
 420			}
 421
 422			refcount_inc(&req->ref);
 423			err = rtrs_inv_rkey(req);
 424			if (err) {
 425				rtrs_err_rl(con->c.path, "Send INV WR key=%#x: %d\n",
 426					  req->mr->rkey, err);
 427			} else if (can_wait) {
 428				wait_for_completion(&req->inv_comp);
 429			}
 430			if (!refcount_dec_and_test(&req->ref))
 431				return;
 432		}
 433		ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
 434				req->sg_cnt, req->dir);
 435	}
 436	if (!refcount_dec_and_test(&req->ref))
 437		return;
 438	if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
 439		atomic_dec(&clt_path->stats->inflight);
 440
 441	req->in_use = false;
 442	req->con = NULL;
 443
 444	if (errno) {
 445		rtrs_err_rl(con->c.path,
 446			    "IO %s request failed: error=%d path=%s [%s:%u] notify=%d\n",
 447			    req->dir == DMA_TO_DEVICE ? "write" : "read", errno,
 448			    kobject_name(&clt_path->kobj), clt_path->hca_name,
 449			    clt_path->hca_port, notify);
 450	}
 451
 452	if (notify)
 453		req->conf(req->priv, errno);
 454}
 455
 456static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
 457				struct rtrs_clt_io_req *req,
 458				struct rtrs_rbuf *rbuf, u32 off,
 459				u32 imm, struct ib_send_wr *wr)
 460{
 461	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 462	enum ib_send_flags flags;
 463	struct ib_sge sge;
 464
 465	if (!req->sg_size) {
 466		rtrs_wrn(con->c.path,
 467			 "Doing RDMA Write failed, no data supplied\n");
 468		return -EINVAL;
 469	}
 470
 471	/* user data and user message in the first list element */
 472	sge.addr   = req->iu->dma_addr;
 473	sge.length = req->sg_size;
 474	sge.lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
 475
 476	/*
 477	 * From time to time we have to post signalled sends,
 478	 * or send queue will fill up and only QP reset can help.
 479	 */
 480	flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
 481			0 : IB_SEND_SIGNALED;
 482
 483	ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
 484				      req->iu->dma_addr,
 485				      req->sg_size, DMA_TO_DEVICE);
 486
 487	return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
 488					    rbuf->rkey, rbuf->addr + off,
 489					    imm, flags, wr, NULL);
 490}
 491
 492static void process_io_rsp(struct rtrs_clt_path *clt_path, u32 msg_id,
 493			   s16 errno, bool w_inval)
 494{
 495	struct rtrs_clt_io_req *req;
 496
 497	if (WARN_ON(msg_id >= clt_path->queue_depth))
 498		return;
 499
 500	req = &clt_path->reqs[msg_id];
 501	/* Drop need_inv if server responded with send with invalidation */
 502	req->mr->need_inval &= !w_inval;
 503	complete_rdma_req(req, errno, true, false);
 504}
 505
 506static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
 507{
 508	struct rtrs_iu *iu;
 509	int err;
 510	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 511
 512	WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
 513	iu = container_of(wc->wr_cqe, struct rtrs_iu,
 514			  cqe);
 515	err = rtrs_iu_post_recv(&con->c, iu);
 516	if (err) {
 517		rtrs_err(con->c.path, "post iu failed %d\n", err);
 518		rtrs_rdma_error_recovery(con);
 519	}
 520}
 521
 522static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
 523{
 524	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 525	struct rtrs_msg_rkey_rsp *msg;
 526	u32 imm_type, imm_payload;
 527	bool w_inval = false;
 528	struct rtrs_iu *iu;
 529	u32 buf_id;
 530	int err;
 531
 532	WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
 533
 534	iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
 535
 536	if (wc->byte_len < sizeof(*msg)) {
 537		rtrs_err(con->c.path, "rkey response is malformed: size %d\n",
 538			  wc->byte_len);
 539		goto out;
 540	}
 541	ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
 542				   iu->size, DMA_FROM_DEVICE);
 543	msg = iu->buf;
 544	if (le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP) {
 545		rtrs_err(clt_path->clt,
 546			  "rkey response is malformed: type %d\n",
 547			  le16_to_cpu(msg->type));
 548		goto out;
 549	}
 550	buf_id = le16_to_cpu(msg->buf_id);
 551	if (WARN_ON(buf_id >= clt_path->queue_depth))
 552		goto out;
 553
 554	rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
 555	if (imm_type == RTRS_IO_RSP_IMM ||
 556	    imm_type == RTRS_IO_RSP_W_INV_IMM) {
 557		u32 msg_id;
 558
 559		w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 560		rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 561
 562		if (WARN_ON(buf_id != msg_id))
 563			goto out;
 564		clt_path->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
 565		process_io_rsp(clt_path, msg_id, err, w_inval);
 566	}
 567	ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, iu->dma_addr,
 568				      iu->size, DMA_FROM_DEVICE);
 569	return rtrs_clt_recv_done(con, wc);
 570out:
 571	rtrs_rdma_error_recovery(con);
 572}
 573
 574static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
 575
 576static struct ib_cqe io_comp_cqe = {
 577	.done = rtrs_clt_rdma_done
 578};
 579
 580/*
 581 * Post x2 empty WRs: first is for this RDMA with IMM,
 582 * second is for RECV with INV, which happened earlier.
 583 */
 584static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
 585{
 586	struct ib_recv_wr wr_arr[2], *wr;
 587	int i;
 588
 589	memset(wr_arr, 0, sizeof(wr_arr));
 590	for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
 591		wr = &wr_arr[i];
 592		wr->wr_cqe  = cqe;
 593		if (i)
 594			/* Chain backwards */
 595			wr->next = &wr_arr[i - 1];
 596	}
 597
 598	return ib_post_recv(con->qp, wr, NULL);
 599}
 600
 601static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
 602{
 603	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
 604	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 605	u32 imm_type, imm_payload;
 606	bool w_inval = false;
 607	int err;
 608
 609	if (wc->status != IB_WC_SUCCESS) {
 610		if (wc->status != IB_WC_WR_FLUSH_ERR) {
 611			rtrs_err(clt_path->clt, "RDMA failed: %s\n",
 612				  ib_wc_status_msg(wc->status));
 613			rtrs_rdma_error_recovery(con);
 614		}
 615		return;
 616	}
 617	rtrs_clt_update_wc_stats(con);
 618
 619	switch (wc->opcode) {
 620	case IB_WC_RECV_RDMA_WITH_IMM:
 621		/*
 622		 * post_recv() RDMA write completions of IO reqs (read/write)
 623		 * and hb
 624		 */
 625		if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
 626			return;
 627		clt_path->s.hb_missed_cnt = 0;
 628		rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
 629			       &imm_type, &imm_payload);
 630		if (imm_type == RTRS_IO_RSP_IMM ||
 631		    imm_type == RTRS_IO_RSP_W_INV_IMM) {
 632			u32 msg_id;
 633
 634			w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
 635			rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
 636
 637			process_io_rsp(clt_path, msg_id, err, w_inval);
 638		} else if (imm_type == RTRS_HB_MSG_IMM) {
 639			WARN_ON(con->c.cid);
 640			rtrs_send_hb_ack(&clt_path->s);
 641			if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
 642				return  rtrs_clt_recv_done(con, wc);
 643		} else if (imm_type == RTRS_HB_ACK_IMM) {
 644			WARN_ON(con->c.cid);
 645			clt_path->s.hb_cur_latency =
 646				ktime_sub(ktime_get(), clt_path->s.hb_last_sent);
 647			if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
 648				return  rtrs_clt_recv_done(con, wc);
 649		} else {
 650			rtrs_wrn(con->c.path, "Unknown IMM type %u\n",
 651				  imm_type);
 652		}
 653		if (w_inval)
 654			/*
 655			 * Post x2 empty WRs: first is for this RDMA with IMM,
 656			 * second is for RECV with INV, which happened earlier.
 657			 */
 658			err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
 659		else
 660			err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 661		if (err) {
 662			rtrs_err(con->c.path, "rtrs_post_recv_empty(): %d\n",
 663				  err);
 664			rtrs_rdma_error_recovery(con);
 665		}
 666		break;
 667	case IB_WC_RECV:
 668		/*
 669		 * Key invalidations from server side
 670		 */
 671		clt_path->s.hb_missed_cnt = 0;
 672		WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
 673			  wc->wc_flags & IB_WC_WITH_IMM));
 674		WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
 675		if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
 676			if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
 677				return  rtrs_clt_recv_done(con, wc);
 678
 679			return  rtrs_clt_rkey_rsp_done(con, wc);
 680		}
 681		break;
 682	case IB_WC_RDMA_WRITE:
 683		/*
 684		 * post_send() RDMA write completions of IO reqs (read/write)
 685		 * and hb.
 686		 */
 687		break;
 688
 689	default:
 690		rtrs_wrn(clt_path->clt, "Unexpected WC type: %d\n", wc->opcode);
 691		return;
 692	}
 693}
 694
 695static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
 696{
 697	int err, i;
 698	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
 699
 700	for (i = 0; i < q_size; i++) {
 701		if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
 702			struct rtrs_iu *iu = &con->rsp_ius[i];
 703
 704			err = rtrs_iu_post_recv(&con->c, iu);
 705		} else {
 706			err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
 707		}
 708		if (err)
 709			return err;
 710	}
 711
 712	return 0;
 713}
 714
 715static int post_recv_path(struct rtrs_clt_path *clt_path)
 716{
 717	size_t q_size = 0;
 718	int err, cid;
 719
 720	for (cid = 0; cid < clt_path->s.con_num; cid++) {
 721		if (cid == 0)
 722			q_size = SERVICE_CON_QUEUE_DEPTH;
 723		else
 724			q_size = clt_path->queue_depth;
 725
 726		/*
 727		 * x2 for RDMA read responses + FR key invalidations,
 728		 * RDMA writes do not require any FR registrations.
 729		 */
 730		q_size *= 2;
 731
 732		err = post_recv_io(to_clt_con(clt_path->s.con[cid]), q_size);
 733		if (err) {
 734			rtrs_err(clt_path->clt, "post_recv_io(), err: %d\n",
 735				 err);
 736			return err;
 737		}
 738	}
 739
 740	return 0;
 741}
 742
 743struct path_it {
 744	int i;
 745	struct list_head skip_list;
 746	struct rtrs_clt_sess *clt;
 747	struct rtrs_clt_path *(*next_path)(struct path_it *it);
 748};
 749
 750/*
 751 * rtrs_clt_get_next_path_or_null - get clt path from the list or return NULL
 752 * @head:	the head for the list.
 753 * @clt_path:	The element to take the next clt_path from.
 754 *
 755 * Next clt path returned in round-robin fashion, i.e. head will be skipped,
 756 * but if list is observed as empty, NULL will be returned.
 757 *
 758 * This function may safely run concurrently with the _rcu list-mutation
 759 * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
 760 */
 761static inline struct rtrs_clt_path *
 762rtrs_clt_get_next_path_or_null(struct list_head *head, struct rtrs_clt_path *clt_path)
 763{
 764	return list_next_or_null_rcu(head, &clt_path->s.entry, typeof(*clt_path), s.entry) ?:
 765				     list_next_or_null_rcu(head,
 766							   READ_ONCE((&clt_path->s.entry)->next),
 767							   typeof(*clt_path), s.entry);
 768}
 769
 770/**
 771 * get_next_path_rr() - Returns path in round-robin fashion.
 772 * @it:	the path pointer
 773 *
 774 * Related to @MP_POLICY_RR
 775 *
 776 * Locks:
 777 *    rcu_read_lock() must be held.
 778 */
 779static struct rtrs_clt_path *get_next_path_rr(struct path_it *it)
 780{
 781	struct rtrs_clt_path __rcu **ppcpu_path;
 782	struct rtrs_clt_path *path;
 783	struct rtrs_clt_sess *clt;
 784
 785	/*
 786	 * Assert that rcu lock must be held
 787	 */
 788	RCU_LOCKDEP_WARN(!rcu_read_lock_held(), "no rcu read lock held");
 789
 790	clt = it->clt;
 791
 792	/*
 793	 * Here we use two RCU objects: @paths_list and @pcpu_path
 794	 * pointer.  See rtrs_clt_remove_path_from_arr() for details
 795	 * how that is handled.
