1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * RDMA Transport Layer
4  *
5  * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved.
6  * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved.
7  * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved.
8  */
9 
10 #undef pr_fmt
11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
12 
13 #include <linux/module.h>
14 #include <linux/rculist.h>
15 #include <linux/random.h>
16 
17 #include "rtrs-clt.h"
18 #include "rtrs-log.h"
19 
20 #define RTRS_CONNECT_TIMEOUT_MS 30000
21 /*
22  * Wait a bit before trying to reconnect after a failure
23  * in order to give server time to finish clean up which
24  * leads to "false positives" failed reconnect attempts
25  */
26 #define RTRS_RECONNECT_BACKOFF 1000
27 /*
28  * Wait for additional random time between 0 and 8 seconds
29  * before starting to reconnect to avoid clients reconnecting
30  * all at once in case of a major network outage
31  */
32 #define RTRS_RECONNECT_SEED 8
33 
34 #define FIRST_CONN 0x01
35 /* limit to 128 * 4k = 512k max IO */
36 #define RTRS_MAX_SEGMENTS          128
37 
38 MODULE_DESCRIPTION("RDMA Transport Client");
39 MODULE_LICENSE("GPL");
40 
41 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops;
42 static struct rtrs_rdma_dev_pd dev_pd = {
43 	.ops = &dev_pd_ops
44 };
45 
46 static struct workqueue_struct *rtrs_wq;
47 static struct class *rtrs_clt_dev_class;
48 
49 static inline bool rtrs_clt_is_connected(const struct rtrs_clt_sess *clt)
50 {
51 	struct rtrs_clt_path *clt_path;
52 	bool connected = false;
53 
54 	rcu_read_lock();
55 	list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry)
56 		connected |= READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED;
57 	rcu_read_unlock();
58 
59 	return connected;
60 }
61 
62 static struct rtrs_permit *
63 __rtrs_get_permit(struct rtrs_clt_sess *clt, enum rtrs_clt_con_type con_type)
64 {
65 	size_t max_depth = clt->queue_depth;
66 	struct rtrs_permit *permit;
67 	int bit;
68 
69 	/*
70 	 * Adapted from null_blk get_tag(). Callers from different cpus may
71 	 * grab the same bit, since find_first_zero_bit is not atomic.
72 	 * But then the test_and_set_bit_lock will fail for all the
73 	 * callers but one, so that they will loop again.
74 	 * This way an explicit spinlock is not required.
75 	 */
76 	do {
77 		bit = find_first_zero_bit(clt->permits_map, max_depth);
78 		if (bit >= max_depth)
79 			return NULL;
80 	} while (test_and_set_bit_lock(bit, clt->permits_map));
81 
82 	permit = get_permit(clt, bit);
83 	WARN_ON(permit->mem_id != bit);
84 	permit->cpu_id = raw_smp_processor_id();
85 	permit->con_type = con_type;
86 
87 	return permit;
88 }
89 
90 static inline void __rtrs_put_permit(struct rtrs_clt_sess *clt,
91 				      struct rtrs_permit *permit)
92 {
93 	clear_bit_unlock(permit->mem_id, clt->permits_map);
94 }
95 
96 /**
97  * rtrs_clt_get_permit() - allocates permit for future RDMA operation
98  * @clt:	Current session
99  * @con_type:	Type of connection to use with the permit
100  * @can_wait:	Wait type
101  *
102  * Description:
103  *    Allocates permit for the following RDMA operation.  Permit is used
104  *    to preallocate all resources and to propagate memory pressure
105  *    up earlier.
106  *
107  * Context:
108  *    Can sleep if @wait == RTRS_PERMIT_WAIT
109  */
110 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt_sess *clt,
111 					  enum rtrs_clt_con_type con_type,
112 					  enum wait_type can_wait)
113 {
114 	struct rtrs_permit *permit;
115 	DEFINE_WAIT(wait);
116 
117 	permit = __rtrs_get_permit(clt, con_type);
118 	if (permit || !can_wait)
119 		return permit;
120 
121 	do {
122 		prepare_to_wait(&clt->permits_wait, &wait,
123 				TASK_UNINTERRUPTIBLE);
124 		permit = __rtrs_get_permit(clt, con_type);
125 		if (permit)
126 			break;
127 
128 		io_schedule();
129 	} while (1);
130 
131 	finish_wait(&clt->permits_wait, &wait);
132 
133 	return permit;
134 }
135 EXPORT_SYMBOL(rtrs_clt_get_permit);
136 
137 /**
138  * rtrs_clt_put_permit() - puts allocated permit
139  * @clt:	Current session
140  * @permit:	Permit to be freed
141  *
142  * Context:
143  *    Does not matter
144  */
145 void rtrs_clt_put_permit(struct rtrs_clt_sess *clt,
146 			 struct rtrs_permit *permit)
147 {
148 	if (WARN_ON(!test_bit(permit->mem_id, clt->permits_map)))
149 		return;
150 
151 	__rtrs_put_permit(clt, permit);
152 
153 	/*
154 	 * rtrs_clt_get_permit() adds itself to the &clt->permits_wait list
155 	 * before calling schedule(). So if rtrs_clt_get_permit() is sleeping
156 	 * it must have added itself to &clt->permits_wait before
157 	 * __rtrs_put_permit() finished.
158 	 * Hence it is safe to guard wake_up() with a waitqueue_active() test.
159 	 */
160 	if (waitqueue_active(&clt->permits_wait))
161 		wake_up(&clt->permits_wait);
162 }
163 EXPORT_SYMBOL(rtrs_clt_put_permit);
164 
165 /**
166  * rtrs_permit_to_clt_con() - returns RDMA connection pointer by the permit
167  * @clt_path: client path pointer
168  * @permit: permit for the allocation of the RDMA buffer
169  * Note:
170  *     IO connection starts from 1.
171  *     0 connection is for user messages.
172  */
173 static
174 struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_path *clt_path,
175 					    struct rtrs_permit *permit)
176 {
177 	int id = 0;
178 
179 	if (permit->con_type == RTRS_IO_CON)
180 		id = (permit->cpu_id % (clt_path->s.irq_con_num - 1)) + 1;
181 
182 	return to_clt_con(clt_path->s.con[id]);
183 }
184 
185 /**
186  * rtrs_clt_change_state() - change the session state through session state
187  * machine.
188  *
189  * @clt_path: client path to change the state of.
190  * @new_state: state to change to.
191  *
192  * returns true if sess's state is changed to new state, otherwise return false.
193  *
194  * Locks:
195  * state_wq lock must be hold.
196  */
197 static bool rtrs_clt_change_state(struct rtrs_clt_path *clt_path,
198 				     enum rtrs_clt_state new_state)
199 {
200 	enum rtrs_clt_state old_state;
201 	bool changed = false;
202 
203 	lockdep_assert_held(&clt_path->state_wq.lock);
204 
205 	old_state = clt_path->state;
206 	switch (new_state) {
207 	case RTRS_CLT_CONNECTING:
208 		switch (old_state) {
209 		case RTRS_CLT_RECONNECTING:
210 			changed = true;
211 			fallthrough;
212 		default:
213 			break;
214 		}
215 		break;
216 	case RTRS_CLT_RECONNECTING:
217 		switch (old_state) {
218 		case RTRS_CLT_CONNECTED:
219 		case RTRS_CLT_CONNECTING_ERR:
220 		case RTRS_CLT_CLOSED:
221 			changed = true;
222 			fallthrough;
223 		default:
224 			break;
225 		}
226 		break;
227 	case RTRS_CLT_CONNECTED:
228 		switch (old_state) {
229 		case RTRS_CLT_CONNECTING:
230 			changed = true;
231 			fallthrough;
232 		default:
233 			break;
234 		}
235 		break;
236 	case RTRS_CLT_CONNECTING_ERR:
237 		switch (old_state) {
238 		case RTRS_CLT_CONNECTING:
239 			changed = true;
240 			fallthrough;
241 		default:
242 			break;
243 		}
244 		break;
245 	case RTRS_CLT_CLOSING:
246 		switch (old_state) {
247 		case RTRS_CLT_CONNECTING:
248 		case RTRS_CLT_CONNECTING_ERR:
249 		case RTRS_CLT_RECONNECTING:
250 		case RTRS_CLT_CONNECTED:
251 			changed = true;
252 			fallthrough;
253 		default:
254 			break;
255 		}
256 		break;
257 	case RTRS_CLT_CLOSED:
258 		switch (old_state) {
259 		case RTRS_CLT_CLOSING:
260 			changed = true;
261 			fallthrough;
262 		default:
263 			break;
264 		}
265 		break;
266 	case RTRS_CLT_DEAD:
267 		switch (old_state) {
268 		case RTRS_CLT_CLOSED:
269 			changed = true;
270 			fallthrough;
271 		default:
272 			break;
273 		}
274 		break;
275 	default:
276 		break;
277 	}
278 	if (changed) {
279 		clt_path->state = new_state;
280 		wake_up_locked(&clt_path->state_wq);
281 	}
282 
283 	return changed;
284 }
285 
286 static bool rtrs_clt_change_state_from_to(struct rtrs_clt_path *clt_path,
287 					   enum rtrs_clt_state old_state,
288 					   enum rtrs_clt_state new_state)
289 {
290 	bool changed = false;
291 
292 	spin_lock_irq(&clt_path->state_wq.lock);
293 	if (clt_path->state == old_state)
294 		changed = rtrs_clt_change_state(clt_path, new_state);
295 	spin_unlock_irq(&clt_path->state_wq.lock);
296 
297 	return changed;
298 }
299 
300 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path);
301 static void rtrs_rdma_error_recovery(struct rtrs_clt_con *con)
302 {
303 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
304 
305 	if (rtrs_clt_change_state_from_to(clt_path,
306 					   RTRS_CLT_CONNECTED,
307 					   RTRS_CLT_RECONNECTING)) {
308 		queue_work(rtrs_wq, &clt_path->err_recovery_work);
309 	} else {
310 		/*
311 		 * Error can happen just on establishing new connection,
312 		 * so notify waiter with error state, waiter is responsible
313 		 * for cleaning the rest and reconnect if needed.
314 		 */
315 		rtrs_clt_change_state_from_to(clt_path,
316 					       RTRS_CLT_CONNECTING,
317 					       RTRS_CLT_CONNECTING_ERR);
318 	}
319 }
320 
321 static void rtrs_clt_fast_reg_done(struct ib_cq *cq, struct ib_wc *wc)
322 {
323 	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
324 
325 	if (wc->status != IB_WC_SUCCESS) {
326 		rtrs_err(con->c.path, "Failed IB_WR_REG_MR: %s\n",
327 			  ib_wc_status_msg(wc->status));
328 		rtrs_rdma_error_recovery(con);
329 	}
330 }
331 
332 static struct ib_cqe fast_reg_cqe = {
333 	.done = rtrs_clt_fast_reg_done
334 };
335 
336 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
337 			      bool notify, bool can_wait);
338 
339 static void rtrs_clt_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
340 {
341 	struct rtrs_clt_io_req *req =
342 		container_of(wc->wr_cqe, typeof(*req), inv_cqe);
343 	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
344 
345 	if (wc->status != IB_WC_SUCCESS) {
346 		rtrs_err(con->c.path, "Failed IB_WR_LOCAL_INV: %s\n",
347 			  ib_wc_status_msg(wc->status));
348 		rtrs_rdma_error_recovery(con);
349 	}
350 	req->need_inv = false;
351 	if (req->need_inv_comp)
352 		complete(&req->inv_comp);
353 	else
354 		/* Complete request from INV callback */
355 		complete_rdma_req(req, req->inv_errno, true, false);
356 }
357 
358 static int rtrs_inv_rkey(struct rtrs_clt_io_req *req)
359 {
360 	struct rtrs_clt_con *con = req->con;
361 	struct ib_send_wr wr = {
362 		.opcode		    = IB_WR_LOCAL_INV,
363 		.wr_cqe		    = &req->inv_cqe,
364 		.send_flags	    = IB_SEND_SIGNALED,
365 		.ex.invalidate_rkey = req->mr->rkey,
366 	};
367 	req->inv_cqe.done = rtrs_clt_inv_rkey_done;
368 
369 	return ib_post_send(con->c.qp, &wr, NULL);
370 }
371 
372 static void complete_rdma_req(struct rtrs_clt_io_req *req, int errno,
373 			      bool notify, bool can_wait)
374 {
375 	struct rtrs_clt_con *con = req->con;
376 	struct rtrs_clt_path *clt_path;
377 	int err;
378 
379 	if (WARN_ON(!req->in_use))
380 		return;
381 	if (WARN_ON(!req->con))
382 		return;
383 	clt_path = to_clt_path(con->c.path);
384 
385 	if (req->sg_cnt) {
386 		if (req->dir == DMA_FROM_DEVICE && req->need_inv) {
387 			/*
388 			 * We are here to invalidate read requests
389 			 * ourselves.  In normal scenario server should
390 			 * send INV for all read requests, but
391 			 * we are here, thus two things could happen:
392 			 *
393 			 *    1.  this is failover, when errno != 0
394 			 *        and can_wait == 1,
395 			 *
396 			 *    2.  something totally bad happened and
397 			 *        server forgot to send INV, so we
398 			 *        should do that ourselves.
399 			 */
400 
401 			if (can_wait) {
402 				req->need_inv_comp = true;
403 			} else {
404 				/* This should be IO path, so always notify */
405 				WARN_ON(!notify);
406 				/* Save errno for INV callback */
407 				req->inv_errno = errno;
408 			}
409 
410 			refcount_inc(&req->ref);
411 			err = rtrs_inv_rkey(req);
412 			if (err) {
413 				rtrs_err(con->c.path, "Send INV WR key=%#x: %d\n",
414 					  req->mr->rkey, err);
415 			} else if (can_wait) {
416 				wait_for_completion(&req->inv_comp);
417 			} else {
418 				/*
419 				 * Something went wrong, so request will be
420 				 * completed from INV callback.
421 				 */
422 				WARN_ON_ONCE(1);
423 
424 				return;
425 			}
426 			if (!refcount_dec_and_test(&req->ref))
427 				return;
428 		}
429 		ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
430 				req->sg_cnt, req->dir);
431 	}
432 	if (!refcount_dec_and_test(&req->ref))
433 		return;
434 	if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
435 		atomic_dec(&clt_path->stats->inflight);
436 
437 	req->in_use = false;
438 	req->con = NULL;
439 
440 	if (errno) {
441 		rtrs_err_rl(con->c.path, "IO request failed: error=%d path=%s [%s:%u] notify=%d\n",
442 			    errno, kobject_name(&clt_path->kobj), clt_path->hca_name,
443 			    clt_path->hca_port, notify);
444 	}
445 
446 	if (notify)
447 		req->conf(req->priv, errno);
448 }
449 
450 static int rtrs_post_send_rdma(struct rtrs_clt_con *con,
451 				struct rtrs_clt_io_req *req,
452 				struct rtrs_rbuf *rbuf, u32 off,
453 				u32 imm, struct ib_send_wr *wr)
454 {
455 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
456 	enum ib_send_flags flags;
457 	struct ib_sge sge;
458 
459 	if (!req->sg_size) {
460 		rtrs_wrn(con->c.path,
461 			 "Doing RDMA Write failed, no data supplied\n");
462 		return -EINVAL;
463 	}
464 
465 	/* user data and user message in the first list element */
466 	sge.addr   = req->iu->dma_addr;
467 	sge.length = req->sg_size;
468 	sge.lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
469 
470 	/*
471 	 * From time to time we have to post signalled sends,
472 	 * or send queue will fill up and only QP reset can help.
