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