 796	 */
 797
 798	ppcpu_path = this_cpu_ptr(clt->pcpu_path);
 799	path = rcu_dereference(*ppcpu_path);
 800	if (!path)
 801		path = list_first_or_null_rcu(&clt->paths_list,
 802					      typeof(*path), s.entry);
 803	else
 804		path = rtrs_clt_get_next_path_or_null(&clt->paths_list, path);
 805
 806	rcu_assign_pointer(*ppcpu_path, path);
 807
 808	return path;
 809}
 810
 811/**
 812 * get_next_path_min_inflight() - Returns path with minimal inflight count.
 813 * @it:	the path pointer
 814 *
 815 * Related to @MP_POLICY_MIN_INFLIGHT
 816 *
 817 * Locks:
 818 *    rcu_read_lock() must be hold.
 819 */
 820static struct rtrs_clt_path *get_next_path_min_inflight(struct path_it *it)
 821{
 822	struct rtrs_clt_path *min_path = NULL;
 823	struct rtrs_clt_sess *clt = it->clt;
 824	struct rtrs_clt_path *clt_path;
 825	int min_inflight = INT_MAX;
 826	int inflight;
 827
 828	list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
 829		if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
 830			continue;
 831
 832		if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
 833			continue;
 834
 835		inflight = atomic_read(&clt_path->stats->inflight);
 836
 837		if (inflight < min_inflight) {
 838			min_inflight = inflight;
 839			min_path = clt_path;
 840		}
 841	}
 842
 843	/*
 844	 * add the path to the skip list, so that next time we can get
 845	 * a different one
 846	 */
 847	if (min_path)
 848		list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
 849
 850	return min_path;
 851}
 852
 853/**
 854 * get_next_path_min_latency() - Returns path with minimal latency.
 855 * @it:	the path pointer
 856 *
 857 * Return: a path with the lowest latency or NULL if all paths are tried
 858 *
 859 * Locks:
 860 *    rcu_read_lock() must be hold.
 861 *
 862 * Related to @MP_POLICY_MIN_LATENCY
 863 *
 864 * This DOES skip an already-tried path.
 865 * There is a skip-list to skip a path if the path has tried but failed.
 866 * It will try the minimum latency path and then the second minimum latency
 867 * path and so on. Finally it will return NULL if all paths are tried.
 868 * Therefore the caller MUST check the returned
 869 * path is NULL and trigger the IO error.
 870 */
 871static struct rtrs_clt_path *get_next_path_min_latency(struct path_it *it)
 872{
 873	struct rtrs_clt_path *min_path = NULL;
 874	struct rtrs_clt_sess *clt = it->clt;
 875	struct rtrs_clt_path *clt_path;
 876	ktime_t min_latency = KTIME_MAX;
 877	ktime_t latency;
 878
 879	list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
 880		if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
 881			continue;
 882
 883		if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
 884			continue;
 885
 886		latency = clt_path->s.hb_cur_latency;
 887
 888		if (latency < min_latency) {
 889			min_latency = latency;
 890			min_path = clt_path;
 891		}
 892	}
 893
 894	/*
 895	 * add the path to the skip list, so that next time we can get
 896	 * a different one
 897	 */
 898	if (min_path)
 899		list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
 900
 901	return min_path;
 902}
 903
 904static inline void path_it_init(struct path_it *it, struct rtrs_clt_sess *clt)
 905{
 906	INIT_LIST_HEAD(&it->skip_list);
 907	it->clt = clt;
 908	it->i = 0;
 909
 910	if (clt->mp_policy == MP_POLICY_RR)
 911		it->next_path = get_next_path_rr;
 912	else if (clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
 913		it->next_path = get_next_path_min_inflight;
 914	else
 915		it->next_path = get_next_path_min_latency;
 916}
 917
 918static inline void path_it_deinit(struct path_it *it)
 919{
 920	struct list_head *skip, *tmp;
 921	/*
 922	 * The skip_list is used only for the MIN_INFLIGHT and MIN_LATENCY policies.
 923	 * We need to remove paths from it, so that next IO can insert
 924	 * paths (->mp_skip_entry) into a skip_list again.
 925	 */
 926	list_for_each_safe(skip, tmp, &it->skip_list)
 927		list_del_init(skip);
 928}
 929
 930/**
 931 * rtrs_clt_init_req() - Initialize an rtrs_clt_io_req holding information
 932 * about an inflight IO.
 933 * The user buffer holding user control message (not data) is copied into
 934 * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
 935 * also hold the control message of rtrs.
 936 * @req: an io request holding information about IO.
 937 * @clt_path: client path
 938 * @conf: conformation callback function to notify upper layer.
 939 * @permit: permit for allocation of RDMA remote buffer
 940 * @priv: private pointer
 941 * @vec: kernel vector containing control message
 942 * @usr_len: length of the user message
 943 * @sg: scater list for IO data
 944 * @sg_cnt: number of scater list entries
 945 * @data_len: length of the IO data
 946 * @dir: direction of the IO.
 947 */
 948static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
 949			      struct rtrs_clt_path *clt_path,
 950			      void (*conf)(void *priv, int errno),
 951			      struct rtrs_permit *permit, void *priv,
 952			      const struct kvec *vec, size_t usr_len,
 953			      struct scatterlist *sg, size_t sg_cnt,
 954			      size_t data_len, int dir)
 955{
 956	struct iov_iter iter;
 957	size_t len;
 958
 959	req->permit = permit;
 960	req->in_use = true;
 961	req->usr_len = usr_len;
 962	req->data_len = data_len;
 963	req->sglist = sg;
 964	req->sg_cnt = sg_cnt;
 965	req->priv = priv;
 966	req->dir = dir;
 967	req->con = rtrs_permit_to_clt_con(clt_path, permit);
 968	req->conf = conf;
 969	req->mr->need_inval = false;
 970	req->need_inv_comp = false;
 971	req->inv_errno = 0;
 972	refcount_set(&req->ref, 1);
 973	req->mp_policy = clt_path->clt->mp_policy;
 974
 975	iov_iter_kvec(&iter, ITER_SOURCE, vec, 1, usr_len);
 976	len = _copy_from_iter(req->iu->buf, usr_len, &iter);
 977	WARN_ON(len != usr_len);
 978
 979	reinit_completion(&req->inv_comp);
 980}
 981
 982static struct rtrs_clt_io_req *
 983rtrs_clt_get_req(struct rtrs_clt_path *clt_path,
 984		 void (*conf)(void *priv, int errno),
 985		 struct rtrs_permit *permit, void *priv,
 986		 const struct kvec *vec, size_t usr_len,
 987		 struct scatterlist *sg, size_t sg_cnt,
 988		 size_t data_len, int dir)
 989{
 990	struct rtrs_clt_io_req *req;
 991
 992	req = &clt_path->reqs[permit->mem_id];
 993	rtrs_clt_init_req(req, clt_path, conf, permit, priv, vec, usr_len,
 994			   sg, sg_cnt, data_len, dir);
 995	return req;
 996}
 997
 998static struct rtrs_clt_io_req *
 999rtrs_clt_get_copy_req(struct rtrs_clt_path *alive_path,
1000		       struct rtrs_clt_io_req *fail_req)
1001{
1002	struct rtrs_clt_io_req *req;
1003	struct kvec vec = {
1004		.iov_base = fail_req->iu->buf,
1005		.iov_len  = fail_req->usr_len
1006	};
1007
1008	req = &alive_path->reqs[fail_req->permit->mem_id];
1009	rtrs_clt_init_req(req, alive_path, fail_req->conf, fail_req->permit,
1010			   fail_req->priv, &vec, fail_req->usr_len,
1011			   fail_req->sglist, fail_req->sg_cnt,
1012			   fail_req->data_len, fail_req->dir);
1013	return req;
1014}
1015
1016static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
1017				   struct rtrs_clt_io_req *req,
1018				   struct rtrs_rbuf *rbuf, bool fr_en,
1019				   u32 count, u32 size, u32 imm,
1020				   struct ib_send_wr *wr,
1021				   struct ib_send_wr *tail)
1022{
1023	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1024	struct ib_sge *sge = req->sge;
1025	enum ib_send_flags flags;
1026	struct scatterlist *sg;
1027	size_t num_sge;
1028	int i;
1029	struct ib_send_wr *ptail = NULL;
1030
1031	if (fr_en) {
1032		i = 0;
1033		sge[i].addr   = req->mr->iova;
1034		sge[i].length = req->mr->length;
1035		sge[i].lkey   = req->mr->lkey;
1036		i++;
1037		num_sge = 2;
1038		ptail = tail;
1039	} else {
1040		for_each_sg(req->sglist, sg, count, i) {
1041			sge[i].addr   = sg_dma_address(sg);
1042			sge[i].length = sg_dma_len(sg);
1043			sge[i].lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
1044		}
1045		num_sge = 1 + count;
1046	}
1047	sge[i].addr   = req->iu->dma_addr;
1048	sge[i].length = size;
1049	sge[i].lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
1050
1051	/*
1052	 * From time to time we have to post signalled sends,
1053	 * or send queue will fill up and only QP reset can help.
1054	 */
1055	flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
1056			0 : IB_SEND_SIGNALED;
1057
1058	ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
1059				      req->iu->dma_addr,
1060				      size, DMA_TO_DEVICE);
1061
1062	return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
1063					    rbuf->rkey, rbuf->addr, imm,
1064					    flags, wr, ptail);
1065}
1066
1067static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1068{
1069	int nr;
1070
1071	/* Align the MR to a 4K page size to match the block virt boundary */
1072	nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1073	if (nr != count)
1074		return nr < 0 ? nr : -EINVAL;
1075	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1076
1077	return nr;
1078}
1079
1080static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
1081{
1082	struct rtrs_clt_con *con = req->con;
1083	struct rtrs_path *s = con->c.path;
1084	struct rtrs_clt_path *clt_path = to_clt_path(s);
1085	struct rtrs_msg_rdma_write *msg;
1086
1087	struct rtrs_rbuf *rbuf;
1088	int ret, count = 0;
1089	u32 imm, buf_id;
1090	struct ib_reg_wr rwr;
1091	struct ib_send_wr *wr = NULL;
1092	bool fr_en = false;
1093
1094	const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1095
1096	if (tsize > clt_path->chunk_size) {
1097		rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
1098			  tsize, clt_path->chunk_size);
1099		return -EMSGSIZE;
1100	}
1101	if (req->sg_cnt) {
1102		count = ib_dma_map_sg(clt_path->s.dev->ib_dev, req->sglist,
1103				      req->sg_cnt, req->dir);
1104		if (!count) {
1105			rtrs_wrn(s, "Write request failed, map failed\n");
1106			return -EINVAL;
1107		}
1108	}
1109	/* put rtrs msg after sg and user message */
1110	msg = req->iu->buf + req->usr_len;
1111	msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1112	msg->usr_len = cpu_to_le16(req->usr_len);
1113
1114	/* rtrs message on server side will be after user data and message */
1115	imm = req->permit->mem_off + req->data_len + req->usr_len;
1116	imm = rtrs_to_io_req_imm(imm);
1117	buf_id = req->permit->mem_id;
1118	req->sg_size = tsize;
1119	rbuf = &clt_path->rbufs[buf_id];
1120
1121	if (count) {
1122		ret = rtrs_map_sg_fr(req, count);
1123		if (ret < 0) {
1124			rtrs_err_rl(s,
1125				    "Write request failed, failed to map fast reg. data, err: %d\n",
1126				    ret);
1127			ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1128					req->sg_cnt, req->dir);
1129			return ret;
1130		}
1131		rwr = (struct ib_reg_wr) {
1132			.wr.opcode = IB_WR_REG_MR,
1133			.wr.wr_cqe = &fast_reg_cqe,
1134			.mr = req->mr,
1135			.key = req->mr->rkey,
1136			.access = (IB_ACCESS_LOCAL_WRITE),
1137		};
1138		wr = &rwr.wr;
1139		fr_en = true;
1140		req->mr->need_inval = true;
1141	}
1142	/*
1143	 * Update stats now, after request is successfully sent it is not
1144	 * safe anymore to touch it.