473 	 */
474 	flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
475 			0 : IB_SEND_SIGNALED;
476 
477 	ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
478 				      req->iu->dma_addr,
479 				      req->sg_size, DMA_TO_DEVICE);
480 
481 	return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, &sge, 1,
482 					    rbuf->rkey, rbuf->addr + off,
483 					    imm, flags, wr, NULL);
484 }
485 
486 static void process_io_rsp(struct rtrs_clt_path *clt_path, u32 msg_id,
487 			   s16 errno, bool w_inval)
488 {
489 	struct rtrs_clt_io_req *req;
490 
491 	if (WARN_ON(msg_id >= clt_path->queue_depth))
492 		return;
493 
494 	req = &clt_path->reqs[msg_id];
495 	/* Drop need_inv if server responded with send with invalidation */
496 	req->need_inv &= !w_inval;
497 	complete_rdma_req(req, errno, true, false);
498 }
499 
500 static void rtrs_clt_recv_done(struct rtrs_clt_con *con, struct ib_wc *wc)
501 {
502 	struct rtrs_iu *iu;
503 	int err;
504 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
505 
506 	WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
507 	iu = container_of(wc->wr_cqe, struct rtrs_iu,
508 			  cqe);
509 	err = rtrs_iu_post_recv(&con->c, iu);
510 	if (err) {
511 		rtrs_err(con->c.path, "post iu failed %d\n", err);
512 		rtrs_rdma_error_recovery(con);
513 	}
514 }
515 
516 static void rtrs_clt_rkey_rsp_done(struct rtrs_clt_con *con, struct ib_wc *wc)
517 {
518 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
519 	struct rtrs_msg_rkey_rsp *msg;
520 	u32 imm_type, imm_payload;
521 	bool w_inval = false;
522 	struct rtrs_iu *iu;
523 	u32 buf_id;
524 	int err;
525 
526 	WARN_ON((clt_path->flags & RTRS_MSG_NEW_RKEY_F) == 0);
527 
528 	iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
529 
530 	if (wc->byte_len < sizeof(*msg)) {
531 		rtrs_err(con->c.path, "rkey response is malformed: size %d\n",
532 			  wc->byte_len);
533 		goto out;
534 	}
535 	ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
536 				   iu->size, DMA_FROM_DEVICE);
537 	msg = iu->buf;
538 	if (le16_to_cpu(msg->type) != RTRS_MSG_RKEY_RSP) {
539 		rtrs_err(clt_path->clt,
540 			  "rkey response is malformed: type %d\n",
541 			  le16_to_cpu(msg->type));
542 		goto out;
543 	}
544 	buf_id = le16_to_cpu(msg->buf_id);
545 	if (WARN_ON(buf_id >= clt_path->queue_depth))
546 		goto out;
547 
548 	rtrs_from_imm(be32_to_cpu(wc->ex.imm_data), &imm_type, &imm_payload);
549 	if (imm_type == RTRS_IO_RSP_IMM ||
550 	    imm_type == RTRS_IO_RSP_W_INV_IMM) {
551 		u32 msg_id;
552 
553 		w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
554 		rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
555 
556 		if (WARN_ON(buf_id != msg_id))
557 			goto out;
558 		clt_path->rbufs[buf_id].rkey = le32_to_cpu(msg->rkey);
559 		process_io_rsp(clt_path, msg_id, err, w_inval);
560 	}
561 	ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev, iu->dma_addr,
562 				      iu->size, DMA_FROM_DEVICE);
563 	return rtrs_clt_recv_done(con, wc);
564 out:
565 	rtrs_rdma_error_recovery(con);
566 }
567 
568 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
569 
570 static struct ib_cqe io_comp_cqe = {
571 	.done = rtrs_clt_rdma_done
572 };
573 
574 /*
575  * Post x2 empty WRs: first is for this RDMA with IMM,
576  * second is for RECV with INV, which happened earlier.
577  */
578 static int rtrs_post_recv_empty_x2(struct rtrs_con *con, struct ib_cqe *cqe)
579 {
580 	struct ib_recv_wr wr_arr[2], *wr;
581 	int i;
582 
583 	memset(wr_arr, 0, sizeof(wr_arr));
584 	for (i = 0; i < ARRAY_SIZE(wr_arr); i++) {
585 		wr = &wr_arr[i];
586 		wr->wr_cqe  = cqe;
587 		if (i)
588 			/* Chain backwards */
589 			wr->next = &wr_arr[i - 1];
590 	}
591 
592 	return ib_post_recv(con->qp, wr, NULL);
593 }
594 
595 static void rtrs_clt_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
596 {
597 	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
598 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
599 	u32 imm_type, imm_payload;
600 	bool w_inval = false;
601 	int err;
602 
603 	if (wc->status != IB_WC_SUCCESS) {
604 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
605 			rtrs_err(clt_path->clt, "RDMA failed: %s\n",
606 				  ib_wc_status_msg(wc->status));
607 			rtrs_rdma_error_recovery(con);
608 		}
609 		return;
610 	}
611 	rtrs_clt_update_wc_stats(con);
612 
613 	switch (wc->opcode) {
614 	case IB_WC_RECV_RDMA_WITH_IMM:
615 		/*
616 		 * post_recv() RDMA write completions of IO reqs (read/write)
617 		 * and hb
618 		 */
619 		if (WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done))
620 			return;
621 		rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
622 			       &imm_type, &imm_payload);
623 		if (imm_type == RTRS_IO_RSP_IMM ||
624 		    imm_type == RTRS_IO_RSP_W_INV_IMM) {
625 			u32 msg_id;
626 
627 			w_inval = (imm_type == RTRS_IO_RSP_W_INV_IMM);
628 			rtrs_from_io_rsp_imm(imm_payload, &msg_id, &err);
629 
630 			process_io_rsp(clt_path, msg_id, err, w_inval);
631 		} else if (imm_type == RTRS_HB_MSG_IMM) {
632 			WARN_ON(con->c.cid);
633 			rtrs_send_hb_ack(&clt_path->s);
634 			if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
635 				return  rtrs_clt_recv_done(con, wc);
636 		} else if (imm_type == RTRS_HB_ACK_IMM) {
637 			WARN_ON(con->c.cid);
638 			clt_path->s.hb_missed_cnt = 0;
639 			clt_path->s.hb_cur_latency =
640 				ktime_sub(ktime_get(), clt_path->s.hb_last_sent);
641 			if (clt_path->flags & RTRS_MSG_NEW_RKEY_F)
642 				return  rtrs_clt_recv_done(con, wc);
643 		} else {
644 			rtrs_wrn(con->c.path, "Unknown IMM type %u\n",
645 				  imm_type);
646 		}
647 		if (w_inval)
648 			/*
649 			 * Post x2 empty WRs: first is for this RDMA with IMM,
650 			 * second is for RECV with INV, which happened earlier.
651 			 */
652 			err = rtrs_post_recv_empty_x2(&con->c, &io_comp_cqe);
653 		else
654 			err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
655 		if (err) {
656 			rtrs_err(con->c.path, "rtrs_post_recv_empty(): %d\n",
657 				  err);
658 			rtrs_rdma_error_recovery(con);
659 		}
660 		break;
661 	case IB_WC_RECV:
662 		/*
663 		 * Key invalidations from server side
664 		 */
665 		WARN_ON(!(wc->wc_flags & IB_WC_WITH_INVALIDATE ||
666 			  wc->wc_flags & IB_WC_WITH_IMM));
667 		WARN_ON(wc->wr_cqe->done != rtrs_clt_rdma_done);
668 		if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
669 			if (wc->wc_flags & IB_WC_WITH_INVALIDATE)
670 				return  rtrs_clt_recv_done(con, wc);
671 
672 			return  rtrs_clt_rkey_rsp_done(con, wc);
673 		}
674 		break;
675 	case IB_WC_RDMA_WRITE:
676 		/*
677 		 * post_send() RDMA write completions of IO reqs (read/write)
678 		 * and hb.
679 		 */
680 		break;
681 
682 	default:
683 		rtrs_wrn(clt_path->clt, "Unexpected WC type: %d\n", wc->opcode);
684 		return;
685 	}
686 }
687 
688 static int post_recv_io(struct rtrs_clt_con *con, size_t q_size)
689 {
690 	int err, i;
691 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
692 
693 	for (i = 0; i < q_size; i++) {
694 		if (clt_path->flags & RTRS_MSG_NEW_RKEY_F) {
695 			struct rtrs_iu *iu = &con->rsp_ius[i];
696 
697 			err = rtrs_iu_post_recv(&con->c, iu);
698 		} else {
699 			err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
700 		}
701 		if (err)
702 			return err;
703 	}
704 
705 	return 0;
706 }
707 
708 static int post_recv_path(struct rtrs_clt_path *clt_path)
709 {
710 	size_t q_size = 0;
711 	int err, cid;
712 
713 	for (cid = 0; cid < clt_path->s.con_num; cid++) {
714 		if (cid == 0)
715 			q_size = SERVICE_CON_QUEUE_DEPTH;
716 		else
717 			q_size = clt_path->queue_depth;
718 
719 		/*
720 		 * x2 for RDMA read responses + FR key invalidations,
721 		 * RDMA writes do not require any FR registrations.
722 		 */
723 		q_size *= 2;
724 
725 		err = post_recv_io(to_clt_con(clt_path->s.con[cid]), q_size);
726 		if (err) {
727 			rtrs_err(clt_path->clt, "post_recv_io(), err: %d\n",
728 				 err);
729 			return err;
730 		}
731 	}
732 
733 	return 0;
734 }
735 
736 struct path_it {
737 	int i;
738 	struct list_head skip_list;
739 	struct rtrs_clt_sess *clt;
740 	struct rtrs_clt_path *(*next_path)(struct path_it *it);
741 };
742 
743 /*
744  * rtrs_clt_get_next_path_or_null - get clt path from the list or return NULL
745  * @head:	the head for the list.
746  * @clt_path:	The element to take the next clt_path from.
747  *
748  * Next clt path returned in round-robin fashion, i.e. head will be skipped,
749  * but if list is observed as empty, NULL will be returned.
750  *
751  * This function may safely run concurrently with the _rcu list-mutation
752  * primitives such as list_add_rcu() as long as it's guarded by rcu_read_lock().
753  */
754 static inline struct rtrs_clt_path *
755 rtrs_clt_get_next_path_or_null(struct list_head *head, struct rtrs_clt_path *clt_path)
756 {
757 	return list_next_or_null_rcu(head, &clt_path->s.entry, typeof(*clt_path), s.entry) ?:
758 				     list_next_or_null_rcu(head,
759 							   READ_ONCE((&clt_path->s.entry)->next),
760 							   typeof(*clt_path), s.entry);
761 }
762 
763 /**
764  * get_next_path_rr() - Returns path in round-robin fashion.
765  * @it:	the path pointer
766  *
767  * Related to @MP_POLICY_RR
768  *
769  * Locks:
770  *    rcu_read_lock() must be hold.
771  */
772 static struct rtrs_clt_path *get_next_path_rr(struct path_it *it)
773 {
774 	struct rtrs_clt_path __rcu **ppcpu_path;
775 	struct rtrs_clt_path *path;
776 	struct rtrs_clt_sess *clt;
777 
778 	clt = it->clt;
779 
780 	/*
781 	 * Here we use two RCU objects: @paths_list and @pcpu_path
782 	 * pointer.  See rtrs_clt_remove_path_from_arr() for details
783 	 * how that is handled.
784 	 */
785 
786 	ppcpu_path = this_cpu_ptr(clt->pcpu_path);
787 	path = rcu_dereference(*ppcpu_path);
788 	if (!path)
789 		path = list_first_or_null_rcu(&clt->paths_list,
790 					      typeof(*path), s.entry);
791 	else
792 		path = rtrs_clt_get_next_path_or_null(&clt->paths_list, path);
793 
794 	rcu_assign_pointer(*ppcpu_path, path);
795 
796 	return path;
797 }
798 
799 /**
800  * get_next_path_min_inflight() - Returns path with minimal inflight count.
801  * @it:	the path pointer
802  *
803  * Related to @MP_POLICY_MIN_INFLIGHT
804  *
805  * Locks:
806  *    rcu_read_lock() must be hold.
807  */
808 static struct rtrs_clt_path *get_next_path_min_inflight(struct path_it *it)
809 {
810 	struct rtrs_clt_path *min_path = NULL;
811 	struct rtrs_clt_sess *clt = it->clt;
812 	struct rtrs_clt_path *clt_path;
813 	int min_inflight = INT_MAX;
814 	int inflight;
815 
816 	list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
817 		if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
818 			continue;
819 
820 		if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
821 			continue;
822 
823 		inflight = atomic_read(&clt_path->stats->inflight);
824 
825 		if (inflight < min_inflight) {
826 			min_inflight = inflight;
827 			min_path = clt_path;
828 		}
829 	}
830 
831 	/*
832 	 * add the path to the skip list, so that next time we can get
833 	 * a different one
834 	 */
835 	if (min_path)
836 		list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
837 
838 	return min_path;
839 }
840 
841 /**
842  * get_next_path_min_latency() - Returns path with minimal latency.
843  * @it:	the path pointer
844  *
845  * Return: a path with the lowest latency or NULL if all paths are tried
846  *
847  * Locks:
848  *    rcu_read_lock() must be hold.
849  *
850  * Related to @MP_POLICY_MIN_LATENCY
851  *
852  * This DOES skip an already-tried path.
853  * There is a skip-list to skip a path if the path has tried but failed.
854  * It will try the minimum latency path and then the second minimum latency
855  * path and so on. Finally it will return NULL if all paths are tried.
856  * Therefore the caller MUST check the returned
857  * path is NULL and trigger the IO error.
858  */
859 static struct rtrs_clt_path *get_next_path_min_latency(struct path_it *it)
860 {
861 	struct rtrs_clt_path *min_path = NULL;
862 	struct rtrs_clt_sess *clt = it->clt;
863 	struct rtrs_clt_path *clt_path;
864 	ktime_t min_latency = KTIME_MAX;
865 	ktime_t latency;
866 
867 	list_for_each_entry_rcu(clt_path, &clt->paths_list, s.entry) {
868 		if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
869 			continue;
870 
871 		if (!list_empty(raw_cpu_ptr(clt_path->mp_skip_entry)))
872 			continue;
873 
874 		latency = clt_path->s.hb_cur_latency;
875 
876 		if (latency < min_latency) {
877 			min_latency = latency;
878 			min_path = clt_path;
879 		}
880 	}
881 
882 	/*
883 	 * add the path to the skip list, so that next time we can get
884 	 * a different one
885 	 */
886 	if (min_path)
887 		list_add(raw_cpu_ptr(min_path->mp_skip_entry), &it->skip_list);
888 
889 	return min_path;
890 }
891 
892 static inline void path_it_init(struct path_it *it, struct rtrs_clt_sess *clt)
893 {
894 	INIT_LIST_HEAD(&it->skip_list);
895 	it->clt = clt;
896 	it->i = 0;
897 
898 	if (clt->mp_policy == MP_POLICY_RR)
899 		it->next_path = get_next_path_rr;
900 	else if (clt->mp_policy == MP_POLICY_MIN_INFLIGHT)
901 		it->next_path = get_next_path_min_inflight;
902 	else
903 		it->next_path = get_next_path_min_latency;
904 }
905 
906 static inline void path_it_deinit(struct path_it *it)
907 {
908 	struct list_head *skip, *tmp;
909 	/*
910 	 * The skip_list is used only for the MIN_INFLIGHT and MIN_LATENCY policies.
911 	 * We need to remove paths from it, so that next IO can insert
912 	 * paths (->mp_skip_entry) into a skip_list again.
913 	 */
914 	list_for_each_safe(skip, tmp, &it->skip_list)
915 		list_del_init(skip);
916 }
917 
918 /**
919  * rtrs_clt_init_req() - Initialize an rtrs_clt_io_req holding information
920  * about an inflight IO.
921  * The user buffer holding user control message (not data) is copied into
922  * the corresponding buffer of rtrs_iu (req->iu->buf), which later on will
923  * also hold the control message of rtrs.
924  * @req: an io request holding information about IO.
925  * @clt_path: client path
926  * @conf: conformation callback function to notify upper layer.
927  * @permit: permit for allocation of RDMA remote buffer
928  * @priv: private pointer
929  * @vec: kernel vector containing control message
930  * @usr_len: length of the user message
931  * @sg: scater list for IO data
932  * @sg_cnt: number of scater list entries
933  * @data_len: length of the IO data
934  * @dir: direction of the IO.
935  */
936 static void rtrs_clt_init_req(struct rtrs_clt_io_req *req,
937 			      struct rtrs_clt_path *clt_path,
938 			      void (*conf)(void *priv, int errno),
939 			      struct rtrs_permit *permit, void *priv,
940 			      const struct kvec *vec, size_t usr_len,
941 			      struct scatterlist *sg, size_t sg_cnt,
942 			      size_t data_len, int dir)
943 {
944 	struct iov_iter iter;
945 	size_t len;
946 
947 	req->permit = permit;
948 	req->in_use = true;
949 	req->usr_len = usr_len;
950 	req->data_len = data_len;
951 	req->sglist = sg;
952 	req->sg_cnt = sg_cnt;
953 	req->priv = priv;
954 	req->dir = dir;
955 	req->con = rtrs_permit_to_clt_con(clt_path, permit);
956 	req->conf = conf;
957 	req->need_inv = false;
958 	req->need_inv_comp = false;
959 	req->inv_errno = 0;
960 	refcount_set(&req->ref, 1);
961 	req->mp_policy = clt_path->clt->mp_policy;
962 
963 	iov_iter_kvec(&iter, READ, vec, 1, usr_len);
964 	len = _copy_from_iter(req->iu->buf, usr_len, &iter);
965 	WARN_ON(len != usr_len);
966 
967 	reinit_completion(&req->inv_comp);
968 }
969 
970 static struct rtrs_clt_io_req *
971 rtrs_clt_get_req(struct rtrs_clt_path *clt_path,
972 		 void (*conf)(void *priv, int errno),
973 		 struct rtrs_permit *permit, void *priv,
974 		 const struct kvec *vec, size_t usr_len,
975 		 struct scatterlist *sg, size_t sg_cnt,
976 		 size_t data_len, int dir)
977 {
978 	struct rtrs_clt_io_req *req;
979 
980 	req = &clt_path->reqs[permit->mem_id];
981 	rtrs_clt_init_req(req, clt_path, conf, permit, priv, vec, usr_len,
982 			   sg, sg_cnt, data_len, dir);
983 	return req;
984 }
985 
986 static struct rtrs_clt_io_req *
987 rtrs_clt_get_copy_req(struct rtrs_clt_path *alive_path,
988 		       struct rtrs_clt_io_req *fail_req)
989 {
990 	struct rtrs_clt_io_req *req;
991 	struct kvec vec = {
992 		.iov_base = fail_req->iu->buf,
993 		.iov_len  = fail_req->usr_len
994 	};
995 
996 	req = &alive_path->reqs[fail_req->permit->mem_id];
997 	rtrs_clt_init_req(req, alive_path, fail_req->conf, fail_req->permit,
998 			   fail_req->priv, &vec, fail_req->usr_len,
999 			   fail_req->sglist, fail_req->sg_cnt,
1000 			   fail_req->data_len, fail_req->dir);
1001 	return req;
1002 }
1003 
1004 static int rtrs_post_rdma_write_sg(struct rtrs_clt_con *con,
1005 				   struct rtrs_clt_io_req *req,
1006 				   struct rtrs_rbuf *rbuf, bool fr_en,
1007 				   u32 count, u32 size, u32 imm,
1008 				   struct ib_send_wr *wr,
1009 				   struct ib_send_wr *tail)
1010 {
1011 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1012 	struct ib_sge *sge = req->sge;
1013 	enum ib_send_flags flags;
1014 	struct scatterlist *sg;
1015 	size_t num_sge;
1016 	int i;
1017 	struct ib_send_wr *ptail = NULL;
1018 
1019 	if (fr_en) {
1020 		i = 0;
1021 		sge[i].addr   = req->mr->iova;
1022 		sge[i].length = req->mr->length;
1023 		sge[i].lkey   = req->mr->lkey;
1024 		i++;
1025 		num_sge = 2;
1026 		ptail = tail;
1027 	} else {
1028 		for_each_sg(req->sglist, sg, count, i) {
1029 			sge[i].addr   = sg_dma_address(sg);
1030 			sge[i].length = sg_dma_len(sg);
1031 			sge[i].lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
1032 		}
1033 		num_sge = 1 + count;
1034 	}
1035 	sge[i].addr   = req->iu->dma_addr;
1036 	sge[i].length = size;
1037 	sge[i].lkey   = clt_path->s.dev->ib_pd->local_dma_lkey;
1038 
1039 	/*
1040 	 * From time to time we have to post signalled sends,
1041 	 * or send queue will fill up and only QP reset can help.
1042 	 */
1043 	flags = atomic_inc_return(&con->c.wr_cnt) % clt_path->s.signal_interval ?
1044 			0 : IB_SEND_SIGNALED;
1045 
1046 	ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
1047 				      req->iu->dma_addr,
1048 				      size, DMA_TO_DEVICE);
1049 
1050 	return rtrs_iu_post_rdma_write_imm(&con->c, req->iu, sge, num_sge,
1051 					    rbuf->rkey, rbuf->addr, imm,
1052 					    flags, wr, ptail);
1053 }
1054 
1055 static int rtrs_map_sg_fr(struct rtrs_clt_io_req *req, size_t count)
1056 {
1057 	int nr;
1058 
1059 	/* Align the MR to a 4K page size to match the block virt boundary */
1060 	nr = ib_map_mr_sg(req->mr, req->sglist, count, NULL, SZ_4K);
1061 	if (nr < 0)
1062 		return nr;
1063 	if (nr < req->sg_cnt)
1064 		return -EINVAL;
1065 	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1066 
1067 	return nr;
1068 }
1069 
1070 static int rtrs_clt_write_req(struct rtrs_clt_io_req *req)
1071 {
1072 	struct rtrs_clt_con *con = req->con;
1073 	struct rtrs_path *s = con->c.path;
1074 	struct rtrs_clt_path *clt_path = to_clt_path(s);
1075 	struct rtrs_msg_rdma_write *msg;
1076 
1077 	struct rtrs_rbuf *rbuf;
1078 	int ret, count = 0;
1079 	u32 imm, buf_id;
1080 	struct ib_reg_wr rwr;
1081 	struct ib_send_wr inv_wr;
1082 	struct ib_send_wr *wr = NULL;
1083 	bool fr_en = false;
1084 
1085 	const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1086 
1087 	if (tsize > clt_path->chunk_size) {
1088 		rtrs_wrn(s, "Write request failed, size too big %zu > %d\n",
1089 			  tsize, clt_path->chunk_size);
1090 		return -EMSGSIZE;
1091 	}
1092 	if (req->sg_cnt) {
1093 		count = ib_dma_map_sg(clt_path->s.dev->ib_dev, req->sglist,
1094 				      req->sg_cnt, req->dir);
1095 		if (!count) {
1096 			rtrs_wrn(s, "Write request failed, map failed\n");
1097 			return -EINVAL;
1098 		}
1099 	}
1100 	/* put rtrs msg after sg and user message */
1101 	msg = req->iu->buf + req->usr_len;
1102 	msg->type = cpu_to_le16(RTRS_MSG_WRITE);
1103 	msg->usr_len = cpu_to_le16(req->usr_len);
1104 
1105 	/* rtrs message on server side will be after user data and message */
1106 	imm = req->permit->mem_off + req->data_len + req->usr_len;
1107 	imm = rtrs_to_io_req_imm(imm);
1108 	buf_id = req->permit->mem_id;
1109 	req->sg_size = tsize;
1110 	rbuf = &clt_path->rbufs[buf_id];
1111 
1112 	if (count) {
1113 		ret = rtrs_map_sg_fr(req, count);
1114 		if (ret < 0) {
1115 			rtrs_err_rl(s,
1116 				    "Write request failed, failed to map fast reg. data, err: %d\n",
1117 				    ret);
1118 			ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1119 					req->sg_cnt, req->dir);
1120 			return ret;
1121 		}
1122 		inv_wr = (struct ib_send_wr) {
1123 			.opcode		    = IB_WR_LOCAL_INV,
1124 			.wr_cqe		    = &req->inv_cqe,
1125 			.send_flags	    = IB_SEND_SIGNALED,
1126 			.ex.invalidate_rkey = req->mr->rkey,
1127 		};
1128 		req->inv_cqe.done = rtrs_clt_inv_rkey_done;
1129 		rwr = (struct ib_reg_wr) {
1130 			.wr.opcode = IB_WR_REG_MR,
1131 			.wr.wr_cqe = &fast_reg_cqe,
1132 			.mr = req->mr,
1133 			.key = req->mr->rkey,
1134 			.access = (IB_ACCESS_LOCAL_WRITE),
1135 		};
1136 		wr = &rwr.wr;
1137 		fr_en = true;
1138 		refcount_inc(&req->ref);
1139 	}
1140 	/*
1141 	 * Update stats now, after request is successfully sent it is not
1142 	 * safe anymore to touch it.
1143 	 */
1144 	rtrs_clt_update_all_stats(req, WRITE);
1145 
1146 	ret = rtrs_post_rdma_write_sg(req->con, req, rbuf, fr_en, count,
1147 				      req->usr_len + sizeof(*msg),
1148 				      imm, wr, &inv_wr);
1149 	if (ret) {
1150 		rtrs_err_rl(s,
1151 			    "Write request failed: error=%d path=%s [%s:%u]\n",
1152 			    ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1153 			    clt_path->hca_port);
1154 		if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1155 			atomic_dec(&clt_path->stats->inflight);
1156 		if (req->sg_cnt)
1157 			ib_dma_unmap_sg(clt_path->s.dev->ib_dev, req->sglist,
1158 					req->sg_cnt, req->dir);
1159 	}
1160 
1161 	return ret;
1162 }
1163 
1164 static int rtrs_clt_read_req(struct rtrs_clt_io_req *req)
1165 {
1166 	struct rtrs_clt_con *con = req->con;
1167 	struct rtrs_path *s = con->c.path;
1168 	struct rtrs_clt_path *clt_path = to_clt_path(s);
1169 	struct rtrs_msg_rdma_read *msg;
1170 	struct rtrs_ib_dev *dev = clt_path->s.dev;
1171 
1172 	struct ib_reg_wr rwr;
1173 	struct ib_send_wr *wr = NULL;
1174 
1175 	int ret, count = 0;
1176 	u32 imm, buf_id;
1177 
1178 	const size_t tsize = sizeof(*msg) + req->data_len + req->usr_len;
1179 
1180 	if (tsize > clt_path->chunk_size) {
1181 		rtrs_wrn(s,
1182 			  "Read request failed, message size is %zu, bigger than CHUNK_SIZE %d\n",
1183 			  tsize, clt_path->chunk_size);
1184 		return -EMSGSIZE;
1185 	}
1186 
1187 	if (req->sg_cnt) {
1188 		count = ib_dma_map_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1189 				      req->dir);
1190 		if (!count) {
1191 			rtrs_wrn(s,
1192 				  "Read request failed, dma map failed\n");
1193 			return -EINVAL;
1194 		}
1195 	}
1196 	/* put our message into req->buf after user message*/
1197 	msg = req->iu->buf + req->usr_len;
1198 	msg->type = cpu_to_le16(RTRS_MSG_READ);
1199 	msg->usr_len = cpu_to_le16(req->usr_len);
1200 
1201 	if (count) {
1202 		ret = rtrs_map_sg_fr(req, count);
1203 		if (ret < 0) {
1204 			rtrs_err_rl(s,
1205 				     "Read request failed, failed to map  fast reg. data, err: %d\n",
1206 				     ret);
1207 			ib_dma_unmap_sg(dev->ib_dev, req->sglist, req->sg_cnt,
1208 					req->dir);
1209 			return ret;
1210 		}
1211 		rwr = (struct ib_reg_wr) {
1212 			.wr.opcode = IB_WR_REG_MR,
1213 			.wr.wr_cqe = &fast_reg_cqe,
1214 			.mr = req->mr,
1215 			.key = req->mr->rkey,
1216 			.access = (IB_ACCESS_LOCAL_WRITE |
1217 				   IB_ACCESS_REMOTE_WRITE),
1218 		};
1219 		wr = &rwr.wr;
1220 
1221 		msg->sg_cnt = cpu_to_le16(1);
1222 		msg->flags = cpu_to_le16(RTRS_MSG_NEED_INVAL_F);
1223 
1224 		msg->desc[0].addr = cpu_to_le64(req->mr->iova);
1225 		msg->desc[0].key = cpu_to_le32(req->mr->rkey);
1226 		msg->desc[0].len = cpu_to_le32(req->mr->length);
1227 
1228 		/* Further invalidation is required */
1229 		req->need_inv = !!RTRS_MSG_NEED_INVAL_F;
1230 
1231 	} else {
1232 		msg->sg_cnt = 0;
1233 		msg->flags = 0;
1234 	}
1235 	/*
1236 	 * rtrs message will be after the space reserved for disk data and
1237 	 * user message
1238 	 */
1239 	imm = req->permit->mem_off + req->data_len + req->usr_len;
1240 	imm = rtrs_to_io_req_imm(imm);
1241 	buf_id = req->permit->mem_id;
1242 
1243 	req->sg_size  = sizeof(*msg);
1244 	req->sg_size += le16_to_cpu(msg->sg_cnt) * sizeof(struct rtrs_sg_desc);
1245 	req->sg_size += req->usr_len;
1246 
1247 	/*
1248 	 * Update stats now, after request is successfully sent it is not
1249 	 * safe anymore to touch it.
1250 	 */
1251 	rtrs_clt_update_all_stats(req, READ);
1252 
1253 	ret = rtrs_post_send_rdma(req->con, req, &clt_path->rbufs[buf_id],
1254 				   req->data_len, imm, wr);
1255 	if (ret) {
1256 		rtrs_err_rl(s,
1257 			    "Read request failed: error=%d path=%s [%s:%u]\n",
1258 			    ret, kobject_name(&clt_path->kobj), clt_path->hca_name,
1259 			    clt_path->hca_port);
1260 		if (req->mp_policy == MP_POLICY_MIN_INFLIGHT)
1261 			atomic_dec(&clt_path->stats->inflight);
1262 		req->need_inv = false;
1263 		if (req->sg_cnt)
1264 			ib_dma_unmap_sg(dev->ib_dev, req->sglist,
1265 					req->sg_cnt, req->dir);
1266 	}
1267 
1268 	return ret;
1269 }
1270 
1271 /**
1272  * rtrs_clt_failover_req() - Try to find an active path for a failed request
1273  * @clt: clt context
1274  * @fail_req: a failed io request.