1145	 */
1146	rtrs_clt_update_all_stats(req, WRITE);
1147
1148	ret = rtrs_post_rdma_write_sg(req->con, req, rbuf, fr_en, count,
1149				      req->usr_len + sizeof(*msg),
1150				      imm, wr, NULL);
1151	if (ret) {
1152		rtrs_err_rl(s,
1153			    "Write request failed: error=%d path=%s [%s:%u]\n",
1154			    ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1155			    clt_path->hca_port);
1156		if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1157			atomic_dec(&clt_path->stats->inflight);
1158		if (req->mr->need_inval) {
1159			req->mr->need_inval = false;
1160			refcount_dec(&req->ref);
1161		}
1162		if (req->sg_cnt)
1163			ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1164					req->sg_cnt, req->dir);
1165	}
1166
1167	return ret;
1168}
1169
1170static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1171{
1172	struct rtrs_clt_con *con = req->con;
1173	struct rtrs_path *s = con->c.path;
1174	struct rtrs_clt_path *clt_path = to_clt_path(s);
1175	struct rtrs_msg_rdma_read *msg;
1176	struct rtrs_ib_dev *dev = clt_path->s.dev;
1177
1178	struct ib_reg_wr rwr;
1179	struct ib_send_wr *wr = NULL;
1180
1181	int ret, count = 0;
1182	u32 imm, buf_id;
1183
1184	const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1185
1186	if (tsize > clt_path->chunk_size) {
1187		rtrs_wrn(s,
1188			  "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1189			  tsize, clt_path->chunk_size);
1190		return -EMSGSIZE;
1191	}
1192
1193	if (req->sg_cnt) {
1194		count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1195				      req->dir);
1196		if (!count) {
1197			rtrs_wrn(s,
1198				  "Read request failed, dma map failed\n");
1199			return -EINVAL;
1200		}
1201	}
1202	/* put our message into req->buf after user message*/
1203	msg = req->iu->buf + req->usr_len;
1204	msg->type = cpu_to_le16(RTRS_MSG_READ);
1205	msg->usr_len = cpu_to_le16(req->usr_len);
1206
1207	if (count) {
1208		ret = rtrs_map_sg_fr(req, count);
1209		if (ret < 0) {
1210			rtrs_err_rl(s,
1211				     "Read request failed, failed to map fast reg. data, err: %d\n",
1212				     ret);
1213			ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1214					req->dir);
1215			return ret;
1216		}
1217		rwr = (struct ib_reg_wr) {
1218			.wr.opcode = IB_WR_REG_MR,
1219			.wr.wr_cqe = &fast_reg_cqe,
1220			.mr = req->mr,
1221			.key = req->mr->rkey,
1222			.access = (IB_ACCESS_LOCAL_WRITE |
1223				   IB_ACCESS_REMOTE_WRITE),
1224		};
1225		wr = &rwr.wr;
1226
1227		msg->sg_cnt = cpu_to_le16(1);
1228		msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1229
1230		msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1231		msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1232		msg->desc[0].len = cpu_to_le32(req->mr->length);
1233
1234		/* Further invalidation is required */
1235		req->mr->need_inval = !!RTRS_MSG_NEED_INVAL_F;
1236
1237	} else {
1238		msg->sg_cnt = 0;
1239		msg->flags = 0;
1240	}
1241	/*
1242	 * rtrs message will be after the space reserved for disk data and
1243	 * user message
1244	 */
1245	imm = req->permit->mem_off + req->data_len + req->usr_len;
1246	imm = rtrs_to_io_req_imm(imm);
1247	buf_id = req->permit->mem_id;
1248
1249	req->sg_size  = sizeof(*msg);
1250	req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1251	req->sg_size += req->usr_len;
1252
1253	/*
1254	 * Update stats now, after request is successfully sent it is not
1255	 * safe anymore to touch it.
1256	 */
1257	rtrs_clt_update_all_stats(req, READ);
1258
1259	ret = rtrs_post_send_rdma(req->con, req, &clt_path->rbufs[buf_id],
1260				   req->data_len, imm, wr);
1261	if (ret) {
1262		rtrs_err_rl(s,
1263			    "Read request failed: error=%d path=%s [%s:%u]\n",
1264			    ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1265			    clt_path->hca_port);
1266		if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1267			atomic_dec(&clt_path->stats->inflight);
1268		req->mr->need_inval = false;
1269		if (req->sg_cnt)
1270			ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1271					req->sg_cnt, req->dir);
1272	}
1273
1274	return ret;
1275}
1276
1277/**
1278 * rtrs_clt_failover_req() - Try to find an active path for a failed request
1279 * @clt: clt context
1280 * @fail_req: a failed io request.
1281 */
1282static int rtrs_clt_failover_req(struct rtrs_clt_sess *clt,
1283				 struct rtrs_clt_io_req *fail_req)
1284{
1285	struct rtrs_clt_path *alive_path;
1286	struct rtrs_clt_io_req *req;
1287	int err = -ECONNABORTED;
1288	struct path_it it;
1289
1290	rcu_read_lock();
1291	for (path_it_init(&it, clt);
1292	     (alive_path = it.next_path(&it)) && it.i < it.clt->paths_num;
1293	     it.i++) {
1294		if (READ_ONCE(alive_path->state) != RTRS_CLT_CONNECTED)
1295			continue;
1296		req = rtrs_clt_get_copy_req(alive_path, fail_req);
1297		if (req->dir == DMA_TO_DEVICE)
1298			err = rtrs_clt_write_req(req);
1299		else
1300			err = rtrs_clt_read_req(req);
1301		if (err) {
1302			req->in_use = false;
1303			continue;
1304		}
1305		/* Success path */
1306		rtrs_clt_inc_failover_cnt(alive_path->stats);
1307		break;
1308	}
1309	path_it_deinit(&it);
1310	rcu_read_unlock();
1311
1312	return err;
1313}
1314
1315static void fail_all_outstanding_reqs(struct rtrs_clt_path *clt_path)
1316{
1317	struct rtrs_clt_sess *clt = clt_path->clt;
1318	struct rtrs_clt_io_req *req;
1319	int i, err;
1320
1321	if (!clt_path->reqs)
1322		return;
1323	for (i = 0; i < clt_path->queue_depth; ++i) {
1324		req = &clt_path->reqs[i];
1325		if (!req->in_use)
1326			continue;
1327
1328		/*
1329		 * Safely (without notification) complete failed request.
1330		 * After completion this request is still useble and can
1331		 * be failovered to another path.
1332		 */
1333		complete_rdma_req(req, -ECONNABORTED, false, true);
1334
1335		err = rtrs_clt_failover_req(clt, req);
1336		if (err)
1337			/* Failover failed, notify anyway */
1338			req->conf(req->priv, err);
1339	}
1340}
1341
1342static void free_path_reqs(struct rtrs_clt_path *clt_path)
1343{
1344	struct rtrs_clt_io_req *req;
1345	int i;
1346
1347	if (!clt_path->reqs)
1348		return;
1349	for (i = 0; i < clt_path->queue_depth; ++i) {
1350		req = &clt_path->reqs[i];
1351		if (req->mr)
1352			ib_dereg_mr(req->mr);
1353		kfree(req->sge);
1354		rtrs_iu_free(req->iu, clt_path->s.dev->ib_dev, 1);
1355	}
1356	kfree(clt_path->reqs);
1357	clt_path->reqs = NULL;
1358}
1359
1360static int alloc_path_reqs(struct rtrs_clt_path *clt_path)
1361{
1362	struct rtrs_clt_io_req *req;
1363	int i, err = -ENOMEM;
1364
1365	clt_path->reqs = kcalloc(clt_path->queue_depth,
1366				 sizeof(*clt_path->reqs),
1367				 GFP_KERNEL);
1368	if (!clt_path->reqs)
1369		return -ENOMEM;
1370
1371	for (i = 0; i < clt_path->queue_depth; ++i) {
1372		req = &clt_path->reqs[i];
1373		req->iu = rtrs_iu_alloc(1, clt_path->max_hdr_size, GFP_KERNEL,
1374					 clt_path->s.dev->ib_dev,
1375					 DMA_TO_DEVICE,
1376					 rtrs_clt_rdma_done);
1377		if (!req->iu)
1378			goto out;
1379
1380		req->sge = kcalloc(2, sizeof(*req->sge), GFP_KERNEL);
1381		if (!req->sge)
1382			goto out;
1383
1384		req->mr = ib_alloc_mr(clt_path->s.dev->ib_pd,
1385				      IB_MR_TYPE_MEM_REG,
1386				      clt_path->max_pages_per_mr);
1387		if (IS_ERR(req->mr)) {
1388			err = PTR_ERR(req->mr);
1389			pr_err("Failed to alloc clt_path->max_pages_per_mr %d: %pe\n",
1390			       clt_path->max_pages_per_mr, req->mr);
1391			req->mr = NULL;
1392			goto out;
1393		}
1394
1395		init_completion(&req->inv_comp);
1396	}
1397
1398	return 0;
1399
1400out:
1401	free_path_reqs(clt_path);
1402
1403	return err;
1404}
1405
1406static int alloc_permits(struct rtrs_clt_sess *clt)
1407{
1408	unsigned int chunk_bits;
1409	int err, i;
1410
1411	clt->permits_map = bitmap_zalloc(clt->queue_depth, GFP_KERNEL);
1412	if (!clt->permits_map) {
1413		err = -ENOMEM;
1414		goto out_err;
1415	}
1416	clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1417	if (!clt->permits) {
1418		err = -ENOMEM;
1419		goto err_map;
1420	}
1421	chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1422	for (i = 0; i < clt->queue_depth; i++) {
1423		struct rtrs_permit *permit;
1424
1425		permit = get_permit(clt, i);
1426		permit->mem_id = i;
1427		permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1428	}
1429
1430	return 0;
1431
1432err_map:
1433	bitmap_free(clt->permits_map);
1434	clt->permits_map = NULL;
1435out_err:
1436	return err;
1437}
1438
1439static void free_permits(struct rtrs_clt_sess *clt)
1440{
1441	if (clt->permits_map)
1442		wait_event(clt->permits_wait,
1443			   bitmap_empty(clt->permits_map, clt->queue_depth));
1444
1445	bitmap_free(clt->permits_map);
1446	clt->permits_map = NULL;
1447	kfree(clt->permits);
1448	clt->permits = NULL;
1449}
1450
1451static void query_fast_reg_mode(struct rtrs_clt_path *clt_path)
1452{
1453	struct ib_device *ib_dev;
1454	u64 max_pages_per_mr;
1455	int mr_page_shift;
1456
1457	ib_dev = clt_path->s.dev->ib_dev;
1458
1459	/*
1460	 * Use the smallest page size supported by the HCA, down to a
1461	 * minimum of 4096 bytes. We're unlikely to build large sglists
1462	 * out of smaller entries.