1275  */
1276 static int rtrs_clt_failover_req(struct rtrs_clt_sess *clt,
1277 				 struct rtrs_clt_io_req *fail_req)
1278 {
1279 	struct rtrs_clt_path *alive_path;
1280 	struct rtrs_clt_io_req *req;
1281 	int err = -ECONNABORTED;
1282 	struct path_it it;
1283 
1284 	rcu_read_lock();
1285 	for (path_it_init(&it, clt);
1286 	     (alive_path = it.next_path(&it)) && it.i < it.clt->paths_num;
1287 	     it.i++) {
1288 		if (READ_ONCE(alive_path->state) != RTRS_CLT_CONNECTED)
1289 			continue;
1290 		req = rtrs_clt_get_copy_req(alive_path, fail_req);
1291 		if (req->dir == DMA_TO_DEVICE)
1292 			err = rtrs_clt_write_req(req);
1293 		else
1294 			err = rtrs_clt_read_req(req);
1295 		if (err) {
1296 			req->in_use = false;
1297 			continue;
1298 		}
1299 		/* Success path */
1300 		rtrs_clt_inc_failover_cnt(alive_path->stats);
1301 		break;
1302 	}
1303 	path_it_deinit(&it);
1304 	rcu_read_unlock();
1305 
1306 	return err;
1307 }
1308 
1309 static void fail_all_outstanding_reqs(struct rtrs_clt_path *clt_path)
1310 {
1311 	struct rtrs_clt_sess *clt = clt_path->clt;
1312 	struct rtrs_clt_io_req *req;
1313 	int i, err;
1314 
1315 	if (!clt_path->reqs)
1316 		return;
1317 	for (i = 0; i < clt_path->queue_depth; ++i) {
1318 		req = &clt_path->reqs[i];
1319 		if (!req->in_use)
1320 			continue;
1321 
1322 		/*
1323 		 * Safely (without notification) complete failed request.
1324 		 * After completion this request is still useble and can
1325 		 * be failovered to another path.
1326 		 */
1327 		complete_rdma_req(req, -ECONNABORTED, false, true);
1328 
1329 		err = rtrs_clt_failover_req(clt, req);
1330 		if (err)
1331 			/* Failover failed, notify anyway */
1332 			req->conf(req->priv, err);
1333 	}
1334 }
1335 
1336 static void free_path_reqs(struct rtrs_clt_path *clt_path)
1337 {
1338 	struct rtrs_clt_io_req *req;
1339 	int i;
1340 
1341 	if (!clt_path->reqs)
1342 		return;
1343 	for (i = 0; i < clt_path->queue_depth; ++i) {
1344 		req = &clt_path->reqs[i];
1345 		if (req->mr)
1346 			ib_dereg_mr(req->mr);
1347 		kfree(req->sge);
1348 		rtrs_iu_free(req->iu, clt_path->s.dev->ib_dev, 1);
1349 	}
1350 	kfree(clt_path->reqs);
1351 	clt_path->reqs = NULL;
1352 }
1353 
1354 static int alloc_path_reqs(struct rtrs_clt_path *clt_path)
1355 {
1356 	struct rtrs_clt_io_req *req;
1357 	int i, err = -ENOMEM;
1358 
1359 	clt_path->reqs = kcalloc(clt_path->queue_depth,
1360 				 sizeof(*clt_path->reqs),
1361 				 GFP_KERNEL);
1362 	if (!clt_path->reqs)
1363 		return -ENOMEM;
1364 
1365 	for (i = 0; i < clt_path->queue_depth; ++i) {
1366 		req = &clt_path->reqs[i];
1367 		req->iu = rtrs_iu_alloc(1, clt_path->max_hdr_size, GFP_KERNEL,
1368 					 clt_path->s.dev->ib_dev,
1369 					 DMA_TO_DEVICE,
1370 					 rtrs_clt_rdma_done);
1371 		if (!req->iu)
1372 			goto out;
1373 
1374 		req->sge = kcalloc(2, sizeof(*req->sge), GFP_KERNEL);
1375 		if (!req->sge)
1376 			goto out;
1377 
1378 		req->mr = ib_alloc_mr(clt_path->s.dev->ib_pd,
1379 				      IB_MR_TYPE_MEM_REG,
1380 				      clt_path->max_pages_per_mr);
1381 		if (IS_ERR(req->mr)) {
1382 			err = PTR_ERR(req->mr);
1383 			req->mr = NULL;
1384 			pr_err("Failed to alloc clt_path->max_pages_per_mr %d\n",
1385 			       clt_path->max_pages_per_mr);
1386 			goto out;
1387 		}
1388 
1389 		init_completion(&req->inv_comp);
1390 	}
1391 
1392 	return 0;
1393 
1394 out:
1395 	free_path_reqs(clt_path);
1396 
1397 	return err;
1398 }
1399 
1400 static int alloc_permits(struct rtrs_clt_sess *clt)
1401 {
1402 	unsigned int chunk_bits;
1403 	int err, i;
1404 
1405 	clt->permits_map = bitmap_zalloc(clt->queue_depth, GFP_KERNEL);
1406 	if (!clt->permits_map) {
1407 		err = -ENOMEM;
1408 		goto out_err;
1409 	}
1410 	clt->permits = kcalloc(clt->queue_depth, permit_size(clt), GFP_KERNEL);
1411 	if (!clt->permits) {
1412 		err = -ENOMEM;
1413 		goto err_map;
1414 	}
1415 	chunk_bits = ilog2(clt->queue_depth - 1) + 1;
1416 	for (i = 0; i < clt->queue_depth; i++) {
1417 		struct rtrs_permit *permit;
1418 
1419 		permit = get_permit(clt, i);
1420 		permit->mem_id = i;
1421 		permit->mem_off = i << (MAX_IMM_PAYL_BITS - chunk_bits);
1422 	}
1423 
1424 	return 0;
1425 
1426 err_map:
1427 	bitmap_free(clt->permits_map);
1428 	clt->permits_map = NULL;
1429 out_err:
1430 	return err;
1431 }
1432 
1433 static void free_permits(struct rtrs_clt_sess *clt)
1434 {
1435 	if (clt->permits_map)
1436 		wait_event(clt->permits_wait,
1437 			   bitmap_empty(clt->permits_map, clt->queue_depth));
1438 
1439 	bitmap_free(clt->permits_map);
1440 	clt->permits_map = NULL;
1441 	kfree(clt->permits);
1442 	clt->permits = NULL;
1443 }
1444 
1445 static void query_fast_reg_mode(struct rtrs_clt_path *clt_path)
1446 {
1447 	struct ib_device *ib_dev;
1448 	u64 max_pages_per_mr;
1449 	int mr_page_shift;
1450 
1451 	ib_dev = clt_path->s.dev->ib_dev;
1452 
1453 	/*
1454 	 * Use the smallest page size supported by the HCA, down to a
1455 	 * minimum of 4096 bytes. We're unlikely to build large sglists
1456 	 * out of smaller entries.
1457 	 */
1458 	mr_page_shift      = max(12, ffs(ib_dev->attrs.page_size_cap) - 1);
1459 	max_pages_per_mr   = ib_dev->attrs.max_mr_size;
1460 	do_div(max_pages_per_mr, (1ull << mr_page_shift));
1461 	clt_path->max_pages_per_mr =
1462 		min3(clt_path->max_pages_per_mr, (u32)max_pages_per_mr,
1463 		     ib_dev->attrs.max_fast_reg_page_list_len);
1464 	clt_path->clt->max_segments =
1465 		min(clt_path->max_pages_per_mr, clt_path->clt->max_segments);
1466 }
1467 
1468 static bool rtrs_clt_change_state_get_old(struct rtrs_clt_path *clt_path,
1469 					   enum rtrs_clt_state new_state,
1470 					   enum rtrs_clt_state *old_state)
1471 {
1472 	bool changed;
1473 
1474 	spin_lock_irq(&clt_path->state_wq.lock);
1475 	if (old_state)
1476 		*old_state = clt_path->state;
1477 	changed = rtrs_clt_change_state(clt_path, new_state);
1478 	spin_unlock_irq(&clt_path->state_wq.lock);
1479 
1480 	return changed;
1481 }
1482 
1483 static void rtrs_clt_hb_err_handler(struct rtrs_con *c)
1484 {
1485 	struct rtrs_clt_con *con = container_of(c, typeof(*con), c);
1486 
1487 	rtrs_rdma_error_recovery(con);
1488 }
1489 
1490 static void rtrs_clt_init_hb(struct rtrs_clt_path *clt_path)
1491 {
1492 	rtrs_init_hb(&clt_path->s, &io_comp_cqe,
1493 		      RTRS_HB_INTERVAL_MS,
1494 		      RTRS_HB_MISSED_MAX,
1495 		      rtrs_clt_hb_err_handler,
1496 		      rtrs_wq);
1497 }
1498 
1499 static void rtrs_clt_reconnect_work(struct work_struct *work);
1500 static void rtrs_clt_close_work(struct work_struct *work);
1501 
1502 static void rtrs_clt_err_recovery_work(struct work_struct *work)
1503 {
1504 	struct rtrs_clt_path *clt_path;
1505 	struct rtrs_clt_sess *clt;
1506 	int delay_ms;
1507 
1508 	clt_path = container_of(work, struct rtrs_clt_path, err_recovery_work);
1509 	clt = clt_path->clt;
1510 	delay_ms = clt->reconnect_delay_sec * 1000;
1511 	rtrs_clt_stop_and_destroy_conns(clt_path);
1512 	queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork,
1513 			   msecs_to_jiffies(delay_ms +
1514 					    prandom_u32() %
1515 					    RTRS_RECONNECT_SEED));
1516 }
1517 
1518 static struct rtrs_clt_path *alloc_path(struct rtrs_clt_sess *clt,
1519 					const struct rtrs_addr *path,
1520 					size_t con_num, u32 nr_poll_queues)
1521 {
1522 	struct rtrs_clt_path *clt_path;
1523 	int err = -ENOMEM;
1524 	int cpu;
1525 	size_t total_con;
1526 
1527 	clt_path = kzalloc(sizeof(*clt_path), GFP_KERNEL);
1528 	if (!clt_path)
1529 		goto err;
1530 
1531 	/*
1532 	 * irqmode and poll
1533 	 * +1: Extra connection for user messages
1534 	 */
1535 	total_con = con_num + nr_poll_queues + 1;
1536 	clt_path->s.con = kcalloc(total_con, sizeof(*clt_path->s.con),
1537 				  GFP_KERNEL);
1538 	if (!clt_path->s.con)
1539 		goto err_free_path;
1540 
1541 	clt_path->s.con_num = total_con;
1542 	clt_path->s.irq_con_num = con_num + 1;
1543 
1544 	clt_path->stats = kzalloc(sizeof(*clt_path->stats), GFP_KERNEL);
1545 	if (!clt_path->stats)
1546 		goto err_free_con;
1547 
1548 	mutex_init(&clt_path->init_mutex);
1549 	uuid_gen(&clt_path->s.uuid);
1550 	memcpy(&clt_path->s.dst_addr, path->dst,
1551 	       rdma_addr_size((struct sockaddr *)path->dst));
1552 
1553 	/*
1554 	 * rdma_resolve_addr() passes src_addr to cma_bind_addr, which
1555 	 * checks the sa_family to be non-zero. If user passed src_addr=NULL
1556 	 * the sess->src_addr will contain only zeros, which is then fine.
1557 	 */
1558 	if (path->src)
1559 		memcpy(&clt_path->s.src_addr, path->src,
1560 		       rdma_addr_size((struct sockaddr *)path->src));
1561 	strscpy(clt_path->s.sessname, clt->sessname,
1562 		sizeof(clt_path->s.sessname));
1563 	clt_path->clt = clt;
1564 	clt_path->max_pages_per_mr = RTRS_MAX_SEGMENTS;
1565 	init_waitqueue_head(&clt_path->state_wq);
1566 	clt_path->state = RTRS_CLT_CONNECTING;
1567 	atomic_set(&clt_path->connected_cnt, 0);
1568 	INIT_WORK(&clt_path->close_work, rtrs_clt_close_work);
1569 	INIT_WORK(&clt_path->err_recovery_work, rtrs_clt_err_recovery_work);
1570 	INIT_DELAYED_WORK(&clt_path->reconnect_dwork, rtrs_clt_reconnect_work);
1571 	rtrs_clt_init_hb(clt_path);
1572 
1573 	clt_path->mp_skip_entry = alloc_percpu(typeof(*clt_path->mp_skip_entry));
1574 	if (!clt_path->mp_skip_entry)
1575 		goto err_free_stats;
1576 
1577 	for_each_possible_cpu(cpu)
1578 		INIT_LIST_HEAD(per_cpu_ptr(clt_path->mp_skip_entry, cpu));
1579 
1580 	err = rtrs_clt_init_stats(clt_path->stats);
1581 	if (err)
1582 		goto err_free_percpu;
1583 
1584 	return clt_path;
1585 
1586 err_free_percpu:
1587 	free_percpu(clt_path->mp_skip_entry);
1588 err_free_stats:
1589 	kfree(clt_path->stats);
1590 err_free_con:
1591 	kfree(clt_path->s.con);
1592 err_free_path:
1593 	kfree(clt_path);
1594 err:
1595 	return ERR_PTR(err);
1596 }
1597 
1598 void free_path(struct rtrs_clt_path *clt_path)
1599 {
1600 	free_percpu(clt_path->mp_skip_entry);
1601 	mutex_destroy(&clt_path->init_mutex);
1602 	kfree(clt_path->s.con);
1603 	kfree(clt_path->rbufs);
1604 	kfree(clt_path);
1605 }
1606 
1607 static int create_con(struct rtrs_clt_path *clt_path, unsigned int cid)
1608 {
1609 	struct rtrs_clt_con *con;
1610 
1611 	con = kzalloc(sizeof(*con), GFP_KERNEL);
1612 	if (!con)
1613 		return -ENOMEM;
1614 
1615 	/* Map first two connections to the first CPU */
1616 	con->cpu  = (cid ? cid - 1 : 0) % nr_cpu_ids;
1617 	con->c.cid = cid;
1618 	con->c.path = &clt_path->s;
1619 	/* Align with srv, init as 1 */
1620 	atomic_set(&con->c.wr_cnt, 1);
1621 	mutex_init(&con->con_mutex);
1622 
1623 	clt_path->s.con[cid] = &con->c;
1624 
1625 	return 0;
1626 }
1627 
1628 static void destroy_con(struct rtrs_clt_con *con)
1629 {
1630 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1631 
1632 	clt_path->s.con[con->c.cid] = NULL;
1633 	mutex_destroy(&con->con_mutex);
1634 	kfree(con);
1635 }
1636 
1637 static int create_con_cq_qp(struct rtrs_clt_con *con)
1638 {
1639 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1640 	u32 max_send_wr, max_recv_wr, cq_num, max_send_sge, wr_limit;
1641 	int err, cq_vector;
1642 	struct rtrs_msg_rkey_rsp *rsp;
1643 
1644 	lockdep_assert_held(&con->con_mutex);
1645 	if (con->c.cid == 0) {
1646 		max_send_sge = 1;
1647 		/* We must be the first here */
1648 		if (WARN_ON(clt_path->s.dev))
1649 			return -EINVAL;
1650 
1651 		/*
1652 		 * The whole session uses device from user connection.
1653 		 * Be careful not to close user connection before ib dev
1654 		 * is gracefully put.
1655 		 */
1656 		clt_path->s.dev = rtrs_ib_dev_find_or_add(con->c.cm_id->device,
1657 						       &dev_pd);
1658 		if (!clt_path->s.dev) {
1659 			rtrs_wrn(clt_path->clt,
1660 				  "rtrs_ib_dev_find_get_or_add(): no memory\n");
1661 			return -ENOMEM;
1662 		}
1663 		clt_path->s.dev_ref = 1;
1664 		query_fast_reg_mode(clt_path);
1665 		wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1666 		/*
1667 		 * Two (request + registration) completion for send
1668 		 * Two for recv if always_invalidate is set on server
1669 		 * or one for recv.
1670 		 * + 2 for drain and heartbeat
1671 		 * in case qp gets into error state.
1672 		 */
1673 		max_send_wr =
1674 			min_t(int, wr_limit, SERVICE_CON_QUEUE_DEPTH * 2 + 2);
1675 		max_recv_wr = max_send_wr;
1676 	} else {
1677 		/*
1678 		 * Here we assume that session members are correctly set.
1679 		 * This is always true if user connection (cid == 0) is
1680 		 * established first.
1681 		 */
1682 		if (WARN_ON(!clt_path->s.dev))
1683 			return -EINVAL;
1684 		if (WARN_ON(!clt_path->queue_depth))
1685 			return -EINVAL;
1686 
1687 		wr_limit = clt_path->s.dev->ib_dev->attrs.max_qp_wr;
1688 		/* Shared between connections */
1689 		clt_path->s.dev_ref++;
1690 		max_send_wr = min_t(int, wr_limit,
1691 			      /* QD * (REQ + RSP + FR REGS or INVS) + drain */
1692 			      clt_path->queue_depth * 3 + 1);
1693 		max_recv_wr = min_t(int, wr_limit,
1694 			      clt_path->queue_depth * 3 + 1);
1695 		max_send_sge = 2;
1696 	}
1697 	atomic_set(&con->c.sq_wr_avail, max_send_wr);
1698 	cq_num = max_send_wr + max_recv_wr;
1699 	/* alloc iu to recv new rkey reply when server reports flags set */
1700 	if (clt_path->flags & RTRS_MSG_NEW_RKEY_F || con->c.cid == 0) {
1701 		con->rsp_ius = rtrs_iu_alloc(cq_num, sizeof(*rsp),
1702 					      GFP_KERNEL,
1703 					      clt_path->s.dev->ib_dev,
1704 					      DMA_FROM_DEVICE,
1705 					      rtrs_clt_rdma_done);
1706 		if (!con->rsp_ius)
1707 			return -ENOMEM;
1708 		con->queue_num = cq_num;
1709 	}
1710 	cq_num = max_send_wr + max_recv_wr;
1711 	cq_vector = con->cpu % clt_path->s.dev->ib_dev->num_comp_vectors;
1712 	if (con->c.cid >= clt_path->s.irq_con_num)
1713 		err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1714 					cq_vector, cq_num, max_send_wr,
1715 					max_recv_wr, IB_POLL_DIRECT);
1716 	else
1717 		err = rtrs_cq_qp_create(&clt_path->s, &con->c, max_send_sge,
1718 					cq_vector, cq_num, max_send_wr,
1719 					max_recv_wr, IB_POLL_SOFTIRQ);
1720 	/*
1721 	 * In case of error we do not bother to clean previous allocations,
1722 	 * since destroy_con_cq_qp() must be called.
1723 	 */
1724 	return err;
1725 }
1726 
1727 static void destroy_con_cq_qp(struct rtrs_clt_con *con)
1728 {
1729 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1730 
1731 	/*
1732 	 * Be careful here: destroy_con_cq_qp() can be called even
1733 	 * create_con_cq_qp() failed, see comments there.
1734 	 */
1735 	lockdep_assert_held(&con->con_mutex);
1736 	rtrs_cq_qp_destroy(&con->c);
1737 	if (con->rsp_ius) {
1738 		rtrs_iu_free(con->rsp_ius, clt_path->s.dev->ib_dev,
1739 			     con->queue_num);
1740 		con->rsp_ius = NULL;
1741 		con->queue_num = 0;
1742 	}
1743 	if (clt_path->s.dev_ref && !--clt_path->s.dev_ref) {
1744 		rtrs_ib_dev_put(clt_path->s.dev);
1745 		clt_path->s.dev = NULL;
1746 	}
1747 }
1748 
1749 static void stop_cm(struct rtrs_clt_con *con)
1750 {
1751 	rdma_disconnect(con->c.cm_id);
1752 	if (con->c.qp)
1753 		ib_drain_qp(con->c.qp);
1754 }
1755 
1756 static void destroy_cm(struct rtrs_clt_con *con)
1757 {
1758 	rdma_destroy_id(con->c.cm_id);
1759 	con->c.cm_id = NULL;
1760 }
1761 
1762 static int rtrs_rdma_addr_resolved(struct rtrs_clt_con *con)
1763 {
1764 	struct rtrs_path *s = con->c.path;
1765 	int err;
1766 
1767 	mutex_lock(&con->con_mutex);
1768 	err = create_con_cq_qp(con);
1769 	mutex_unlock(&con->con_mutex);
1770 	if (err) {
1771 		rtrs_err(s, "create_con_cq_qp(), err: %d\n", err);
1772 		return err;
1773 	}
1774 	err = rdma_resolve_route(con->c.cm_id, RTRS_CONNECT_TIMEOUT_MS);
1775 	if (err)
1776 		rtrs_err(s, "Resolving route failed, err: %d\n", err);
1777 
1778 	return err;
1779 }
1780 
1781 static int rtrs_rdma_route_resolved(struct rtrs_clt_con *con)
1782 {
1783 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1784 	struct rtrs_clt_sess *clt = clt_path->clt;
1785 	struct rtrs_msg_conn_req msg;
1786 	struct rdma_conn_param param;
1787 
1788 	int err;
1789 
1790 	param = (struct rdma_conn_param) {
1791 		.retry_count = 7,
1792 		.rnr_retry_count = 7,
1793 		.private_data = &msg,
1794 		.private_data_len = sizeof(msg),
1795 	};
1796 
1797 	msg = (struct rtrs_msg_conn_req) {
1798 		.magic = cpu_to_le16(RTRS_MAGIC),
1799 		.version = cpu_to_le16(RTRS_PROTO_VER),
1800 		.cid = cpu_to_le16(con->c.cid),
1801 		.cid_num = cpu_to_le16(clt_path->s.con_num),
1802 		.recon_cnt = cpu_to_le16(clt_path->s.recon_cnt),
1803 	};
1804 	msg.first_conn = clt_path->for_new_clt ? FIRST_CONN : 0;
1805 	uuid_copy(&msg.sess_uuid, &clt_path->s.uuid);
1806 	uuid_copy(&msg.paths_uuid, &clt->paths_uuid);
1807 
1808 	err = rdma_connect_locked(con->c.cm_id, &param);
1809 	if (err)
1810 		rtrs_err(clt, "rdma_connect_locked(): %d\n", err);
1811 
1812 	return err;
1813 }
1814 
1815 static int rtrs_rdma_conn_established(struct rtrs_clt_con *con,
1816 				       struct rdma_cm_event *ev)
1817 {
1818 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1819 	struct rtrs_clt_sess *clt = clt_path->clt;
1820 	const struct rtrs_msg_conn_rsp *msg;
1821 	u16 version, queue_depth;
1822 	int errno;
1823 	u8 len;
1824 
1825 	msg = ev->param.conn.private_data;
1826 	len = ev->param.conn.private_data_len;
1827 	if (len < sizeof(*msg)) {
1828 		rtrs_err(clt, "Invalid RTRS connection response\n");
1829 		return -ECONNRESET;
1830 	}
1831 	if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1832 		rtrs_err(clt, "Invalid RTRS magic\n");
1833 		return -ECONNRESET;
1834 	}
1835 	version = le16_to_cpu(msg->version);
1836 	if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1837 		rtrs_err(clt, "Unsupported major RTRS version: %d, expected %d\n",
1838 			  version >> 8, RTRS_PROTO_VER_MAJOR);
1839 		return -ECONNRESET;
1840 	}
1841 	errno = le16_to_cpu(msg->errno);
1842 	if (errno) {
1843 		rtrs_err(clt, "Invalid RTRS message: errno %d\n",
1844 			  errno);
1845 		return -ECONNRESET;
1846 	}
1847 	if (con->c.cid == 0) {
1848 		queue_depth = le16_to_cpu(msg->queue_depth);
1849 
1850 		if (clt_path->queue_depth > 0 && queue_depth != clt_path->queue_depth) {
1851 			rtrs_err(clt, "Error: queue depth changed\n");
1852 
1853 			/*
1854 			 * Stop any more reconnection attempts
1855 			 */
1856 			clt_path->reconnect_attempts = -1;
1857 			rtrs_err(clt,
1858 				"Disabling auto-reconnect. Trigger a manual reconnect after issue is resolved\n");
1859 			return -ECONNRESET;
1860 		}
1861 
1862 		if (!clt_path->rbufs) {
1863 			clt_path->rbufs = kcalloc(queue_depth,
1864 						  sizeof(*clt_path->rbufs),
1865 						  GFP_KERNEL);
1866 			if (!clt_path->rbufs)
1867 				return -ENOMEM;
1868 		}
1869 		clt_path->queue_depth = queue_depth;
1870 		clt_path->s.signal_interval = min_not_zero(queue_depth,
1871 						(unsigned short) SERVICE_CON_QUEUE_DEPTH);
1872 		clt_path->max_hdr_size = le32_to_cpu(msg->max_hdr_size);
1873 		clt_path->max_io_size = le32_to_cpu(msg->max_io_size);
1874 		clt_path->flags = le32_to_cpu(msg->flags);
1875 		clt_path->chunk_size = clt_path->max_io_size + clt_path->max_hdr_size;
1876 
1877 		/*
1878 		 * Global IO size is always a minimum.
1879 		 * If while a reconnection server sends us a value a bit
1880 		 * higher - client does not care and uses cached minimum.
1881 		 *
1882 		 * Since we can have several sessions (paths) restablishing
1883 		 * connections in parallel, use lock.
1884 		 */
1885 		mutex_lock(&clt->paths_mutex);
1886 		clt->queue_depth = clt_path->queue_depth;
1887 		clt->max_io_size = min_not_zero(clt_path->max_io_size,
1888 						clt->max_io_size);
1889 		mutex_unlock(&clt->paths_mutex);
1890 
1891 		/*
1892 		 * Cache the hca_port and hca_name for sysfs
1893 		 */
1894 		clt_path->hca_port = con->c.cm_id->port_num;
1895 		scnprintf(clt_path->hca_name, sizeof(clt_path->hca_name),
1896 			  clt_path->s.dev->ib_dev->name);
1897 		clt_path->s.src_addr = con->c.cm_id->route.addr.src_addr;
1898 		/* set for_new_clt, to allow future reconnect on any path */
1899 		clt_path->for_new_clt = 1;
1900 	}
1901 
1902 	return 0;
1903 }
1904 
1905 static inline void flag_success_on_conn(struct rtrs_clt_con *con)
1906 {
1907 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
1908 
1909 	atomic_inc(&clt_path->connected_cnt);
1910 	con->cm_err = 1;
1911 }
1912 
1913 static int rtrs_rdma_conn_rejected(struct rtrs_clt_con *con,
1914 				    struct rdma_cm_event *ev)
1915 {
1916 	struct rtrs_path *s = con->c.path;
1917 	const struct rtrs_msg_conn_rsp *msg;
1918 	const char *rej_msg;
1919 	int status, errno;
1920 	u8 data_len;
1921 
1922 	status = ev->status;
1923 	rej_msg = rdma_reject_msg(con->c.cm_id, status);
1924 	msg = rdma_consumer_reject_data(con->c.cm_id, ev, &data_len);
1925 
1926 	if (msg && data_len >= sizeof(*msg)) {
1927 		errno = (int16_t)le16_to_cpu(msg->errno);
1928 		if (errno == -EBUSY)
1929 			rtrs_err(s,
1930 				  "Previous session is still exists on the server, please reconnect later\n");
1931 		else
1932 			rtrs_err(s,
1933 				  "Connect rejected: status %d (%s), rtrs errno %d\n",
1934 				  status, rej_msg, errno);
1935 	} else {
1936 		rtrs_err(s,
1937 			  "Connect rejected but with malformed message: status %d (%s)\n",
1938 			  status, rej_msg);
1939 	}
1940 
1941 	return -ECONNRESET;
1942 }
1943 
1944 void rtrs_clt_close_conns(struct rtrs_clt_path *clt_path, bool wait)
1945 {
1946 	if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSING, NULL))
1947 		queue_work(rtrs_wq, &clt_path->close_work);
1948 	if (wait)
1949 		flush_work(&clt_path->close_work);
1950 }
1951 
1952 static inline void flag_error_on_conn(struct rtrs_clt_con *con, int cm_err)
1953 {
1954 	if (con->cm_err == 1) {
1955 		struct rtrs_clt_path *clt_path;
1956 
1957 		clt_path = to_clt_path(con->c.path);
1958 		if (atomic_dec_and_test(&clt_path->connected_cnt))
1959 
1960 			wake_up(&clt_path->state_wq);
1961 	}
1962 	con->cm_err = cm_err;
1963 }
1964 
1965 static int rtrs_clt_rdma_cm_handler(struct rdma_cm_id *cm_id,
1966 				     struct rdma_cm_event *ev)
1967 {
1968 	struct rtrs_clt_con *con = cm_id->context;
1969 	struct rtrs_path *s = con->c.path;
1970 	struct rtrs_clt_path *clt_path = to_clt_path(s);
1971 	int cm_err = 0;
1972 
1973 	switch (ev->event) {
1974 	case RDMA_CM_EVENT_ADDR_RESOLVED:
1975 		cm_err = rtrs_rdma_addr_resolved(con);
1976 		break;
1977 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1978 		cm_err = rtrs_rdma_route_resolved(con);
1979 		break;
1980 	case RDMA_CM_EVENT_ESTABLISHED:
1981 		cm_err = rtrs_rdma_conn_established(con, ev);
1982 		if (!cm_err) {
1983 			/*
1984 			 * Report success and wake up. Here we abuse state_wq,
1985 			 * i.e. wake up without state change, but we set cm_err.
1986 			 */
1987 			flag_success_on_conn(con);
1988 			wake_up(&clt_path->state_wq);
1989 			return 0;
1990 		}
1991 		break;
1992 	case RDMA_CM_EVENT_REJECTED:
1993 		cm_err = rtrs_rdma_conn_rejected(con, ev);
1994 		break;
1995 	case RDMA_CM_EVENT_DISCONNECTED:
1996 		/* No message for disconnecting */
1997 		cm_err = -ECONNRESET;
1998 		break;
1999 	case RDMA_CM_EVENT_CONNECT_ERROR:
2000 	case RDMA_CM_EVENT_UNREACHABLE:
2001 	case RDMA_CM_EVENT_ADDR_CHANGE:
2002 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2003 		rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2004 			 rdma_event_msg(ev->event), ev->status);
2005 		cm_err = -ECONNRESET;
2006 		break;
2007 	case RDMA_CM_EVENT_ADDR_ERROR:
2008 	case RDMA_CM_EVENT_ROUTE_ERROR:
2009 		rtrs_wrn(s, "CM error (CM event: %s, err: %d)\n",
2010 			 rdma_event_msg(ev->event), ev->status);
2011 		cm_err = -EHOSTUNREACH;
2012 		break;
2013 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
2014 		/*
2015 		 * Device removal is a special case.  Queue close and return 0.
2016 		 */
2017 		rtrs_clt_close_conns(clt_path, false);
2018 		return 0;
2019 	default:
2020 		rtrs_err(s, "Unexpected RDMA CM error (CM event: %s, err: %d)\n",
2021 			 rdma_event_msg(ev->event), ev->status);
2022 		cm_err = -ECONNRESET;
2023 		break;
2024 	}
2025 
2026 	if (cm_err) {
2027 		/*
2028 		 * cm error makes sense only on connection establishing,
2029 		 * in other cases we rely on normal procedure of reconnecting.