1463	 */
1464	mr_page_shift      = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1465	max_pages_per_mr   = ib_dev->attrs.max_mr_size;
1466	do_div(max_pages_per_mr, (1ull << mr_page_shift));
1467	max_pages_per_mr = min_not_zero((u32)max_pages_per_mr, U32_MAX);
1468	clt_path->max_pages_per_mr =
1469		min3(clt_path->max_pages_per_mr, (u32)max_pages_per_mr,
1470		     ib_dev->attrs.max_fast_reg_page_list_len);
1471	clt_path->clt->max_segments =
1472		min(clt_path->max_pages_per_mr, clt_path->clt->max_segments);
1473}
1474
1475static bool rtrs_clt_change_state_get_old(struct rtrs_clt_path *clt_path,
1476					   enum rtrs_clt_state new_state,
1477					   enum rtrs_clt_state *old_state)
1478{
1479	bool changed;
1480
1481	spin_lock_irq(&clt_path->state_wq.lock);
1482	if (old_state)
1483		*old_state = clt_path->state;
1484	changed = rtrs_clt_change_state(clt_path, new_state);
1485	spin_unlock_irq(&clt_path->state_wq.lock);
1486
1487	return changed;
1488}
1489
1490static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1491{
1492	struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1493	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1494
1495	rtrs_err(con->c.path, "HB err handler for path=%s\n", kobject_name(&clt_path->kobj));
1496	rtrs_rdma_error_recovery(con);
1497}
1498
1499static void rtrs_clt_init_hb(struct rtrs_clt_path *clt_path)
1500{
1501	rtrs_init_hb(&clt_path->s, &io_comp_cqe,
1502		      RTRS_HB_INTERVAL_MS,
1503		      RTRS_HB_MISSED_MAX,
1504		      rtrs_clt_hb_err_handler,
1505		      rtrs_wq);
1506}
1507
1508static void rtrs_clt_reconnect_work(struct work_struct *work);
1509static void rtrs_clt_close_work(struct work_struct *work);
1510
1511static void rtrs_clt_err_recovery_work(struct work_struct *work)
1512{
1513	struct rtrs_clt_path *clt_path;
1514	struct rtrs_clt_sess *clt;
1515	int delay_ms;
1516
1517	clt_path = container_of(work, struct rtrs_clt_path, err_recovery_work);
1518	clt = clt_path->clt;
1519	delay_ms = clt->reconnect_delay_sec * 1000;
1520	rtrs_clt_stop_and_destroy_conns(clt_path);
1521	queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork,
1522			   msecs_to_jiffies(delay_ms +
1523					    get_random_u32_below(RTRS_RECONNECT_SEED)));
1524}
1525
1526static struct rtrs_clt_path *alloc_path(struct rtrs_clt_sess *clt,
1527					const struct rtrs_addr *path,
1528					size_t con_num, u32 nr_poll_queues)
1529{
1530	struct rtrs_clt_path *clt_path;
1531	int err = -ENOMEM;
1532	int cpu;
1533	size_t total_con;
1534
1535	clt_path = kzalloc(sizeof(*clt_path), GFP_KERNEL);
1536	if (!clt_path)
1537		goto err;
1538
1539	/*
1540	 * irqmode and poll
1541	 * +1: Extra connection for user messages
1542	 */
1543	total_con = con_num + nr_poll_queues + 1;
1544	clt_path->s.con = kcalloc(total_con, sizeof(*clt_path->s.con),
1545				  GFP_KERNEL);
1546	if (!clt_path->s.con)
1547		goto err_free_path;
1548
1549	clt_path->s.con_num = total_con;
1550	clt_path->s.irq_con_num = con_num + 1;
1551
1552	clt_path->stats = kzalloc(sizeof(*clt_path->stats), GFP_KERNEL);
1553	if (!clt_path->stats)
1554		goto err_free_con;
1555
1556	mutex_init(&clt_path->init_mutex);
1557	uuid_gen(&clt_path->s.uuid);
1558	memcpy(&clt_path->s.dst_addr, path->dst,
1559	       rdma_addr_size((struct sockaddr *)path->dst));
1560
1561	/*
1562	 * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1563	 * checks the sa_family to be non-zero. If user passed src_addr=NULL
1564	 * the sess->src_addr will contain only zeros, which is then fine.
1565	 */
1566	if (path->src)
1567		memcpy(&clt_path->s.src_addr, path->src,
1568		       rdma_addr_size((struct sockaddr *)path->src));
1569	strscpy(clt_path->s.sessname, clt->sessname,
1570		sizeof(clt_path->s.sessname));
1571	clt_path->clt = clt;
1572	clt_path->max_pages_per_mr = RTRS_MAX_SEGMENTS;
1573	init_waitqueue_head(&clt_path->state_wq);
1574	clt_path->state = RTRS_CLT_CONNECTING;
1575	atomic_set(&clt_path->connected_cnt, 0);
1576	INIT_WORK(&clt_path->close_work, rtrs_clt_close_work);
1577	INIT_WORK(&clt_path->err_recovery_work, rtrs_clt_err_recovery_work);
1578	INIT_DELAYED_WORK(&clt_path->reconnect_dwork, rtrs_clt_reconnect_work);
1579	rtrs_clt_init_hb(clt_path);
1580
1581	clt_path->mp_skip_entry = alloc_percpu(typeof(*clt_path->mp_skip_entry));
1582	if (!clt_path->mp_skip_entry)
1583		goto err_free_stats;
1584
1585	for_each_possible_cpu(cpu)
1586		INIT_LIST_HEAD(per_cpu_ptr(clt_path->mp_skip_entry, cpu));
1587
1588	err = rtrs_clt_init_stats(clt_path->stats);
1589	if (err)
1590		goto err_free_percpu;
1591
1592	return clt_path;
1593
1594err_free_percpu:
1595	free_percpu(clt_path->mp_skip_entry);
1596err_free_stats:
1597	kfree(clt_path->stats);
1598err_free_con:
1599	kfree(clt_path->s.con);
1600err_free_path:
1601	kfree(clt_path);
1602err:
1603	return ERR_PTR(err);
1604}
1605
1606void free_path(struct rtrs_clt_path *clt_path)
1607{
1608	free_percpu(clt_path->mp_skip_entry);
1609	mutex_destroy(&clt_path->init_mutex);
1610	kfree(clt_path->s.con);
1611	kfree(clt_path->rbufs);
1612	kfree(clt_path);
1613}
1614
1615static int create_con(struct rtrs_clt_path *clt_path, unsigned int cid)
1616{
1617	struct rtrs_clt_con *con;
1618
1619	con = kzalloc(sizeof(*con), GFP_KERNEL);
1620	if (!con)
1621		return -ENOMEM;
1622
1623	/* Map first two connections to the first CPU */
1624	con->cpu  = (cid ? cid - 1 : 0) % nr_cpu_ids;
1625	con->c.cid = cid;
1626	con->c.path = &clt_path->s;
1627	/* Align with srv, init as 1 */
1628	atomic_set(&con->c.wr_cnt, 1);
1629	mutex_init(&con->con_mutex);
1630
1631	clt_path->s.con[cid] = &con->c;
1632
1633	return 0;
1634}
1635
1636static void destroy_con(struct rtrs_clt_con *con)
1637{
1638	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1639
1640	clt_path->s.con[con->c.cid] = NULL;
1641	mutex_destroy(&con->con_mutex);
1642	kfree(con);
1643}
1644
1645static int create_con_cq_qp(struct rtrs_clt_con *con)
1646{
1647	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1648	u32 max_send_wr, max_recv_wr, cq_num, max_send_sge, wr_limit;
1649	int err, cq_vector;
1650	struct rtrs_msg_rkey_rsp *rsp;
1651
1652	lockdep_assert_held(&con->con_mutex);
1653	if (con->c.cid == 0) {
1654		max_send_sge = 1;
1655		/* We must be the first here */
1656		if (WARN_ON(clt_path->s.dev))
1657			return -EINVAL;
1658
1659		/*
1660		 * The whole session uses device from user connection.
1661		 * Be careful not to close user connection before ib dev
1662		 * is gracefully put.
1663		 */
1664		clt_path->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1665						       &dev_pd);
1666		if (!clt_path->s.dev) {
1667			rtrs_wrn(clt_path->clt,
1668				  "rtrs_ib_dev_find_get_or_add(): no memory\n");
1669			return -ENOMEM;
1670		}
1671		clt_path->s.dev_ref = 1;
1672		query_fast_reg_mode(clt_path);
1673		wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1674		/*
1675		 * Two (request + registration) completion for send
1676		 * Two for recv if always_invalidate is set on server
1677		 * or one for recv.
1678		 * + 2 for drain and heartbeat
1679		 * in case qp gets into error state.
1680		 */
1681		max_send_wr =
1682			min_t(int, wr_limit, SERVICE_CON_QUEUE_DEPTH * 2 + 2);
1683		max_recv_wr = max_send_wr;
1684	} else {
1685		/*
1686		 * Here we assume that session members are correctly set.
1687		 * This is always true if user connection (cid == 0) is
1688		 * established first.
1689		 */
1690		if (WARN_ON(!clt_path->s.dev))
1691			return -EINVAL;
1692		if (WARN_ON(!clt_path->queue_depth))
1693			return -EINVAL;
1694
1695		wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1696		/* Shared between connections */
1697		clt_path->s.dev_ref++;
1698		max_send_wr = min_t(int, wr_limit,
1699			      /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1700			      clt_path->queue_depth * 4 + 1);
1701		max_recv_wr = min_t(int, wr_limit,
1702			      clt_path->queue_depth * 3 + 1);
1703		max_send_sge = 2;
1704	}
1705	atomic_set(&con->c.sq_wr_avail, max_send_wr);
1706	cq_num = max_send_wr + max_recv_wr;
1707	/* alloc iu to recv new rkey reply when server reports flags set */
1708	if (clt_path->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1709		con->rsp_ius = rtrs_iu_alloc(cq_num, sizeof(*rsp),
1710					      GFP_KERNEL,
1711					      clt_path->s.dev->ib_dev,
1712					      DMA_FROM_DEVICE,
1713					      rtrs_clt_rdma_done);
1714		if (!con->rsp_ius)
1715			return -ENOMEM;
1716		con->queue_num = cq_num;
1717	}
1718	cq_vector = con->cpu % clt_path->s.dev->ib_dev->num_comp_vectors;
1719	if (con->c.cid >= clt_path->s.irq_con_num)
1720		err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1721					cq_vector, cq_num, max_send_wr,
1722					max_recv_wr, IB_POLL_DIRECT);
1723	else
1724		err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1725					cq_vector, cq_num, max_send_wr,
1726					max_recv_wr, IB_POLL_SOFTIRQ);
1727	/*
1728	 * In case of error we do not bother to clean previous allocations,
1729	 * since destroy_con_cq_qp() must be called.
1730	 */
1731	return err;
1732}
1733
1734static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1735{
1736	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1737
1738	/*
1739	 * Be careful here: destroy_con_cq_qp() can be called even
1740	 * create_con_cq_qp() failed, see comments there.
1741	 */
1742	lockdep_assert_held(&con->con_mutex);
1743	rtrs_cq_qp_destroy(&con->c);
1744	if (con->rsp_ius) {
1745		rtrs_iu_free(con->rsp_ius, clt_path->s.dev->ib_dev,
1746			     con->queue_num);
1747		con->rsp_ius = NULL;
1748		con->queue_num = 0;
1749	}
1750	if (clt_path->s.dev_ref && !--clt_path->s.dev_ref) {
1751		rtrs_ib_dev_put(clt_path->s.dev);
1752		clt_path->s.dev = NULL;
1753	}
1754}
1755
1756static void stop_cm(struct rtrs_clt_con *con)
1757{
1758	rdma_disconnect(con->c.cm_id);
1759	if (con->c.qp)
1760		ib_drain_qp(con->c.qp);
1761}
1762
1763static void destroy_cm(struct rtrs_clt_con *con)
1764{
1765	rdma_destroy_id(con->c.cm_id);
1766	con->c.cm_id = NULL;
1767}
1768
1769static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1770{
1771	struct rtrs_path *s = con->c.path;
1772	int err;
1773
1774	mutex_lock(&con->con_mutex);
1775	err = create_con_cq_qp(con);
1776	mutex_unlock(&con->con_mutex);
1777	if (err) {
1778		rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1779		return err;
1780	}
1781	err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1782	if (err)
1783		rtrs_err(s, "Resolving route failed, err: %d\n", err);
1784
1785	return err;
1786}
1787
1788static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1789{
1790	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1791	struct rtrs_clt_sess *clt = clt_path->clt;
1792	struct rtrs_msg_conn_req msg;
1793	struct rdma_conn_param param;
1794
1795	int err;
1796
1797	param = (struct rdma_conn_param) {
1798		.retry_count = 7,
1799		.rnr_retry_count = 7,
1800		.private_data = &msg,
1801		.private_data_len = sizeof(msg),
1802	};
1803
1804	msg = (struct rtrs_msg_conn_req) {
1805		.magic = cpu_to_le16(RTRS_MAGIC),
1806		.version = cpu_to_le16(RTRS_PROTO_VER),
1807		.cid = cpu_to_le16(con->c.cid),
1808		.cid_num = cpu_to_le16(clt_path->s.con_num),
1809		.recon_cnt = cpu_to_le16(clt_path->s.recon_cnt),
1810	};
1811	msg.first_conn = clt_path->for_new_clt ? FIRST_CONN : 0;
1812	uuid_copy(&msg.sess_uuid, &clt_path->s.uuid);
1813	uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1814
1815	err = rdma_connect_locked(con->c.cm_id, &param);
1816	if (err)
1817		rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1818
1819	return err;
1820}
1821
1822static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1823				       struct rdma_cm_event *ev)
1824{
1825	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1826	struct rtrs_clt_sess *clt = clt_path->clt;
1827	const struct rtrs_msg_conn_rsp *msg;
1828	u16 version, queue_depth;
1829	int errno;
1830	u8 len;
1831
1832	msg = ev->param.conn.private_data;
1833	len = ev->param.conn.private_data_len;
1834	if (len < sizeof(*msg)) {
1835		rtrs_err(clt, "Invalid RTRS connection response\n");
1836		return -ECONNRESET;
1837	}
1838	if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1839		rtrs_err(clt, "Invalid RTRS magic\n");
1840		return -ECONNRESET;
1841	}
1842	version = le16_to_cpu(msg->version);
1843	if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1844		rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1845			  version >> 8, RTRS_PROTO_VER_MAJOR);
1846		return -ECONNRESET;
1847	}
1848	errno = le16_to_cpu(msg->errno);
1849	if (errno) {
1850		rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1851			  errno);
1852		return -ECONNRESET;
1853	}
1854	if (con->c.cid == 0) {
1855		queue_depth = le16_to_cpu(msg->queue_depth);
1856
1857		if (clt_path->queue_depth > 0 && queue_depth != clt_path->queue_depth) {
1858			rtrs_err(clt, "Error: queue depth changed\n");
1859
1860			/*
1861			 * Stop any more reconnection attempts
1862			 */
1863			clt_path->reconnect_attempts = -1;
1864			rtrs_err(clt,
1865				"Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n");
1866			return -ECONNRESET;
1867		}
1868
1869		if (!clt_path->rbufs) {
1870			clt_path->rbufs = kcalloc(queue_depth,
1871						  sizeof(*clt_path->rbufs),
1872						  GFP_KERNEL);
1873			if (!clt_path->rbufs)
1874				return -ENOMEM;
1875		}
1876		clt_path->queue_depth = queue_depth;
1877		clt_path->s.signal_interval = min_not_zero(queue_depth,
1878						(unsigned short) SERVICE_CON_QUEUE_DEPTH);
1879		clt_path->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1880		clt_path->max_io_size = le32_to_cpu(msg->max_io_size);
1881		clt_path->flags = le32_to_cpu(msg->flags);
1882		clt_path->chunk_size = clt_path->max_io_size + clt_path->max_hdr_size;
1883
1884		/*
1885		 * Global IO size is always a minimum.
1886		 * If while a reconnection server sends us a value a bit
1887		 * higher - client does not care and uses cached minimum.
1888		 *
1889		 * Since we can have several sessions (paths) restablishing
1890		 * connections in parallel, use lock.
1891		 */
1892		mutex_lock(&clt->paths_mutex);
1893		clt->queue_depth = clt_path->queue_depth;
1894		clt->max_io_size = min_not_zero(clt_path->max_io_size,
1895						clt->max_io_size);
1896		mutex_unlock(&clt->paths_mutex);
1897
1898		/*
1899		 * Cache the hca_port and hca_name for sysfs
1900		 */
1901		clt_path->hca_port = con->c.cm_id->port_num;
1902		scnprintf(clt_path->hca_name, sizeof(clt_path->hca_name),
1903			  clt_path->s.dev->ib_dev->name);
1904		clt_path->s.src_addr = con->c.cm_id->route.addr.src_addr;
1905		/* set for_new_clt, to allow future reconnect on any path */
1906		clt_path->for_new_clt = 1;
1907	}
1908
1909	return 0;
1910}
1911
1912static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1913{
1914	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1915
1916	atomic_inc(&clt_path->connected_cnt);
1917	con->cm_err = 1;
1918}
1919
1920static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1921				    struct rdma_cm_event *ev)
1922{
1923	struct rtrs_path *s = con->c.path;
1924	const struct rtrs_msg_conn_rsp *msg;
1925	const char *rej_msg;
1926	int status, errno = -ECONNRESET;
1927	u8 data_len;
1928
1929	status = ev->status;
1930	rej_msg = rdma_reject_msg(con->c.cm_id, status);
1931	msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1932
1933	if (msg && data_len >= sizeof(*msg)) {
1934		errno = (int16_t)le16_to_cpu(msg->errno);
1935		if (errno == -EBUSY)
1936			rtrs_err(s,
1937				  "Previous session is still exists on the server, please reconnect later\n");
1938		else
1939			rtrs_err(s,
1940				  "Connect rejected: status %d (%s), rtrs errno %d\n",
1941				  status, rej_msg, errno);
1942	} else {
1943		rtrs_err(s,
1944			  "Connect rejected but with malformed message: status %d (%s)\n",
1945			  status, rej_msg);
1946	}
1947
1948	return errno;
1949}
1950
1951void rtrs_clt_close_conns(struct rtrs_clt_path *clt_path, bool wait)
1952{
1953	trace_rtrs_clt_close_conns(clt_path);
1954
1955	if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSING, NULL))
1956		queue_work(rtrs_wq, &clt_path->close_work);
1957	if (wait)
1958		flush_work(&clt_path->close_work);
1959}
1960
1961static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1962{
1963	if (con->cm_err == 1) {
1964		struct rtrs_clt_path *clt_path;
1965
1966		clt_path = to_clt_path(con->c.path);
1967		if (atomic_dec_and_test(&clt_path->connected_cnt))
1968
1969			wake_up(&clt_path->state_wq);
1970	}
1971	con->cm_err = cm_err;
1972}
1973
1974static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1975				     struct rdma_cm_event *ev)
1976{
1977	struct rtrs_clt_con *con = cm_id->context;
1978	struct rtrs_path *s = con->c.path;
1979	struct rtrs_clt_path *clt_path = to_clt_path(s);
1980	int cm_err = 0;
1981
1982	switch (ev->event) {
1983	case RDMA_CM_EVENT_ADDR_RESOLVED:
1984		cm_err = rtrs_rdma_addr_resolved(con);
1985		break;
1986	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1987		cm_err = rtrs_rdma_route_resolved(con);
1988		break;
1989	case RDMA_CM_EVENT_ESTABLISHED:
1990		cm_err = rtrs_rdma_conn_established(con, ev);
1991		if (!cm_err) {
1992			/*
1993			 * Report success and wake up. Here we abuse state_wq,
1994			 * i.e. wake up without state change, but we set cm_err.
1995			 */
1996			flag_success_on_conn(con);
1997			wake_up(&clt_path->state_wq);
1998			return 0;
1999		}
2000		break;
2001	case RDMA_CM_EVENT_REJECTED:
2002		cm_err = rtrs_rdma_conn_rejected(con, ev);
2003		break;
2004	case RDMA_CM_EVENT_DISCONNECTED:
2005		/* No message for disconnecting */
2006		cm_err = -ECONNRESET;
2007		break;
2008	case RDMA_CM_EVENT_CONNECT_ERROR:
2009	case RDMA_CM_EVENT_UNREACHABLE:
2010	case RDMA_CM_EVENT_ADDR_CHANGE:
2011	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2012		rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2013			 rdma_event_msg(ev->event), ev->status);
2014		cm_err = -ECONNRESET;
2015		break;
2016	case RDMA_CM_EVENT_ADDR_ERROR:
2017	case RDMA_CM_EVENT_ROUTE_ERROR:
2018		rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2019			 rdma_event_msg(ev->event), ev->status);
2020		cm_err = -EHOSTUNREACH;
2021		break;
2022	case RDMA_CM_EVENT_DEVICE_REMOVAL:
2023		/*
2024		 * Device removal is a special case.  Queue close and return 0.
2025		 */
2026		rtrs_wrn_rl(s, "CM event: %s, status: %d\n", rdma_event_msg(ev->event),
2027			    ev->status);
2028		rtrs_clt_close_conns(clt_path, false);
2029		return 0;
2030	default:
2031		rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %d)\n",
2032			 rdma_event_msg(ev->event), ev->status);
2033		cm_err = -ECONNRESET;
2034		break;
2035	}
2036
2037	if (cm_err) {
2038		/*
2039		 * cm error makes sense only on connection establishing,
2040		 * in other cases we rely on normal procedure of reconnecting.
2041		 */
2042		flag_error_on_conn(con, cm_err);
2043		rtrs_rdma_error_recovery(con);
2044	}
2045
2046	return 0;
2047}
2048
2049/* The caller should do the cleanup in case of error */
2050static int create_cm(struct rtrs_clt_con *con)
2051{
2052	struct rtrs_path *s = con->c.path;
2053	struct rtrs_clt_path *clt_path = to_clt_path(s);
2054	struct rdma_cm_id *cm_id;
2055	int err;
2056
2057	cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
2058			       clt_path->s.dst_addr.ss_family == AF_IB ?
2059			       RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
2060	if (IS_ERR(cm_id)) {
2061		rtrs_err(s, "Failed to create CM ID, err: %pe\n", cm_id);
2062		return PTR_ERR(cm_id);
2063	}
2064	con->c.cm_id = cm_id;
2065	con->cm_err = 0;
2066	/* allow the port to be reused */
2067	err = rdma_set_reuseaddr(cm_id, 1);
2068	if (err != 0) {
2069		rtrs_err(s, "Set address reuse failed, err: %d\n", err);
2070		return err;
2071	}
2072	err = rdma_resolve_addr(cm_id, (struct sockaddr *)&clt_path->s.src_addr,
2073				(struct sockaddr *)&clt_path->s.dst_addr,
2074				RTRS_CONNECT_TIMEOUT_MS);
2075	if (err) {
2076		rtrs_err(s, "Failed to resolve address, err: %d\n", err);
2077		return err;
2078	}
2079	/*
2080	 * Combine connection status and session events. This is needed
2081	 * for waiting two possible cases: cm_err has something meaningful
2082	 * or session state was really changed to error by device removal.
2083	 */
2084	err = wait_event_interruptible_timeout(
2085			clt_path->state_wq,
2086			con->cm_err || clt_path->state != RTRS_CLT_CONNECTING,
2087			msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2088	if (err == 0 || err == -ERESTARTSYS) {
2089		if (err == 0)
2090			err = -ETIMEDOUT;
2091		/* Timedout or interrupted */
2092		return err;
2093	}
2094	if (con->cm_err < 0)
2095		return con->cm_err;
2096	if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTING)
2097		/* Device removal */
2098		return -ECONNABORTED;
2099
2100	return 0;
2101}
2102
2103static void rtrs_clt_path_up(struct rtrs_clt_path *clt_path)
2104{
2105	struct rtrs_clt_sess *clt = clt_path->clt;
2106	int up;
2107
2108	/*
2109	 * We can fire RECONNECTED event only when all paths were
2110	 * connected on rtrs_clt_open(), then each was disconnected
2111	 * and the first one connected again.  That's why this nasty
2112	 * game with counter value.
2113	 */
2114
2115	mutex_lock(&clt->paths_ev_mutex);
2116	up = ++clt->paths_up;
2117	/*
2118	 * Here it is safe to access paths num directly since up counter
2119	 * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
2120	 * in progress, thus paths removals are impossible.
2121	 */
2122	if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
2123		clt->paths_up = clt->paths_num;
2124	else if (up == 1)
2125		clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
2126	mutex_unlock(&clt->paths_ev_mutex);
2127
2128	/* Mark session as established */
2129	clt_path->established = true;
2130	clt_path->reconnect_attempts = 0;
2131	clt_path->stats->reconnects.successful_cnt++;
2132}
2133
2134static void rtrs_clt_path_down(struct rtrs_clt_path *clt_path)
2135{
2136	struct rtrs_clt_sess *clt = clt_path->clt;
2137
2138	if (!clt_path->established)
2139		return;
2140
2141	clt_path->established = false;
2142	mutex_lock(&clt->paths_ev_mutex);
2143	WARN_ON(!clt->paths_up);
2144	if (--clt->paths_up == 0)
2145		clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2146	mutex_unlock(&clt->paths_ev_mutex);
2147}
2148
2149static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path)
2150{
2151	struct rtrs_clt_con *con;
2152	unsigned int cid;
2153
2154	WARN_ON(READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED);
2155
2156	/*
2157	 * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2158	 * exactly in between.  Start destroying after it finishes.
2159	 */
2160	mutex_lock(&clt_path->init_mutex);
2161	mutex_unlock(&clt_path->init_mutex);
2162
2163	/*
2164	 * All IO paths must observe !CONNECTED state before we
2165	 * free everything.