2030 		 */
2031 		flag_error_on_conn(con, cm_err);
2032 		rtrs_rdma_error_recovery(con);
2033 	}
2034 
2035 	return 0;
2036 }
2037 
2038 static int create_cm(struct rtrs_clt_con *con)
2039 {
2040 	struct rtrs_path *s = con->c.path;
2041 	struct rtrs_clt_path *clt_path = to_clt_path(s);
2042 	struct rdma_cm_id *cm_id;
2043 	int err;
2044 
2045 	cm_id = rdma_create_id(&init_net, rtrs_clt_rdma_cm_handler, con,
2046 			       clt_path->s.dst_addr.ss_family == AF_IB ?
2047 			       RDMA_PS_IB : RDMA_PS_TCP, IB_QPT_RC);
2048 	if (IS_ERR(cm_id)) {
2049 		err = PTR_ERR(cm_id);
2050 		rtrs_err(s, "Failed to create CM ID, err: %d\n", err);
2051 
2052 		return err;
2053 	}
2054 	con->c.cm_id = cm_id;
2055 	con->cm_err = 0;
2056 	/* allow the port to be reused */
2057 	err = rdma_set_reuseaddr(cm_id, 1);
2058 	if (err != 0) {
2059 		rtrs_err(s, "Set address reuse failed, err: %d\n", err);
2060 		goto destroy_cm;
2061 	}
2062 	err = rdma_resolve_addr(cm_id, (struct sockaddr *)&clt_path->s.src_addr,
2063 				(struct sockaddr *)&clt_path->s.dst_addr,
2064 				RTRS_CONNECT_TIMEOUT_MS);
2065 	if (err) {
2066 		rtrs_err(s, "Failed to resolve address, err: %d\n", err);
2067 		goto destroy_cm;
2068 	}
2069 	/*
2070 	 * Combine connection status and session events. This is needed
2071 	 * for waiting two possible cases: cm_err has something meaningful
2072 	 * or session state was really changed to error by device removal.
2073 	 */
2074 	err = wait_event_interruptible_timeout(
2075 			clt_path->state_wq,
2076 			con->cm_err || clt_path->state != RTRS_CLT_CONNECTING,
2077 			msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2078 	if (err == 0 || err == -ERESTARTSYS) {
2079 		if (err == 0)
2080 			err = -ETIMEDOUT;
2081 		/* Timedout or interrupted */
2082 		goto errr;
2083 	}
2084 	if (con->cm_err < 0) {
2085 		err = con->cm_err;
2086 		goto errr;
2087 	}
2088 	if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTING) {
2089 		/* Device removal */
2090 		err = -ECONNABORTED;
2091 		goto errr;
2092 	}
2093 
2094 	return 0;
2095 
2096 errr:
2097 	stop_cm(con);
2098 	mutex_lock(&con->con_mutex);
2099 	destroy_con_cq_qp(con);
2100 	mutex_unlock(&con->con_mutex);
2101 destroy_cm:
2102 	destroy_cm(con);
2103 
2104 	return err;
2105 }
2106 
2107 static void rtrs_clt_path_up(struct rtrs_clt_path *clt_path)
2108 {
2109 	struct rtrs_clt_sess *clt = clt_path->clt;
2110 	int up;
2111 
2112 	/*
2113 	 * We can fire RECONNECTED event only when all paths were
2114 	 * connected on rtrs_clt_open(), then each was disconnected
2115 	 * and the first one connected again.  That's why this nasty
2116 	 * game with counter value.
2117 	 */
2118 
2119 	mutex_lock(&clt->paths_ev_mutex);
2120 	up = ++clt->paths_up;
2121 	/*
2122 	 * Here it is safe to access paths num directly since up counter
2123 	 * is greater than MAX_PATHS_NUM only while rtrs_clt_open() is
2124 	 * in progress, thus paths removals are impossible.
2125 	 */
2126 	if (up > MAX_PATHS_NUM && up == MAX_PATHS_NUM + clt->paths_num)
2127 		clt->paths_up = clt->paths_num;
2128 	else if (up == 1)
2129 		clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_RECONNECTED);
2130 	mutex_unlock(&clt->paths_ev_mutex);
2131 
2132 	/* Mark session as established */
2133 	clt_path->established = true;
2134 	clt_path->reconnect_attempts = 0;
2135 	clt_path->stats->reconnects.successful_cnt++;
2136 }
2137 
2138 static void rtrs_clt_path_down(struct rtrs_clt_path *clt_path)
2139 {
2140 	struct rtrs_clt_sess *clt = clt_path->clt;
2141 
2142 	if (!clt_path->established)
2143 		return;
2144 
2145 	clt_path->established = false;
2146 	mutex_lock(&clt->paths_ev_mutex);
2147 	WARN_ON(!clt->paths_up);
2148 	if (--clt->paths_up == 0)
2149 		clt->link_ev(clt->priv, RTRS_CLT_LINK_EV_DISCONNECTED);
2150 	mutex_unlock(&clt->paths_ev_mutex);
2151 }
2152 
2153 static void rtrs_clt_stop_and_destroy_conns(struct rtrs_clt_path *clt_path)
2154 {
2155 	struct rtrs_clt_con *con;
2156 	unsigned int cid;
2157 
2158 	WARN_ON(READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTED);
2159 
2160 	/*
2161 	 * Possible race with rtrs_clt_open(), when DEVICE_REMOVAL comes
2162 	 * exactly in between.  Start destroying after it finishes.
2163 	 */
2164 	mutex_lock(&clt_path->init_mutex);
2165 	mutex_unlock(&clt_path->init_mutex);
2166 
2167 	/*
2168 	 * All IO paths must observe !CONNECTED state before we
2169 	 * free everything.
2170 	 */
2171 	synchronize_rcu();
2172 
2173 	rtrs_stop_hb(&clt_path->s);
2174 
2175 	/*
2176 	 * The order it utterly crucial: firstly disconnect and complete all
2177 	 * rdma requests with error (thus set in_use=false for requests),
2178 	 * then fail outstanding requests checking in_use for each, and
2179 	 * eventually notify upper layer about session disconnection.
2180 	 */
2181 
2182 	for (cid = 0; cid < clt_path->s.con_num; cid++) {
2183 		if (!clt_path->s.con[cid])
2184 			break;
2185 		con = to_clt_con(clt_path->s.con[cid]);
2186 		stop_cm(con);
2187 	}
2188 	fail_all_outstanding_reqs(clt_path);
2189 	free_path_reqs(clt_path);
2190 	rtrs_clt_path_down(clt_path);
2191 
2192 	/*
2193 	 * Wait for graceful shutdown, namely when peer side invokes
2194 	 * rdma_disconnect(). 'connected_cnt' is decremented only on
2195 	 * CM events, thus if other side had crashed and hb has detected
2196 	 * something is wrong, here we will stuck for exactly timeout ms,
2197 	 * since CM does not fire anything.  That is fine, we are not in
2198 	 * hurry.
2199 	 */
2200 	wait_event_timeout(clt_path->state_wq,
2201 			   !atomic_read(&clt_path->connected_cnt),
2202 			   msecs_to_jiffies(RTRS_CONNECT_TIMEOUT_MS));
2203 
2204 	for (cid = 0; cid < clt_path->s.con_num; cid++) {
2205 		if (!clt_path->s.con[cid])
2206 			break;
2207 		con = to_clt_con(clt_path->s.con[cid]);
2208 		mutex_lock(&con->con_mutex);
2209 		destroy_con_cq_qp(con);
2210 		mutex_unlock(&con->con_mutex);
2211 		destroy_cm(con);
2212 		destroy_con(con);
2213 	}
2214 }
2215 
2216 static inline bool xchg_paths(struct rtrs_clt_path __rcu **rcu_ppcpu_path,
2217 			      struct rtrs_clt_path *clt_path,
2218 			      struct rtrs_clt_path *next)
2219 {
2220 	struct rtrs_clt_path **ppcpu_path;
2221 
2222 	/* Call cmpxchg() without sparse warnings */
2223 	ppcpu_path = (typeof(ppcpu_path))rcu_ppcpu_path;
2224 	return clt_path == cmpxchg(ppcpu_path, clt_path, next);
2225 }
2226 
2227 static void rtrs_clt_remove_path_from_arr(struct rtrs_clt_path *clt_path)
2228 {
2229 	struct rtrs_clt_sess *clt = clt_path->clt;
2230 	struct rtrs_clt_path *next;
2231 	bool wait_for_grace = false;
2232 	int cpu;
2233 
2234 	mutex_lock(&clt->paths_mutex);
2235 	list_del_rcu(&clt_path->s.entry);
2236 
2237 	/* Make sure everybody observes path removal. */
2238 	synchronize_rcu();
2239 
2240 	/*
2241 	 * At this point nobody sees @sess in the list, but still we have
2242 	 * dangling pointer @pcpu_path which _can_ point to @sess.  Since
2243 	 * nobody can observe @sess in the list, we guarantee that IO path
2244 	 * will not assign @sess to @pcpu_path, i.e. @pcpu_path can be equal
2245 	 * to @sess, but can never again become @sess.
2246 	 */
2247 
2248 	/*
2249 	 * Decrement paths number only after grace period, because
2250 	 * caller of do_each_path() must firstly observe list without
2251 	 * path and only then decremented paths number.
2252 	 *
2253 	 * Otherwise there can be the following situation:
2254 	 *    o Two paths exist and IO is coming.
2255 	 *    o One path is removed:
2256 	 *      CPU#0                          CPU#1
2257 	 *      do_each_path():                rtrs_clt_remove_path_from_arr():
2258 	 *          path = get_next_path()
2259 	 *          ^^^                            list_del_rcu(path)
2260 	 *          [!CONNECTED path]              clt->paths_num--
2261 	 *                                              ^^^^^^^^^
2262 	 *          load clt->paths_num                 from 2 to 1
2263 	 *                    ^^^^^^^^^
2264 	 *                    sees 1
2265 	 *
2266 	 *      path is observed as !CONNECTED, but do_each_path() loop
2267 	 *      ends, because expression i < clt->paths_num is false.
2268 	 */
2269 	clt->paths_num--;
2270 
2271 	/*
2272 	 * Get @next connection from current @sess which is going to be
2273 	 * removed.  If @sess is the last element, then @next is NULL.
2274 	 */
2275 	rcu_read_lock();
2276 	next = rtrs_clt_get_next_path_or_null(&clt->paths_list, clt_path);
2277 	rcu_read_unlock();
2278 
2279 	/*
2280 	 * @pcpu paths can still point to the path which is going to be
2281 	 * removed, so change the pointer manually.
2282 	 */
2283 	for_each_possible_cpu(cpu) {
2284 		struct rtrs_clt_path __rcu **ppcpu_path;
2285 
2286 		ppcpu_path = per_cpu_ptr(clt->pcpu_path, cpu);
2287 		if (rcu_dereference_protected(*ppcpu_path,
2288 			lockdep_is_held(&clt->paths_mutex)) != clt_path)
2289 			/*
2290 			 * synchronize_rcu() was called just after deleting
2291 			 * entry from the list, thus IO code path cannot
2292 			 * change pointer back to the pointer which is going
2293 			 * to be removed, we are safe here.
2294 			 */
2295 			continue;
2296 
2297 		/*
2298 		 * We race with IO code path, which also changes pointer,
2299 		 * thus we have to be careful not to overwrite it.
2300 		 */
2301 		if (xchg_paths(ppcpu_path, clt_path, next))
2302 			/*
2303 			 * @ppcpu_path was successfully replaced with @next,
2304 			 * that means that someone could also pick up the
2305 			 * @sess and dereferencing it right now, so wait for
2306 			 * a grace period is required.
2307 			 */
2308 			wait_for_grace = true;
2309 	}
2310 	if (wait_for_grace)
2311 		synchronize_rcu();
2312 
2313 	mutex_unlock(&clt->paths_mutex);
2314 }
2315 
2316 static void rtrs_clt_add_path_to_arr(struct rtrs_clt_path *clt_path)
2317 {
2318 	struct rtrs_clt_sess *clt = clt_path->clt;
2319 
2320 	mutex_lock(&clt->paths_mutex);
2321 	clt->paths_num++;
2322 
2323 	list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2324 	mutex_unlock(&clt->paths_mutex);
2325 }
2326 
2327 static void rtrs_clt_close_work(struct work_struct *work)
2328 {
2329 	struct rtrs_clt_path *clt_path;
2330 
2331 	clt_path = container_of(work, struct rtrs_clt_path, close_work);
2332 
2333 	cancel_work_sync(&clt_path->err_recovery_work);
2334 	cancel_delayed_work_sync(&clt_path->reconnect_dwork);
2335 	rtrs_clt_stop_and_destroy_conns(clt_path);
2336 	rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CLOSED, NULL);
2337 }
2338 
2339 static int init_conns(struct rtrs_clt_path *clt_path)
2340 {
2341 	unsigned int cid;
2342 	int err;
2343 
2344 	/*
2345 	 * On every new session connections increase reconnect counter
2346 	 * to avoid clashes with previous sessions not yet closed
2347 	 * sessions on a server side.
2348 	 */
2349 	clt_path->s.recon_cnt++;
2350 
2351 	/* Establish all RDMA connections  */
2352 	for (cid = 0; cid < clt_path->s.con_num; cid++) {
2353 		err = create_con(clt_path, cid);
2354 		if (err)
2355 			goto destroy;
2356 
2357 		err = create_cm(to_clt_con(clt_path->s.con[cid]));
2358 		if (err) {
2359 			destroy_con(to_clt_con(clt_path->s.con[cid]));
2360 			goto destroy;
2361 		}
2362 	}
2363 	err = alloc_path_reqs(clt_path);
2364 	if (err)
2365 		goto destroy;
2366 
2367 	rtrs_start_hb(&clt_path->s);
2368 
2369 	return 0;
2370 
2371 destroy:
2372 	while (cid--) {
2373 		struct rtrs_clt_con *con = to_clt_con(clt_path->s.con[cid]);
2374 
2375 		stop_cm(con);
2376 
2377 		mutex_lock(&con->con_mutex);
2378 		destroy_con_cq_qp(con);
2379 		mutex_unlock(&con->con_mutex);
2380 		destroy_cm(con);
2381 		destroy_con(con);
2382 	}
2383 	/*
2384 	 * If we've never taken async path and got an error, say,
2385 	 * doing rdma_resolve_addr(), switch to CONNECTION_ERR state
2386 	 * manually to keep reconnecting.
2387 	 */
2388 	rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2389 
2390 	return err;
2391 }
2392 
2393 static void rtrs_clt_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
2394 {
2395 	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2396 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2397 	struct rtrs_iu *iu;
2398 
2399 	iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2400 	rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2401 
2402 	if (wc->status != IB_WC_SUCCESS) {
2403 		rtrs_err(clt_path->clt, "Path info request send failed: %s\n",
2404 			  ib_wc_status_msg(wc->status));
2405 		rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING_ERR, NULL);
2406 		return;
2407 	}
2408 
2409 	rtrs_clt_update_wc_stats(con);
2410 }
2411 
2412 static int process_info_rsp(struct rtrs_clt_path *clt_path,
2413 			    const struct rtrs_msg_info_rsp *msg)
2414 {
2415 	unsigned int sg_cnt, total_len;
2416 	int i, sgi;
2417 
2418 	sg_cnt = le16_to_cpu(msg->sg_cnt);
2419 	if (!sg_cnt || (clt_path->queue_depth % sg_cnt)) {
2420 		rtrs_err(clt_path->clt,
2421 			  "Incorrect sg_cnt %d, is not multiple\n",
2422 			  sg_cnt);
2423 		return -EINVAL;
2424 	}
2425 
2426 	/*
2427 	 * Check if IB immediate data size is enough to hold the mem_id and
2428 	 * the offset inside the memory chunk.