2166	 */
2167	synchronize_rcu();
2168
2169	rtrs_stop_hb(&clt_path->s);
2170
2171	/*
2172	 * The order it utterly crucial: firstly disconnect and complete all
2173	 * rdma requests with error (thus set in_use=false for requests),
2174	 * then fail outstanding requests checking in_use for each, and
2175	 * eventually notify upper layer about session disconnection.
2176	 */
2177
2178	for (cid = 0; cid < clt_path->s.con_num; cid++) {
2179		if (!clt_path->s.con[cid])
2180			break;
2181		con = to_clt_con(clt_path->s.con[cid]);
2182		stop_cm(con);
2183	}
2184	fail_all_outstanding_reqs(clt_path);
2185	free_path_reqs(clt_path);
2186	rtrs_clt_path_down(clt_path);
2187
2188	/*
2189	 * Wait for graceful shutdown, namely when peer side invokes
2190	 * rdma_disconnect(). 'connected_cnt' is decremented only on
2191	 * CM events, thus if other side had crashed and hb has detected
2192	 * something is wrong, here we will stuck for exactly timeout ms,
2193	 * since CM does not fire anything.  That is fine, we are not in
2194	 * hurry.
2195	 */
2196	wait_event_timeout(clt_path->state_wq,
2197			   !atomic_read(&clt_path->connected_cnt),
2198			   msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2199
2200	for (cid = 0; cid < clt_path->s.con_num; cid++) {
2201		if (!clt_path->s.con[cid])
2202			break;
2203		con = to_clt_con(clt_path->s.con[cid]);
2204		mutex_lock(&con->con_mutex);
2205		destroy_con_cq_qp(con);
2206		mutex_unlock(&con->con_mutex);
2207		destroy_cm(con);
2208		destroy_con(con);
2209	}
2210}
2211
2212static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_path *clt_path)
2213{
2214	struct rtrs_clt_sess *clt = clt_path->clt;
2215	struct rtrs_clt_path *next;
2216	bool wait_for_grace = false;
2217	int cpu;
2218
2219	mutex_lock(&clt->paths_mutex);
2220	list_del_rcu(&clt_path->s.entry);
2221
2222	/* Make sure everybody observes path removal. */
2223	synchronize_rcu();
2224
2225	/*
2226	 * At this point nobody sees @sess in the list, but still we have
2227	 * dangling pointer @pcpu_path which _can_ point to @sess.  Since
2228	 * nobody can observe @sess in the list, we guarantee that IO path
2229	 * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2230	 * to @sess, but can never again become @sess.
2231	 */
2232
2233	/*
2234	 * Decrement paths number only after grace period, because
2235	 * caller of do_each_path() must firstly observe list without
2236	 * path and only then decremented paths number.
2237	 *
2238	 * Otherwise there can be the following situation:
2239	 *    o Two paths exist and IO is coming.
2240	 *    o One path is removed:
2241	 *      CPU#0                          CPU#1
2242	 *      do_each_path():                rtrs_clt_remove_path_from_arr():
2243	 *          path = get_next_path()
2244	 *          ^^^                            list_del_rcu(path)
2245	 *          [!CONNECTED path]              clt->paths_num--
2246	 *                                              ^^^^^^^^^
2247	 *          load clt->paths_num                 from 2 to 1
2248	 *                    ^^^^^^^^^
2249	 *                    sees 1
2250	 *
2251	 *      path is observed as !CONNECTED, but do_each_path() loop
2252	 *      ends, because expression i < clt->paths_num is false.
2253	 */
2254	clt->paths_num--;
2255
2256	/*
2257	 * Get @next connection from current @sess which is going to be
2258	 * removed.  If @sess is the last element, then @next is NULL.
2259	 */
2260	rcu_read_lock();
2261	next = rtrs_clt_get_next_path_or_null(&clt->paths_list, clt_path);
2262	rcu_read_unlock();
2263
2264	/*
2265	 * @pcpu paths can still point to the path which is going to be
2266	 * removed, so change the pointer manually.
2267	 */
2268	for_each_possible_cpu(cpu) {
2269		struct rtrs_clt_path __rcu **ppcpu_path;
2270
2271		ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2272		if (rcu_dereference_protected(*ppcpu_path,
2273			lockdep_is_held(&clt->paths_mutex)) != clt_path)
2274			/*
2275			 * synchronize_rcu() was called just after deleting
2276			 * entry from the list, thus IO code path cannot
2277			 * change pointer back to the pointer which is going
2278			 * to be removed, we are safe here.
2279			 */
2280			continue;
2281
2282		/*
2283		 * We race with IO code path, which also changes pointer,
2284		 * thus we have to be careful not to overwrite it.
2285		 */
2286		if (try_cmpxchg((struct rtrs_clt_path **)ppcpu_path, &clt_path,
2287				next))
2288			/*
2289			 * @ppcpu_path was successfully replaced with @next,
2290			 * that means that someone could also pick up the
2291			 * @sess and dereferencing it right now, so wait for
2292			 * a grace period is required.
2293			 */
2294			wait_for_grace = true;
2295	}
2296	if (wait_for_grace)
2297		synchronize_rcu();
2298
2299	mutex_unlock(&clt->paths_mutex);
2300}
2301
2302static void rtrs_clt_add_path_to_arr(struct rtrs_clt_path *clt_path)
2303{
2304	struct rtrs_clt_sess *clt = clt_path->clt;
2305
2306	mutex_lock(&clt->paths_mutex);
2307	clt->paths_num++;
2308
2309	list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2310	mutex_unlock(&clt->paths_mutex);
2311}
2312
2313static void rtrs_clt_close_work(struct work_struct *work)
2314{
2315	struct rtrs_clt_path *clt_path;
2316
2317	clt_path = container_of(work, struct rtrs_clt_path, close_work);
2318
2319	cancel_work_sync(&clt_path->err_recovery_work);
2320	cancel_delayed_work_sync(&clt_path->reconnect_dwork);
2321	rtrs_clt_stop_and_destroy_conns(clt_path);
2322	rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSED, NULL);
2323}
2324
2325static int init_conns(struct rtrs_clt_path *clt_path)
2326{
2327	unsigned int cid;
2328	int err, i;
2329
2330	/*
2331	 * On every new session connections increase reconnect counter
2332	 * to avoid clashes with previous sessions not yet closed
2333	 * sessions on a server side.
2334	 */
2335	clt_path->s.recon_cnt++;
2336
2337	/* Establish all RDMA connections  */
2338	for (cid = 0; cid < clt_path->s.con_num; cid++) {
2339		err = create_con(clt_path, cid);
2340		if (err)
2341			goto destroy;
2342
2343		err = create_cm(to_clt_con(clt_path->s.con[cid]));
2344		if (err)
2345			goto destroy;
2346	}
2347
2348	/*
2349	 * Set the cid to con_num - 1, since if we fail later, we want to stay in bounds.
2350	 */
2351	cid = clt_path->s.con_num - 1;
2352
2353	err = alloc_path_reqs(clt_path);
2354	if (err)
2355		goto destroy;
2356
2357	return 0;
2358
2359destroy:
2360	/* Make sure we do the cleanup in the order they are created */
2361	for (i = 0; i <= cid; i++) {
2362		struct rtrs_clt_con *con;
2363
2364		if (!clt_path->s.con[i])
2365			break;
2366
2367		con = to_clt_con(clt_path->s.con[i]);
2368		if (con->c.cm_id) {
2369			stop_cm(con);
2370			mutex_lock(&con->con_mutex);
2371			destroy_con_cq_qp(con);
2372			mutex_unlock(&con->con_mutex);
2373			destroy_cm(con);
2374		}
2375		destroy_con(con);
2376	}
2377	/*
2378	 * If we've never taken async path and got an error, say,
2379	 * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2380	 * manually to keep reconnecting.
2381	 */
2382	rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2383
2384	return err;
2385}
2386
2387static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2388{
2389	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2390	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2391	struct rtrs_iu *iu;
2392
2393	iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2394	rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2395
2396	if (wc->status != IB_WC_SUCCESS) {
2397		rtrs_err(clt_path->clt, "Path info request send failed: %s\n",
2398			  ib_wc_status_msg(wc->status));
2399		rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2400		return;
2401	}
2402
2403	rtrs_clt_update_wc_stats(con);
2404}
2405
2406static int process_info_rsp(struct rtrs_clt_path *clt_path,
2407			    const struct rtrs_msg_info_rsp *msg)
2408{
2409	unsigned int sg_cnt, total_len;
2410	int i, sgi;
2411
2412	sg_cnt = le16_to_cpu(msg->sg_cnt);
2413	if (!sg_cnt || (clt_path->queue_depth % sg_cnt)) {
2414		rtrs_err(clt_path->clt,
2415			  "Incorrect sg_cnt %d, is not multiple\n",
2416			  sg_cnt);
2417		return -EINVAL;
2418	}
2419
2420	/*
2421	 * Check if IB immediate data size is enough to hold the mem_id and
2422	 * the offset inside the memory chunk.
2423	 */
2424	if ((ilog2(sg_cnt - 1) + 1) + (ilog2(clt_path->chunk_size - 1) + 1) >
2425	    MAX_IMM_PAYL_BITS) {
2426		rtrs_err(clt_path->clt,
2427			  "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2428			  MAX_IMM_PAYL_BITS, sg_cnt, clt_path->chunk_size);
2429		return -EINVAL;
2430	}
2431	total_len = 0;
2432	for (sgi = 0, i = 0; sgi < sg_cnt && i < clt_path->queue_depth; sgi++) {
2433		const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2434		u32 len, rkey;
2435		u64 addr;
2436
2437		addr = le64_to_cpu(desc->addr);
2438		rkey = le32_to_cpu(desc->key);
2439		len  = le32_to_cpu(desc->len);
2440
2441		total_len += len;
2442
2443		if (!len || (len % clt_path->chunk_size)) {
2444			rtrs_err(clt_path->clt, "Incorrect [%d].len %d\n",
2445				  sgi,
2446				  len);
2447			return -EINVAL;
2448		}
2449		for ( ; len && i < clt_path->queue_depth; i++) {
2450			clt_path->rbufs[i].addr = addr;
2451			clt_path->rbufs[i].rkey = rkey;
2452
2453			len  -= clt_path->chunk_size;
2454			addr += clt_path->chunk_size;
2455		}
2456	}
2457	/* Sanity check */
2458	if (sgi != sg_cnt || i != clt_path->queue_depth) {
2459		rtrs_err(clt_path->clt,
2460			 "Incorrect sg vector, not fully mapped\n");
2461		return -EINVAL;
2462	}
2463	if (total_len != clt_path->chunk_size * clt_path->queue_depth) {
2464		rtrs_err(clt_path->clt, "Incorrect total_len %d\n", total_len);
2465		return -EINVAL;
2466	}
2467
2468	return 0;
2469}
2470
2471static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2472{
2473	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2474	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2475	struct rtrs_msg_info_rsp *msg;
2476	enum rtrs_clt_state state;
2477	struct rtrs_iu *iu;
2478	size_t rx_sz;
2479	int err;
2480
2481	state = RTRS_CLT_CONNECTING_ERR;
2482
2483	WARN_ON(con->c.cid);
2484	iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2485	if (wc->status != IB_WC_SUCCESS) {
2486		rtrs_err(clt_path->clt, "Path info response recv failed: %s\n",
2487			  ib_wc_status_msg(wc->status));
2488		goto out;
2489	}
2490	WARN_ON(wc->opcode != IB_WC_RECV);
2491
2492	if (wc->byte_len < sizeof(*msg)) {
2493		rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2494			  wc->byte_len);
2495		goto out;
2496	}
2497	ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
2498				   iu->size, DMA_FROM_DEVICE);
2499	msg = iu->buf;
2500	if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP) {
2501		rtrs_err(clt_path->clt, "Path info response is malformed: type %d\n",
2502			  le16_to_cpu(msg->type));
2503		goto out;
2504	}
2505	rx_sz  = sizeof(*msg);
2506	rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2507	if (wc->byte_len < rx_sz) {
2508		rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2509			  wc->byte_len);
2510		goto out;
2511	}
2512	err = process_info_rsp(clt_path, msg);
2513	if (err)
2514		goto out;
2515
2516	err = post_recv_path(clt_path);
2517	if (err)
2518		goto out;
2519
2520	state = RTRS_CLT_CONNECTED;
2521
2522out:
2523	rtrs_clt_update_wc_stats(con);
2524	rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2525	rtrs_clt_change_state_get_old(clt_path, state, NULL);
2526}
2527
2528static int rtrs_send_path_info(struct rtrs_clt_path *clt_path)
2529{
2530	struct rtrs_clt_con *usr_con = to_clt_con(clt_path->s.con[0]);
2531	struct rtrs_msg_info_req *msg;
2532	struct rtrs_iu *tx_iu, *rx_iu;
2533	size_t rx_sz;
2534	int err;
2535
2536	rx_sz  = sizeof(struct rtrs_msg_info_rsp);
2537	rx_sz += sizeof(struct rtrs_sg_desc) * clt_path->queue_depth;
2538
2539	tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2540			       clt_path->s.dev->ib_dev, DMA_TO_DEVICE,
2541			       rtrs_clt_info_req_done);
2542	rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, clt_path->s.dev->ib_dev,
2543			       DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2544	if (!tx_iu || !rx_iu) {
2545		err = -ENOMEM;
2546		goto out;
2547	}
2548	/* Prepare for getting info response */
2549	err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2550	if (err) {
2551		rtrs_err(clt_path->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2552		goto out;
2553	}
2554	rx_iu = NULL;
2555
2556	msg = tx_iu->buf;
2557	msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2558	memcpy(msg->pathname, clt_path->s.sessname, sizeof(msg->pathname));
2559
2560	ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
2561				      tx_iu->dma_addr,
2562				      tx_iu->size, DMA_TO_DEVICE);
2563
2564	/* Send info request */
2565	err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2566	if (err) {
2567		rtrs_err(clt_path->clt, "rtrs_iu_post_send(), err: %d\n", err);
2568		goto out;
2569	}
2570	tx_iu = NULL;
2571
2572	/* Wait for state change */
2573	wait_event_interruptible_timeout(clt_path->state_wq,
2574					 clt_path->state != RTRS_CLT_CONNECTING,
2575					 msecs_to_jiffies(
2576						 RTRS_CONNECT_TIMEOUT_MS));
2577	if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) {
2578		if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTING_ERR)
2579			err = -ECONNRESET;
2580		else
2581			err = -ETIMEDOUT;
2582	}
2583
2584out:
2585	if (tx_iu)
2586		rtrs_iu_free(tx_iu, clt_path->s.dev->ib_dev, 1);
2587	if (rx_iu)
2588		rtrs_iu_free(rx_iu, clt_path->s.dev->ib_dev, 1);
2589	if (err)
2590		/* If we've never taken async path because of malloc problems */
2591		rtrs_clt_change_state_get_old(clt_path,
2592					      RTRS_CLT_CONNECTING_ERR, NULL);
2593
2594	return err;
2595}
2596
2597/**
2598 * init_path() - establishes all path connections and does handshake
2599 * @clt_path: client path.
2600 * In case of error full close or reconnect procedure should be taken,
2601 * because reconnect or close async works can be started.
2602 */
2603static int init_path(struct rtrs_clt_path *clt_path)
2604{
2605	int err;
2606	char str[NAME_MAX];
2607	struct rtrs_addr path = {
2608		.src = &clt_path->s.src_addr,
2609		.dst = &clt_path->s.dst_addr,
2610	};
2611
2612	rtrs_addr_to_str(&path, str, sizeof(str));
2613
2614	mutex_lock(&clt_path->init_mutex);
2615	err = init_conns(clt_path);
2616	if (err) {
2617		rtrs_err(clt_path->clt,
2618			 "init_conns() failed: err=%d path=%s [%s:%u]\n", err,
2619			 str, clt_path->hca_name, clt_path->hca_port);
2620		goto out;
2621	}
2622	err = rtrs_send_path_info(clt_path);
2623	if (err) {
2624		rtrs_err(clt_path->clt,
2625			 "rtrs_send_path_info() failed: err=%d path=%s [%s:%u]\n",
2626			 err, str, clt_path->hca_name, clt_path->hca_port);
2627		goto out;
2628	}
2629	rtrs_clt_path_up(clt_path);
2630	rtrs_start_hb(&clt_path->s);
2631out:
2632	mutex_unlock(&clt_path->init_mutex);
2633
2634	return err;
2635}
2636
2637static void rtrs_clt_reconnect_work(struct work_struct *work)
2638{
2639	struct rtrs_clt_path *clt_path;
2640	struct rtrs_clt_sess *clt;
2641	int err;
2642
2643	clt_path = container_of(to_delayed_work(work), struct rtrs_clt_path,
2644				reconnect_dwork);
2645	clt = clt_path->clt;
2646
2647	trace_rtrs_clt_reconnect_work(clt_path);
2648
2649	if (READ_ONCE(clt_path->state) != RTRS_CLT_RECONNECTING)
2650		return;
2651
2652	if (clt_path->reconnect_attempts >= clt->max_reconnect_attempts) {
2653		/* Close a path completely if max attempts is reached */
2654		rtrs_clt_close_conns(clt_path, false);
2655		return;
2656	}
2657	clt_path->reconnect_attempts++;
2658
2659	msleep(RTRS_RECONNECT_BACKOFF);
2660	if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING, NULL)) {
2661		err = init_path(clt_path);
2662		if (err)
2663			goto reconnect_again;
2664	}
2665
2666	return;
2667
2668reconnect_again:
2669	if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_RECONNECTING, NULL)) {
2670		clt_path->stats->reconnects.fail_cnt++;
2671		queue_work(rtrs_wq, &clt_path->err_recovery_work);
2672	}
2673}
2674
2675static void rtrs_clt_dev_release(struct device *dev)
2676{
2677	struct rtrs_clt_sess *clt = container_of(dev, struct rtrs_clt_sess,
2678						 dev);
2679
2680	mutex_destroy(&clt->paths_ev_mutex);
2681	mutex_destroy(&clt->paths_mutex);
2682	kfree(clt);
2683}
2684
2685static struct rtrs_clt_sess *alloc_clt(const char *sessname, size_t paths_num,
2686				  u16 port, size_t pdu_sz, void *priv,
2687				  void	(*link_ev)(void *priv,
2688						   enum rtrs_clt_link_ev ev),
2689				  unsigned int reconnect_delay_sec,
2690				  unsigned int max_reconnect_attempts)
2691{
2692	struct rtrs_clt_sess *clt;
2693	int err;
2694
2695	if (!paths_num || paths_num > MAX_PATHS_NUM)
2696		return ERR_PTR(-EINVAL);
2697
2698	if (strlen(sessname) >= sizeof(clt->sessname))
2699		return ERR_PTR(-EINVAL);
2700
2701	clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2702	if (!clt)
2703		return ERR_PTR(-ENOMEM);
2704
2705	clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2706	if (!clt->pcpu_path) {
2707		kfree(clt);
2708		return ERR_PTR(-ENOMEM);
2709	}
2710
2711	clt->dev.class = &rtrs_clt_dev_class;
2712	clt->dev.release = rtrs_clt_dev_release;
2713	uuid_gen(&clt->paths_uuid);
2714	INIT_LIST_HEAD_RCU(&clt->paths_list);
2715	clt->paths_num = paths_num;
2716	clt->paths_up = MAX_PATHS_NUM;
2717	clt->port = port;
2718	clt->pdu_sz = pdu_sz;
2719	clt->max_segments = RTRS_MAX_SEGMENTS;
2720	clt->reconnect_delay_sec = reconnect_delay_sec;
2721	clt->max_reconnect_attempts = max_reconnect_attempts;
2722	clt->priv = priv;
2723	clt->link_ev = link_ev;
2724	clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2725	strscpy(clt->sessname, sessname, sizeof(clt->sessname));
2726	init_waitqueue_head(&clt->permits_wait);
2727	mutex_init(&clt->paths_ev_mutex);
2728	mutex_init(&clt->paths_mutex);
2729	device_initialize(&clt->dev);
2730
2731	err = dev_set_name(&clt->dev, "%s", sessname);
2732	if (err)
2733		goto err_put;
2734
2735	/*
2736	 * Suppress user space notification until
2737	 * sysfs files are created
2738	 */
2739	dev_set_uevent_suppress(&clt->dev, true);
2740	err = device_add(&clt->dev);
2741	if (err)
2742		goto err_put;
2743
2744	clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2745	if (!clt->kobj_paths) {
2746		err = -ENOMEM;
2747		goto err_del;
2748	}
2749	err = rtrs_clt_create_sysfs_root_files(clt);
2750	if (err) {
2751		kobject_del(clt->kobj_paths);
2752		kobject_put(clt->kobj_paths);
2753		goto err_del;
2754	}
2755	dev_set_uevent_suppress(&clt->dev, false);
2756	kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2757
2758	return clt;
2759err_del:
2760	device_del(&clt->dev);
2761err_put:
2762	free_percpu(clt->pcpu_path);
2763	put_device(&clt->dev);
2764	return ERR_PTR(err);
2765}
2766
2767static void free_clt(struct rtrs_clt_sess *clt)
2768{
2769	free_percpu(clt->pcpu_path);
2770
2771	/*
2772	 * release callback will free clt and destroy mutexes in last put
2773	 */
2774	device_unregister(&clt->dev);
2775}
2776
2777/**
2778 * rtrs_clt_open() - Open a path to an RTRS server
2779 * @ops: holds the link event callback and the private pointer.
2780 * @pathname: name of the path to an RTRS server
2781 * @paths: Paths to be established defined by their src and dst addresses
2782 * @paths_num: Number of elements in the @paths array
2783 * @port: port to be used by the RTRS session
2784 * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2785 * @reconnect_delay_sec: time between reconnect tries
2786 * @max_reconnect_attempts: Number of times to reconnect on error before giving
2787 *			    up, 0 for * disabled, -1 for forever
2788 * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
2789 *
2790 * Starts session establishment with the rtrs_server. The function can block
2791 * up to ~2000ms before it returns.
2792 *
2793 * Return a valid pointer on success otherwise PTR_ERR.
2794 */
2795struct rtrs_clt_sess *rtrs_clt_open(struct rtrs_clt_ops *ops,
2796				 const char *pathname,
2797				 const struct rtrs_addr *paths,
2798				 size_t paths_num, u16 port,
2799				 size_t pdu_sz, u8 reconnect_delay_sec,
2800				 s16 max_reconnect_attempts, u32 nr_poll_queues)
2801{
2802	struct rtrs_clt_path *clt_path, *tmp;
2803	struct rtrs_clt_sess *clt;
2804	int err, i;
2805
2806	if (strchr(pathname, '/') || strchr(pathname, '.')) {
2807		pr_err("pathname cannot contain / and .\n");
2808		err = -EINVAL;
2809		goto out;
2810	}
2811
2812	clt = alloc_clt(pathname, paths_num, port, pdu_sz, ops->priv,
2813			ops->link_ev,
2814			reconnect_delay_sec,
2815			max_reconnect_attempts);
2816	if (IS_ERR(clt)) {
2817		err = PTR_ERR(clt);
2818		goto out;
2819	}
2820	for (i = 0; i < paths_num; i++) {
2821		struct rtrs_clt_path *clt_path;
2822
2823		clt_path = alloc_path(clt, &paths[i], nr_cpu_ids,
2824				  nr_poll_queues);
2825		if (IS_ERR(clt_path)) {
2826			err = PTR_ERR(clt_path);
2827			goto close_all_path;
2828		}
2829		if (!i)
2830			clt_path->for_new_clt = 1;
2831		list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2832
2833		err = init_path(clt_path);
2834		if (err) {
2835			list_del_rcu(&clt_path->s.entry);
2836			rtrs_clt_close_conns(clt_path, true);
2837			free_percpu(clt_path->stats->pcpu_stats);
2838			kfree(clt_path->stats);
2839			free_path(clt_path);
2840			goto close_all_path;
2841		}
2842
2843		err = rtrs_clt_create_path_files(clt_path);
2844		if (err) {
2845			list_del_rcu(&clt_path->s.entry);
2846			rtrs_clt_close_conns(clt_path, true);
2847			free_percpu(clt_path->stats->pcpu_stats);
2848			kfree(clt_path->stats);
2849			free_path(clt_path);
2850			goto close_all_path;
2851		}
2852	}
2853	err = alloc_permits(clt);
2854	if (err)
2855		goto close_all_path;
2856
2857	return clt;
2858
2859close_all_path:
2860	list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2861		rtrs_clt_destroy_path_files(clt_path, NULL);
2862		rtrs_clt_close_conns(clt_path, true);
2863		kobject_put(&clt_path->kobj);
2864	}
2865	rtrs_clt_destroy_sysfs_root(clt);
2866	free_clt(clt);
2867
2868out:
2869	return ERR_PTR(err);
2870}
2871EXPORT_SYMBOL(rtrs_clt_open);
2872
2873/**
2874 * rtrs_clt_close() - Close a path
2875 * @clt: Session handle. Session is freed upon return.