2429 	 */
2430 	if ((ilog2(sg_cnt - 1) + 1) + (ilog2(clt_path->chunk_size - 1) + 1) >
2431 	    MAX_IMM_PAYL_BITS) {
2432 		rtrs_err(clt_path->clt,
2433 			  "RDMA immediate size (%db) not enough to encode %d buffers of size %dB\n",
2434 			  MAX_IMM_PAYL_BITS, sg_cnt, clt_path->chunk_size);
2435 		return -EINVAL;
2436 	}
2437 	total_len = 0;
2438 	for (sgi = 0, i = 0; sgi < sg_cnt && i < clt_path->queue_depth; sgi++) {
2439 		const struct rtrs_sg_desc *desc = &msg->desc[sgi];
2440 		u32 len, rkey;
2441 		u64 addr;
2442 
2443 		addr = le64_to_cpu(desc->addr);
2444 		rkey = le32_to_cpu(desc->key);
2445 		len  = le32_to_cpu(desc->len);
2446 
2447 		total_len += len;
2448 
2449 		if (!len || (len % clt_path->chunk_size)) {
2450 			rtrs_err(clt_path->clt, "Incorrect [%d].len %d\n",
2451 				  sgi,
2452 				  len);
2453 			return -EINVAL;
2454 		}
2455 		for ( ; len && i < clt_path->queue_depth; i++) {
2456 			clt_path->rbufs[i].addr = addr;
2457 			clt_path->rbufs[i].rkey = rkey;
2458 
2459 			len  -= clt_path->chunk_size;
2460 			addr += clt_path->chunk_size;
2461 		}
2462 	}
2463 	/* Sanity check */
2464 	if (sgi != sg_cnt || i != clt_path->queue_depth) {
2465 		rtrs_err(clt_path->clt,
2466 			 "Incorrect sg vector, not fully mapped\n");
2467 		return -EINVAL;
2468 	}
2469 	if (total_len != clt_path->chunk_size * clt_path->queue_depth) {
2470 		rtrs_err(clt_path->clt, "Incorrect total_len %d\n", total_len);
2471 		return -EINVAL;
2472 	}
2473 
2474 	return 0;
2475 }
2476 
2477 static void rtrs_clt_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
2478 {
2479 	struct rtrs_clt_con *con = to_clt_con(wc->qp->qp_context);
2480 	struct rtrs_clt_path *clt_path = to_clt_path(con->c.path);
2481 	struct rtrs_msg_info_rsp *msg;
2482 	enum rtrs_clt_state state;
2483 	struct rtrs_iu *iu;
2484 	size_t rx_sz;
2485 	int err;
2486 
2487 	state = RTRS_CLT_CONNECTING_ERR;
2488 
2489 	WARN_ON(con->c.cid);
2490 	iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
2491 	if (wc->status != IB_WC_SUCCESS) {
2492 		rtrs_err(clt_path->clt, "Path info response recv failed: %s\n",
2493 			  ib_wc_status_msg(wc->status));
2494 		goto out;
2495 	}
2496 	WARN_ON(wc->opcode != IB_WC_RECV);
2497 
2498 	if (wc->byte_len < sizeof(*msg)) {
2499 		rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2500 			  wc->byte_len);
2501 		goto out;
2502 	}
2503 	ib_dma_sync_single_for_cpu(clt_path->s.dev->ib_dev, iu->dma_addr,
2504 				   iu->size, DMA_FROM_DEVICE);
2505 	msg = iu->buf;
2506 	if (le16_to_cpu(msg->type) != RTRS_MSG_INFO_RSP) {
2507 		rtrs_err(clt_path->clt, "Path info response is malformed: type %d\n",
2508 			  le16_to_cpu(msg->type));
2509 		goto out;
2510 	}
2511 	rx_sz  = sizeof(*msg);
2512 	rx_sz += sizeof(msg->desc[0]) * le16_to_cpu(msg->sg_cnt);
2513 	if (wc->byte_len < rx_sz) {
2514 		rtrs_err(clt_path->clt, "Path info response is malformed: size %d\n",
2515 			  wc->byte_len);
2516 		goto out;
2517 	}
2518 	err = process_info_rsp(clt_path, msg);
2519 	if (err)
2520 		goto out;
2521 
2522 	err = post_recv_path(clt_path);
2523 	if (err)
2524 		goto out;
2525 
2526 	state = RTRS_CLT_CONNECTED;
2527 
2528 out:
2529 	rtrs_clt_update_wc_stats(con);
2530 	rtrs_iu_free(iu, clt_path->s.dev->ib_dev, 1);
2531 	rtrs_clt_change_state_get_old(clt_path, state, NULL);
2532 }
2533 
2534 static int rtrs_send_path_info(struct rtrs_clt_path *clt_path)
2535 {
2536 	struct rtrs_clt_con *usr_con = to_clt_con(clt_path->s.con[0]);
2537 	struct rtrs_msg_info_req *msg;
2538 	struct rtrs_iu *tx_iu, *rx_iu;
2539 	size_t rx_sz;
2540 	int err;
2541 
2542 	rx_sz  = sizeof(struct rtrs_msg_info_rsp);
2543 	rx_sz += sizeof(struct rtrs_sg_desc) * clt_path->queue_depth;
2544 
2545 	tx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req), GFP_KERNEL,
2546 			       clt_path->s.dev->ib_dev, DMA_TO_DEVICE,
2547 			       rtrs_clt_info_req_done);
2548 	rx_iu = rtrs_iu_alloc(1, rx_sz, GFP_KERNEL, clt_path->s.dev->ib_dev,
2549 			       DMA_FROM_DEVICE, rtrs_clt_info_rsp_done);
2550 	if (!tx_iu || !rx_iu) {
2551 		err = -ENOMEM;
2552 		goto out;
2553 	}
2554 	/* Prepare for getting info response */
2555 	err = rtrs_iu_post_recv(&usr_con->c, rx_iu);
2556 	if (err) {
2557 		rtrs_err(clt_path->clt, "rtrs_iu_post_recv(), err: %d\n", err);
2558 		goto out;
2559 	}
2560 	rx_iu = NULL;
2561 
2562 	msg = tx_iu->buf;
2563 	msg->type = cpu_to_le16(RTRS_MSG_INFO_REQ);
2564 	memcpy(msg->pathname, clt_path->s.sessname, sizeof(msg->pathname));
2565 
2566 	ib_dma_sync_single_for_device(clt_path->s.dev->ib_dev,
2567 				      tx_iu->dma_addr,
2568 				      tx_iu->size, DMA_TO_DEVICE);
2569 
2570 	/* Send info request */
2571 	err = rtrs_iu_post_send(&usr_con->c, tx_iu, sizeof(*msg), NULL);
2572 	if (err) {
2573 		rtrs_err(clt_path->clt, "rtrs_iu_post_send(), err: %d\n", err);
2574 		goto out;
2575 	}
2576 	tx_iu = NULL;
2577 
2578 	/* Wait for state change */
2579 	wait_event_interruptible_timeout(clt_path->state_wq,
2580 					 clt_path->state != RTRS_CLT_CONNECTING,
2581 					 msecs_to_jiffies(
2582 						 RTRS_CONNECT_TIMEOUT_MS));
2583 	if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED) {
2584 		if (READ_ONCE(clt_path->state) == RTRS_CLT_CONNECTING_ERR)
2585 			err = -ECONNRESET;
2586 		else
2587 			err = -ETIMEDOUT;
2588 	}
2589 
2590 out:
2591 	if (tx_iu)
2592 		rtrs_iu_free(tx_iu, clt_path->s.dev->ib_dev, 1);
2593 	if (rx_iu)
2594 		rtrs_iu_free(rx_iu, clt_path->s.dev->ib_dev, 1);
2595 	if (err)
2596 		/* If we've never taken async path because of malloc problems */
2597 		rtrs_clt_change_state_get_old(clt_path,
2598 					      RTRS_CLT_CONNECTING_ERR, NULL);
2599 
2600 	return err;
2601 }
2602 
2603 /**
2604  * init_path() - establishes all path connections and does handshake
2605  * @clt_path: client path.
2606  * In case of error full close or reconnect procedure should be taken,
2607  * because reconnect or close async works can be started.
2608  */
2609 static int init_path(struct rtrs_clt_path *clt_path)
2610 {
2611 	int err;
2612 	char str[NAME_MAX];
2613 	struct rtrs_addr path = {
2614 		.src = &clt_path->s.src_addr,
2615 		.dst = &clt_path->s.dst_addr,
2616 	};
2617 
2618 	rtrs_addr_to_str(&path, str, sizeof(str));
2619 
2620 	mutex_lock(&clt_path->init_mutex);
2621 	err = init_conns(clt_path);
2622 	if (err) {
2623 		rtrs_err(clt_path->clt,
2624 			 "init_conns() failed: err=%d path=%s [%s:%u]\n", err,
2625 			 str, clt_path->hca_name, clt_path->hca_port);
2626 		goto out;
2627 	}
2628 	err = rtrs_send_path_info(clt_path);
2629 	if (err) {
2630 		rtrs_err(clt_path->clt,
2631 			 "rtrs_send_path_info() failed: err=%d path=%s [%s:%u]\n",
2632 			 err, str, clt_path->hca_name, clt_path->hca_port);
2633 		goto out;
2634 	}
2635 	rtrs_clt_path_up(clt_path);
2636 out:
2637 	mutex_unlock(&clt_path->init_mutex);
2638 
2639 	return err;
2640 }
2641 
2642 static void rtrs_clt_reconnect_work(struct work_struct *work)
2643 {
2644 	struct rtrs_clt_path *clt_path;
2645 	struct rtrs_clt_sess *clt;
2646 	int err;
2647 
2648 	clt_path = container_of(to_delayed_work(work), struct rtrs_clt_path,
2649 				reconnect_dwork);
2650 	clt = clt_path->clt;
2651 
2652 	if (READ_ONCE(clt_path->state) != RTRS_CLT_RECONNECTING)
2653 		return;
2654 
2655 	if (clt_path->reconnect_attempts >= clt->max_reconnect_attempts) {
2656 		/* Close a path completely if max attempts is reached */
2657 		rtrs_clt_close_conns(clt_path, false);
2658 		return;
2659 	}
2660 	clt_path->reconnect_attempts++;
2661 
2662 	msleep(RTRS_RECONNECT_BACKOFF);
2663 	if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_CONNECTING, NULL)) {
2664 		err = init_path(clt_path);
2665 		if (err)
2666 			goto reconnect_again;
2667 	}
2668 
2669 	return;
2670 
2671 reconnect_again:
2672 	if (rtrs_clt_change_state_get_old(clt_path, RTRS_CLT_RECONNECTING, NULL)) {
2673 		clt_path->stats->reconnects.fail_cnt++;
2674 		queue_work(rtrs_wq, &clt_path->err_recovery_work);
2675 	}
2676 }
2677 
2678 static void rtrs_clt_dev_release(struct device *dev)
2679 {
2680 	struct rtrs_clt_sess *clt = container_of(dev, struct rtrs_clt_sess,
2681 						 dev);
2682 
2683 	mutex_destroy(&clt->paths_ev_mutex);
2684 	mutex_destroy(&clt->paths_mutex);
2685 	kfree(clt);
2686 }
2687 
2688 static struct rtrs_clt_sess *alloc_clt(const char *sessname, size_t paths_num,
2689 				  u16 port, size_t pdu_sz, void *priv,
2690 				  void	(*link_ev)(void *priv,
2691 						   enum rtrs_clt_link_ev ev),
2692 				  unsigned int reconnect_delay_sec,
2693 				  unsigned int max_reconnect_attempts)
2694 {
2695 	struct rtrs_clt_sess *clt;
2696 	int err;
2697 
2698 	if (!paths_num || paths_num > MAX_PATHS_NUM)
2699 		return ERR_PTR(-EINVAL);
2700 
2701 	if (strlen(sessname) >= sizeof(clt->sessname))
2702 		return ERR_PTR(-EINVAL);
2703 
2704 	clt = kzalloc(sizeof(*clt), GFP_KERNEL);
2705 	if (!clt)
2706 		return ERR_PTR(-ENOMEM);
2707 
2708 	clt->pcpu_path = alloc_percpu(typeof(*clt->pcpu_path));
2709 	if (!clt->pcpu_path) {
2710 		kfree(clt);
2711 		return ERR_PTR(-ENOMEM);
2712 	}
2713 
2714 	clt->dev.class = rtrs_clt_dev_class;
2715 	clt->dev.release = rtrs_clt_dev_release;
2716 	uuid_gen(&clt->paths_uuid);
2717 	INIT_LIST_HEAD_RCU(&clt->paths_list);
2718 	clt->paths_num = paths_num;
2719 	clt->paths_up = MAX_PATHS_NUM;
2720 	clt->port = port;
2721 	clt->pdu_sz = pdu_sz;
2722 	clt->max_segments = RTRS_MAX_SEGMENTS;
2723 	clt->reconnect_delay_sec = reconnect_delay_sec;
2724 	clt->max_reconnect_attempts = max_reconnect_attempts;
2725 	clt->priv = priv;
2726 	clt->link_ev = link_ev;
2727 	clt->mp_policy = MP_POLICY_MIN_INFLIGHT;
2728 	strscpy(clt->sessname, sessname, sizeof(clt->sessname));
2729 	init_waitqueue_head(&clt->permits_wait);
2730 	mutex_init(&clt->paths_ev_mutex);
2731 	mutex_init(&clt->paths_mutex);
2732 	device_initialize(&clt->dev);
2733 
2734 	err = dev_set_name(&clt->dev, "%s", sessname);
2735 	if (err)
2736 		goto err_put;
2737 
2738 	/*
2739 	 * Suppress user space notification until
2740 	 * sysfs files are created
2741 	 */
2742 	dev_set_uevent_suppress(&clt->dev, true);
2743 	err = device_add(&clt->dev);
2744 	if (err)
2745 		goto err_put;
2746 
2747 	clt->kobj_paths = kobject_create_and_add("paths", &clt->dev.kobj);
2748 	if (!clt->kobj_paths) {
2749 		err = -ENOMEM;
2750 		goto err_del;
2751 	}
2752 	err = rtrs_clt_create_sysfs_root_files(clt);
2753 	if (err) {
2754 		kobject_del(clt->kobj_paths);
2755 		kobject_put(clt->kobj_paths);
2756 		goto err_del;
2757 	}
2758 	dev_set_uevent_suppress(&clt->dev, false);
2759 	kobject_uevent(&clt->dev.kobj, KOBJ_ADD);
2760 
2761 	return clt;
2762 err_del:
2763 	device_del(&clt->dev);
2764 err_put:
2765 	free_percpu(clt->pcpu_path);
2766 	put_device(&clt->dev);
2767 	return ERR_PTR(err);
2768 }
2769 
2770 static void free_clt(struct rtrs_clt_sess *clt)
2771 {
2772 	free_percpu(clt->pcpu_path);
2773 
2774 	/*
2775 	 * release callback will free clt and destroy mutexes in last put
2776 	 */
2777 	device_unregister(&clt->dev);
2778 }
2779 
2780 /**
2781  * rtrs_clt_open() - Open a path to an RTRS server
2782  * @ops: holds the link event callback and the private pointer.
2783  * @pathname: name of the path to an RTRS server
2784  * @paths: Paths to be established defined by their src and dst addresses
2785  * @paths_num: Number of elements in the @paths array
2786  * @port: port to be used by the RTRS session
2787  * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
2788  * @reconnect_delay_sec: time between reconnect tries
2789  * @max_reconnect_attempts: Number of times to reconnect on error before giving
2790  *			    up, 0 for * disabled, -1 for forever
2791  * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
2792  *
2793  * Starts session establishment with the rtrs_server. The function can block
2794  * up to ~2000ms before it returns.
2795  *
2796  * Return a valid pointer on success otherwise PTR_ERR.