2876 */
2877void rtrs_clt_close(struct rtrs_clt_sess *clt)
2878{
2879	struct rtrs_clt_path *clt_path, *tmp;
2880
2881	/* Firstly forbid sysfs access */
2882	rtrs_clt_destroy_sysfs_root(clt);
2883
2884	/* Now it is safe to iterate over all paths without locks */
2885	list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2886		rtrs_clt_close_conns(clt_path, true);
2887		rtrs_clt_destroy_path_files(clt_path, NULL);
2888		kobject_put(&clt_path->kobj);
2889	}
2890	free_permits(clt);
2891	free_clt(clt);
2892}
2893EXPORT_SYMBOL(rtrs_clt_close);
2894
2895int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_path *clt_path)
2896{
2897	enum rtrs_clt_state old_state;
2898	int err = -EBUSY;
2899	bool changed;
2900
2901	changed = rtrs_clt_change_state_get_old(clt_path,
2902						 RTRS_CLT_RECONNECTING,
2903						 &old_state);
2904	if (changed) {
2905		clt_path->reconnect_attempts = 0;
2906		rtrs_clt_stop_and_destroy_conns(clt_path);
2907		queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork, 0);
2908	}
2909	if (changed || old_state == RTRS_CLT_RECONNECTING) {
2910		/*
2911		 * flush_delayed_work() queues pending work for immediate
2912		 * execution, so do the flush if we have queued something
2913		 * right now or work is pending.
2914		 */
2915		flush_delayed_work(&clt_path->reconnect_dwork);
2916		err = (READ_ONCE(clt_path->state) ==
2917		       RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2918	}
2919
2920	return err;
2921}
2922
2923int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_path *clt_path,
2924				     const struct attribute *sysfs_self)
2925{
2926	enum rtrs_clt_state old_state;
2927	bool changed;
2928
2929	/*
2930	 * Continue stopping path till state was changed to DEAD or
2931	 * state was observed as DEAD:
2932	 * 1. State was changed to DEAD - we were fast and nobody
2933	 *    invoked rtrs_clt_reconnect(), which can again start
2934	 *    reconnecting.
2935	 * 2. State was observed as DEAD - we have someone in parallel
2936	 *    removing the path.
2937	 */
2938	do {
2939		rtrs_clt_close_conns(clt_path, true);
2940		changed = rtrs_clt_change_state_get_old(clt_path,
2941							RTRS_CLT_DEAD,
2942							&old_state);
2943	} while (!changed && old_state != RTRS_CLT_DEAD);
2944
2945	if (changed) {
2946		rtrs_clt_remove_path_from_arr(clt_path);
2947		rtrs_clt_destroy_path_files(clt_path, sysfs_self);
2948		kobject_put(&clt_path->kobj);
2949	}
2950
2951	return 0;
2952}
2953
2954void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt_sess *clt, int value)
2955{
2956	clt->max_reconnect_attempts = (unsigned int)value;
2957}
2958
2959int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt_sess *clt)
2960{
2961	return (int)clt->max_reconnect_attempts;
2962}
2963
2964/**
2965 * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2966 *
2967 * @dir:	READ/WRITE
2968 * @ops:	callback function to be called as confirmation, and the pointer.
2969 * @clt:	Session
2970 * @permit:	Preallocated permit
2971 * @vec:	Message that is sent to server together with the request.
2972 *		Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2973 *		Since the msg is copied internally it can be allocated on stack.
2974 * @nr:		Number of elements in @vec.
2975 * @data_len:	length of data sent to/from server
2976 * @sg:		Pages to be sent/received to/from server.
2977 * @sg_cnt:	Number of elements in the @sg
2978 *
2979 * Return:
2980 * 0:		Success
2981 * <0:		Error
2982 *
2983 * On dir=READ rtrs client will request a data transfer from Server to client.
2984 * The data that the server will respond with will be stored in @sg when
2985 * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2986 * On dir=WRITE rtrs client will rdma write data in sg to server side.
2987 */
2988int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2989		     struct rtrs_clt_sess *clt, struct rtrs_permit *permit,
2990		     const struct kvec *vec, size_t nr, size_t data_len,
2991		     struct scatterlist *sg, unsigned int sg_cnt)
2992{
2993	struct rtrs_clt_io_req *req;
2994	struct rtrs_clt_path *clt_path;
2995
2996	enum dma_data_direction dma_dir;
2997	int err = -ECONNABORTED, i;
2998	size_t usr_len, hdr_len;
2999	struct path_it it;
3000
3001	/* Get kvec length */
3002	for (i = 0, usr_len = 0; i < nr; i++)
3003		usr_len += vec[i].iov_len;
3004
3005	if (dir == READ) {
3006		hdr_len = sizeof(struct rtrs_msg_rdma_read) +
3007			  sg_cnt * sizeof(struct rtrs_sg_desc);
3008		dma_dir = DMA_FROM_DEVICE;
3009	} else {
3010		hdr_len = sizeof(struct rtrs_msg_rdma_write);
3011		dma_dir = DMA_TO_DEVICE;
3012	}
3013
3014	rcu_read_lock();
3015	for (path_it_init(&it, clt);
3016	     (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3017		if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3018			continue;
3019
3020		if (usr_len + hdr_len > clt_path->max_hdr_size) {
3021			rtrs_wrn_rl(clt_path->clt,
3022				     "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
3023				     dir == READ ? "Read" : "Write",
3024				     usr_len, hdr_len, clt_path->max_hdr_size);
3025			err = -EMSGSIZE;
3026			break;
3027		}
3028		req = rtrs_clt_get_req(clt_path, ops->conf_fn, permit, ops->priv,
3029				       vec, usr_len, sg, sg_cnt, data_len,
3030				       dma_dir);
3031		if (dir == READ)
3032			err = rtrs_clt_read_req(req);
3033		else
3034			err = rtrs_clt_write_req(req);
3035		if (err) {
3036			req->in_use = false;
3037			continue;
3038		}
3039		/* Success path */
3040		break;
3041	}
3042	path_it_deinit(&it);
3043	rcu_read_unlock();
3044
3045	return err;
3046}
3047EXPORT_SYMBOL(rtrs_clt_request);
3048
3049int rtrs_clt_rdma_cq_direct(struct rtrs_clt_sess *clt, unsigned int index)
3050{
3051	/* If no path, return -1 for block layer not to try again */
3052	int cnt = -1;
3053	struct rtrs_con *con;
3054	struct rtrs_clt_path *clt_path;
3055	struct path_it it;
3056
3057	rcu_read_lock();
3058	for (path_it_init(&it, clt);
3059	     (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3060		if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3061			continue;
3062
3063		con = clt_path->s.con[index + 1];
3064		cnt = ib_process_cq_direct(con->cq, -1);
3065		if (cnt)
3066			break;
3067	}
3068	path_it_deinit(&it);
3069	rcu_read_unlock();
3070
3071	return cnt;
3072}
3073EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
3074
3075/**
3076 * rtrs_clt_query() - queries RTRS session attributes
3077 *@clt: session pointer
3078 *@attr: query results for session attributes.
3079 * Returns:
3080 *    0 on success
3081 *    -ECOMM		no connection to the server
3082 */
3083int rtrs_clt_query(struct rtrs_clt_sess *clt, struct rtrs_attrs *attr)
3084{
3085	if (!rtrs_clt_is_connected(clt))
3086		return -ECOMM;
3087
3088	attr->queue_depth      = clt->queue_depth;
3089	attr->max_segments     = clt->max_segments;
3090	/* Cap max_io_size to min of remote buffer size and the fr pages */
3091	attr->max_io_size = min_t(int, clt->max_io_size,
3092				  clt->max_segments * SZ_4K);
3093
3094	return 0;
3095}
3096EXPORT_SYMBOL(rtrs_clt_query);
3097
3098int rtrs_clt_create_path_from_sysfs(struct rtrs_clt_sess *clt,
3099				     struct rtrs_addr *addr)
3100{
3101	struct rtrs_clt_path *clt_path;
3102	int err;
3103
3104	clt_path = alloc_path(clt, addr, nr_cpu_ids, 0);
3105	if (IS_ERR(clt_path))
3106		return PTR_ERR(clt_path);
3107
3108	mutex_lock(&clt->paths_mutex);
3109	if (clt->paths_num == 0) {
3110		/*
3111		 * When all the paths are removed for a session,
3112		 * the addition of the first path is like a new session for
3113		 * the storage server
3114		 */
3115		clt_path->for_new_clt = 1;
3116	}
3117
3118	mutex_unlock(&clt->paths_mutex);
3119
3120	/*
3121	 * It is totally safe to add path in CONNECTING state: coming
3122	 * IO will never grab it.  Also it is very important to add
3123	 * path before init, since init fires LINK_CONNECTED event.
3124	 */
3125	rtrs_clt_add_path_to_arr(clt_path);
3126
3127	err = init_path(clt_path);
3128	if (err)
3129		goto close_path;
3130
3131	err = rtrs_clt_create_path_files(clt_path);
3132	if (err)
3133		goto close_path;
3134
3135	return 0;
3136
3137close_path:
3138	rtrs_clt_remove_path_from_arr(clt_path);
3139	rtrs_clt_close_conns(clt_path, true);
3140	free_percpu(clt_path->stats->pcpu_stats);
3141	kfree(clt_path->stats);
3142	free_path(clt_path);
3143
3144	return err;
3145}
3146
3147void rtrs_clt_ib_event_handler(struct ib_event_handler *handler,
3148			       struct ib_event *ibevent)
3149{
3150	pr_info("Handling event: %s (%d).\n", ib_event_msg(ibevent->event),
3151		ibevent->event);
3152}
3153
3154
3155static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
3156{
3157	INIT_IB_EVENT_HANDLER(&dev->event_handler, dev->ib_dev,
3158			      rtrs_clt_ib_event_handler);
3159	ib_register_event_handler(&dev->event_handler);
3160
3161	if (!(dev->ib_dev->attrs.device_cap_flags &
3162	      IB_DEVICE_MEM_MGT_EXTENSIONS)) {
3163		pr_err("Memory registrations not supported.\n");
3164		return -ENOTSUPP;
3165	}
3166
3167	return 0;
3168}
3169
3170static void rtrs_clt_ib_dev_deinit(struct rtrs_ib_dev *dev)
3171{
3172	ib_unregister_event_handler(&dev->event_handler);
3173}
3174
3175
3176static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
3177	.init = rtrs_clt_ib_dev_init,
3178	.deinit = rtrs_clt_ib_dev_deinit
3179};
3180
3181static int __init rtrs_client_init(void)
3182{
3183	int ret = 0;
3184
3185	rtrs_rdma_dev_pd_init(0, &dev_pd);
3186	ret = class_register(&rtrs_clt_dev_class);
3187	if (ret) {
3188		pr_err("Failed to create rtrs-client dev class\n");
3189		return ret;
3190	}
3191	rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
3192	if (!rtrs_wq) {
3193		class_unregister(&rtrs_clt_dev_class);
3194		return -ENOMEM;
3195	}
3196
3197	return 0;
3198}
3199
3200static void __exit rtrs_client_exit(void)
3201{
3202	destroy_workqueue(rtrs_wq);
3203	class_unregister(&rtrs_clt_dev_class);
3204	rtrs_rdma_dev_pd_deinit(&dev_pd);
3205}
3206
3207module_init(rtrs_client_init);
3208module_exit(rtrs_client_exit);