2797  */
2798 struct rtrs_clt_sess *rtrs_clt_open(struct rtrs_clt_ops *ops,
2799 				 const char *pathname,
2800 				 const struct rtrs_addr *paths,
2801 				 size_t paths_num, u16 port,
2802 				 size_t pdu_sz, u8 reconnect_delay_sec,
2803 				 s16 max_reconnect_attempts, u32 nr_poll_queues)
2804 {
2805 	struct rtrs_clt_path *clt_path, *tmp;
2806 	struct rtrs_clt_sess *clt;
2807 	int err, i;
2808 
2809 	if (strchr(pathname, '/') || strchr(pathname, '.')) {
2810 		pr_err("pathname cannot contain / and .\n");
2811 		err = -EINVAL;
2812 		goto out;
2813 	}
2814 
2815 	clt = alloc_clt(pathname, paths_num, port, pdu_sz, ops->priv,
2816 			ops->link_ev,
2817 			reconnect_delay_sec,
2818 			max_reconnect_attempts);
2819 	if (IS_ERR(clt)) {
2820 		err = PTR_ERR(clt);
2821 		goto out;
2822 	}
2823 	for (i = 0; i < paths_num; i++) {
2824 		struct rtrs_clt_path *clt_path;
2825 
2826 		clt_path = alloc_path(clt, &paths[i], nr_cpu_ids,
2827 				  nr_poll_queues);
2828 		if (IS_ERR(clt_path)) {
2829 			err = PTR_ERR(clt_path);
2830 			goto close_all_path;
2831 		}
2832 		if (!i)
2833 			clt_path->for_new_clt = 1;
2834 		list_add_tail_rcu(&clt_path->s.entry, &clt->paths_list);
2835 
2836 		err = init_path(clt_path);
2837 		if (err) {
2838 			list_del_rcu(&clt_path->s.entry);
2839 			rtrs_clt_close_conns(clt_path, true);
2840 			free_percpu(clt_path->stats->pcpu_stats);
2841 			kfree(clt_path->stats);
2842 			free_path(clt_path);
2843 			goto close_all_path;
2844 		}
2845 
2846 		err = rtrs_clt_create_path_files(clt_path);
2847 		if (err) {
2848 			list_del_rcu(&clt_path->s.entry);
2849 			rtrs_clt_close_conns(clt_path, true);
2850 			free_percpu(clt_path->stats->pcpu_stats);
2851 			kfree(clt_path->stats);
2852 			free_path(clt_path);
2853 			goto close_all_path;
2854 		}
2855 	}
2856 	err = alloc_permits(clt);
2857 	if (err)
2858 		goto close_all_path;
2859 
2860 	return clt;
2861 
2862 close_all_path:
2863 	list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2864 		rtrs_clt_destroy_path_files(clt_path, NULL);
2865 		rtrs_clt_close_conns(clt_path, true);
2866 		kobject_put(&clt_path->kobj);
2867 	}
2868 	rtrs_clt_destroy_sysfs_root(clt);
2869 	free_clt(clt);
2870 
2871 out:
2872 	return ERR_PTR(err);
2873 }
2874 EXPORT_SYMBOL(rtrs_clt_open);
2875 
2876 /**
2877  * rtrs_clt_close() - Close a path
2878  * @clt: Session handle. Session is freed upon return.
2879  */
2880 void rtrs_clt_close(struct rtrs_clt_sess *clt)
2881 {
2882 	struct rtrs_clt_path *clt_path, *tmp;
2883 
2884 	/* Firstly forbid sysfs access */
2885 	rtrs_clt_destroy_sysfs_root(clt);
2886 
2887 	/* Now it is safe to iterate over all paths without locks */
2888 	list_for_each_entry_safe(clt_path, tmp, &clt->paths_list, s.entry) {
2889 		rtrs_clt_close_conns(clt_path, true);
2890 		rtrs_clt_destroy_path_files(clt_path, NULL);
2891 		kobject_put(&clt_path->kobj);
2892 	}
2893 	free_permits(clt);
2894 	free_clt(clt);
2895 }
2896 EXPORT_SYMBOL(rtrs_clt_close);
2897 
2898 int rtrs_clt_reconnect_from_sysfs(struct rtrs_clt_path *clt_path)
2899 {
2900 	enum rtrs_clt_state old_state;
2901 	int err = -EBUSY;
2902 	bool changed;
2903 
2904 	changed = rtrs_clt_change_state_get_old(clt_path,
2905 						 RTRS_CLT_RECONNECTING,
2906 						 &old_state);
2907 	if (changed) {
2908 		clt_path->reconnect_attempts = 0;
2909 		rtrs_clt_stop_and_destroy_conns(clt_path);
2910 		queue_delayed_work(rtrs_wq, &clt_path->reconnect_dwork, 0);
2911 	}
2912 	if (changed || old_state == RTRS_CLT_RECONNECTING) {
2913 		/*
2914 		 * flush_delayed_work() queues pending work for immediate
2915 		 * execution, so do the flush if we have queued something
2916 		 * right now or work is pending.
2917 		 */
2918 		flush_delayed_work(&clt_path->reconnect_dwork);
2919 		err = (READ_ONCE(clt_path->state) ==
2920 		       RTRS_CLT_CONNECTED ? 0 : -ENOTCONN);
2921 	}
2922 
2923 	return err;
2924 }
2925 
2926 int rtrs_clt_remove_path_from_sysfs(struct rtrs_clt_path *clt_path,
2927 				     const struct attribute *sysfs_self)
2928 {
2929 	enum rtrs_clt_state old_state;
2930 	bool changed;
2931 
2932 	/*
2933 	 * Continue stopping path till state was changed to DEAD or
2934 	 * state was observed as DEAD:
2935 	 * 1. State was changed to DEAD - we were fast and nobody
2936 	 *    invoked rtrs_clt_reconnect(), which can again start
2937 	 *    reconnecting.
2938 	 * 2. State was observed as DEAD - we have someone in parallel
2939 	 *    removing the path.
2940 	 */
2941 	do {
2942 		rtrs_clt_close_conns(clt_path, true);
2943 		changed = rtrs_clt_change_state_get_old(clt_path,
2944 							RTRS_CLT_DEAD,
2945 							&old_state);
2946 	} while (!changed && old_state != RTRS_CLT_DEAD);
2947 
2948 	if (changed) {
2949 		rtrs_clt_remove_path_from_arr(clt_path);
2950 		rtrs_clt_destroy_path_files(clt_path, sysfs_self);
2951 		kobject_put(&clt_path->kobj);
2952 	}
2953 
2954 	return 0;
2955 }
2956 
2957 void rtrs_clt_set_max_reconnect_attempts(struct rtrs_clt_sess *clt, int value)
2958 {
2959 	clt->max_reconnect_attempts = (unsigned int)value;
2960 }
2961 
2962 int rtrs_clt_get_max_reconnect_attempts(const struct rtrs_clt_sess *clt)
2963 {
2964 	return (int)clt->max_reconnect_attempts;
2965 }
2966 
2967 /**
2968  * rtrs_clt_request() - Request data transfer to/from server via RDMA.
2969  *
2970  * @dir:	READ/WRITE
2971  * @ops:	callback function to be called as confirmation, and the pointer.
2972  * @clt:	Session
2973  * @permit:	Preallocated permit
2974  * @vec:	Message that is sent to server together with the request.
2975  *		Sum of len of all @vec elements limited to <= IO_MSG_SIZE.
2976  *		Since the msg is copied internally it can be allocated on stack.
2977  * @nr:		Number of elements in @vec.
2978  * @data_len:	length of data sent to/from server
2979  * @sg:		Pages to be sent/received to/from server.
2980  * @sg_cnt:	Number of elements in the @sg
2981  *
2982  * Return:
2983  * 0:		Success
2984  * <0:		Error
2985  *
2986  * On dir=READ rtrs client will request a data transfer from Server to client.
2987  * The data that the server will respond with will be stored in @sg when
2988  * the user receives an %RTRS_CLT_RDMA_EV_RDMA_REQUEST_WRITE_COMPL event.
2989  * On dir=WRITE rtrs client will rdma write data in sg to server side.
2990  */
2991 int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
2992 		     struct rtrs_clt_sess *clt, struct rtrs_permit *permit,
2993 		     const struct kvec *vec, size_t nr, size_t data_len,
2994 		     struct scatterlist *sg, unsigned int sg_cnt)
2995 {
2996 	struct rtrs_clt_io_req *req;
2997 	struct rtrs_clt_path *clt_path;
2998 
2999 	enum dma_data_direction dma_dir;
3000 	int err = -ECONNABORTED, i;
3001 	size_t usr_len, hdr_len;
3002 	struct path_it it;
3003 
3004 	/* Get kvec length */
3005 	for (i = 0, usr_len = 0; i < nr; i++)
3006 		usr_len += vec[i].iov_len;
3007 
3008 	if (dir == READ) {
3009 		hdr_len = sizeof(struct rtrs_msg_rdma_read) +
3010 			  sg_cnt * sizeof(struct rtrs_sg_desc);
3011 		dma_dir = DMA_FROM_DEVICE;
3012 	} else {
3013 		hdr_len = sizeof(struct rtrs_msg_rdma_write);
3014 		dma_dir = DMA_TO_DEVICE;
3015 	}
3016 
3017 	rcu_read_lock();
3018 	for (path_it_init(&it, clt);
3019 	     (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3020 		if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3021 			continue;
3022 
3023 		if (usr_len + hdr_len > clt_path->max_hdr_size) {
3024 			rtrs_wrn_rl(clt_path->clt,
3025 				     "%s request failed, user message size is %zu and header length %zu, but max size is %u\n",
3026 				     dir == READ ? "Read" : "Write",
3027 				     usr_len, hdr_len, clt_path->max_hdr_size);
3028 			err = -EMSGSIZE;
3029 			break;
3030 		}
3031 		req = rtrs_clt_get_req(clt_path, ops->conf_fn, permit, ops->priv,
3032 				       vec, usr_len, sg, sg_cnt, data_len,
3033 				       dma_dir);
3034 		if (dir == READ)
3035 			err = rtrs_clt_read_req(req);
3036 		else
3037 			err = rtrs_clt_write_req(req);
3038 		if (err) {
3039 			req->in_use = false;
3040 			continue;
3041 		}
3042 		/* Success path */
3043 		break;
3044 	}
3045 	path_it_deinit(&it);
3046 	rcu_read_unlock();
3047 
3048 	return err;
3049 }
3050 EXPORT_SYMBOL(rtrs_clt_request);
3051 
3052 int rtrs_clt_rdma_cq_direct(struct rtrs_clt_sess *clt, unsigned int index)
3053 {
3054 	/* If no path, return -1 for block layer not to try again */
3055 	int cnt = -1;
3056 	struct rtrs_con *con;
3057 	struct rtrs_clt_path *clt_path;
3058 	struct path_it it;
3059 
3060 	rcu_read_lock();
3061 	for (path_it_init(&it, clt);
3062 	     (clt_path = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
3063 		if (READ_ONCE(clt_path->state) != RTRS_CLT_CONNECTED)
3064 			continue;
3065 
3066 		con = clt_path->s.con[index + 1];
3067 		cnt = ib_process_cq_direct(con->cq, -1);
3068 		if (cnt)
3069 			break;
3070 	}
3071 	path_it_deinit(&it);
3072 	rcu_read_unlock();
3073 
3074 	return cnt;
3075 }
3076 EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
3077 
3078 /**
3079  * rtrs_clt_query() - queries RTRS session attributes
3080  *@clt: session pointer
3081  *@attr: query results for session attributes.
3082  * Returns:
3083  *    0 on success
3084  *    -ECOMM		no connection to the server
3085  */
3086 int rtrs_clt_query(struct rtrs_clt_sess *clt, struct rtrs_attrs *attr)
3087 {
3088 	if (!rtrs_clt_is_connected(clt))
3089 		return -ECOMM;
3090 
3091 	attr->queue_depth      = clt->queue_depth;
3092 	attr->max_segments     = clt->max_segments;
3093 	/* Cap max_io_size to min of remote buffer size and the fr pages */
3094 	attr->max_io_size = min_t(int, clt->max_io_size,
3095 				  clt->max_segments * SZ_4K);
3096 
3097 	return 0;
3098 }
3099 EXPORT_SYMBOL(rtrs_clt_query);
3100 
3101 int rtrs_clt_create_path_from_sysfs(struct rtrs_clt_sess *clt,
3102 				     struct rtrs_addr *addr)
3103 {
3104 	struct rtrs_clt_path *clt_path;
3105 	int err;
3106 
3107 	clt_path = alloc_path(clt, addr, nr_cpu_ids, 0);
3108 	if (IS_ERR(clt_path))
3109 		return PTR_ERR(clt_path);
3110 
3111 	mutex_lock(&clt->paths_mutex);
3112 	if (clt->paths_num == 0) {
3113 		/*
3114 		 * When all the paths are removed for a session,
3115 		 * the addition of the first path is like a new session for
3116 		 * the storage server
3117 		 */
3118 		clt_path->for_new_clt = 1;
3119 	}
3120 
3121 	mutex_unlock(&clt->paths_mutex);
3122 
3123 	/*
3124 	 * It is totally safe to add path in CONNECTING state: coming
3125 	 * IO will never grab it.  Also it is very important to add
3126 	 * path before init, since init fires LINK_CONNECTED event.
3127 	 */
3128 	rtrs_clt_add_path_to_arr(clt_path);
3129 
3130 	err = init_path(clt_path);
3131 	if (err)
3132 		goto close_path;
3133 
3134 	err = rtrs_clt_create_path_files(clt_path);
3135 	if (err)
3136 		goto close_path;
3137 
3138 	return 0;
3139 
3140 close_path:
3141 	rtrs_clt_remove_path_from_arr(clt_path);
3142 	rtrs_clt_close_conns(clt_path, true);
3143 	free_percpu(clt_path->stats->pcpu_stats);
3144 	kfree(clt_path->stats);
3145 	free_path(clt_path);
3146 
3147 	return err;
3148 }
3149 
3150 static int rtrs_clt_ib_dev_init(struct rtrs_ib_dev *dev)
3151 {
3152 	if (!(dev->ib_dev->attrs.device_cap_flags &
3153 	      IB_DEVICE_MEM_MGT_EXTENSIONS)) {
3154 		pr_err("Memory registrations not supported.\n");
3155 		return -ENOTSUPP;
3156 	}
3157 
3158 	return 0;
3159 }
3160 
3161 static const struct rtrs_rdma_dev_pd_ops dev_pd_ops = {
3162 	.init = rtrs_clt_ib_dev_init
3163 };
3164 
3165 static int __init rtrs_client_init(void)
3166 {
3167 	rtrs_rdma_dev_pd_init(0, &dev_pd);
3168 
3169 	rtrs_clt_dev_class = class_create(THIS_MODULE, "rtrs-client");
3170 	if (IS_ERR(rtrs_clt_dev_class)) {
3171 		pr_err("Failed to create rtrs-client dev class\n");
3172 		return PTR_ERR(rtrs_clt_dev_class);
3173 	}
3174 	rtrs_wq = alloc_workqueue("rtrs_client_wq", 0, 0);
3175 	if (!rtrs_wq) {
3176 		class_destroy(rtrs_clt_dev_class);
3177 		return -ENOMEM;
3178 	}
3179 
3180 	return 0;
3181 }
3182 
3183 static void __exit rtrs_client_exit(void)
3184 {
3185 	destroy_workqueue(rtrs_wq);
3186 	class_destroy(rtrs_clt_dev_class);
3187 	rtrs_rdma_dev_pd_deinit(&dev_pd);
3188 }
3189 
3190 module_init(rtrs_client_init);
3191 module_exit(rtrs_client_exit);
3192