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/mempool.h>
15 
16 #include "rtrs-srv.h"
17 #include "rtrs-log.h"
18 #include <rdma/ib_cm.h>
19 
20 MODULE_DESCRIPTION("RDMA Transport Server");
21 MODULE_LICENSE("GPL");
22 
23 /* Must be power of 2, see mask from mr->page_size in ib_sg_to_pages() */
24 #define DEFAULT_MAX_CHUNK_SIZE (128 << 10)
25 #define DEFAULT_SESS_QUEUE_DEPTH 512
26 #define MAX_HDR_SIZE PAGE_SIZE
27 
28 /* We guarantee to serve 10 paths at least */
29 #define CHUNK_POOL_SZ 10
30 
31 static struct rtrs_rdma_dev_pd dev_pd;
32 static mempool_t *chunk_pool;
33 struct class *rtrs_dev_class;
34 
35 static int __read_mostly max_chunk_size = DEFAULT_MAX_CHUNK_SIZE;
36 static int __read_mostly sess_queue_depth = DEFAULT_SESS_QUEUE_DEPTH;
37 
38 static bool always_invalidate = true;
39 module_param(always_invalidate, bool, 0444);
40 MODULE_PARM_DESC(always_invalidate,
41 		 "Invalidate memory registration for contiguous memory regions before accessing.");
42 
43 module_param_named(max_chunk_size, max_chunk_size, int, 0444);
44 MODULE_PARM_DESC(max_chunk_size,
45 		 "Max size for each IO request, when change the unit is in byte (default: "
46 		 __stringify(DEFAULT_MAX_CHUNK_SIZE) "KB)");
47 
48 module_param_named(sess_queue_depth, sess_queue_depth, int, 0444);
49 MODULE_PARM_DESC(sess_queue_depth,
50 		 "Number of buffers for pending I/O requests to allocate per session. Maximum: "
51 		 __stringify(MAX_SESS_QUEUE_DEPTH) " (default: "
52 		 __stringify(DEFAULT_SESS_QUEUE_DEPTH) ")");
53 
54 static cpumask_t cq_affinity_mask = { CPU_BITS_ALL };
55 
56 static struct workqueue_struct *rtrs_wq;
57 
58 static inline struct rtrs_srv_con *to_srv_con(struct rtrs_con *c)
59 {
60 	return container_of(c, struct rtrs_srv_con, c);
61 }
62 
63 static inline struct rtrs_srv_sess *to_srv_sess(struct rtrs_sess *s)
64 {
65 	return container_of(s, struct rtrs_srv_sess, s);
66 }
67 
68 static bool __rtrs_srv_change_state(struct rtrs_srv_sess *sess,
69 				     enum rtrs_srv_state new_state)
70 {
71 	enum rtrs_srv_state old_state;
72 	bool changed = false;
73 
74 	lockdep_assert_held(&sess->state_lock);
75 	old_state = sess->state;
76 	switch (new_state) {
77 	case RTRS_SRV_CONNECTED:
78 		switch (old_state) {
79 		case RTRS_SRV_CONNECTING:
80 			changed = true;
81 			fallthrough;
82 		default:
83 			break;
84 		}
85 		break;
86 	case RTRS_SRV_CLOSING:
87 		switch (old_state) {
88 		case RTRS_SRV_CONNECTING:
89 		case RTRS_SRV_CONNECTED:
90 			changed = true;
91 			fallthrough;
92 		default:
93 			break;
94 		}
95 		break;
96 	case RTRS_SRV_CLOSED:
97 		switch (old_state) {
98 		case RTRS_SRV_CLOSING:
99 			changed = true;
100 			fallthrough;
101 		default:
102 			break;
103 		}
104 		break;
105 	default:
106 		break;
107 	}
108 	if (changed)
109 		sess->state = new_state;
110 
111 	return changed;
112 }
113 
114 static bool rtrs_srv_change_state_get_old(struct rtrs_srv_sess *sess,
115 					   enum rtrs_srv_state new_state,
116 					   enum rtrs_srv_state *old_state)
117 {
118 	bool changed;
119 
120 	spin_lock_irq(&sess->state_lock);
121 	*old_state = sess->state;
122 	changed = __rtrs_srv_change_state(sess, new_state);
123 	spin_unlock_irq(&sess->state_lock);
124 
125 	return changed;
126 }
127 
128 static bool rtrs_srv_change_state(struct rtrs_srv_sess *sess,
129 				   enum rtrs_srv_state new_state)
130 {
131 	enum rtrs_srv_state old_state;
132 
133 	return rtrs_srv_change_state_get_old(sess, new_state, &old_state);
134 }
135 
136 static void free_id(struct rtrs_srv_op *id)
137 {
138 	if (!id)
139 		return;
140 	kfree(id);
141 }
142 
143 static void rtrs_srv_free_ops_ids(struct rtrs_srv_sess *sess)
144 {
145 	struct rtrs_srv *srv = sess->srv;
146 	int i;
147 
148 	WARN_ON(atomic_read(&sess->ids_inflight));
149 	if (sess->ops_ids) {
150 		for (i = 0; i < srv->queue_depth; i++)
151 			free_id(sess->ops_ids[i]);
152 		kfree(sess->ops_ids);
153 		sess->ops_ids = NULL;
154 	}
155 }
156 
157 static void rtrs_srv_rdma_done(struct ib_cq *cq, struct ib_wc *wc);
158 
159 static struct ib_cqe io_comp_cqe = {
160 	.done = rtrs_srv_rdma_done
161 };
162 
163 static int rtrs_srv_alloc_ops_ids(struct rtrs_srv_sess *sess)
164 {
165 	struct rtrs_srv *srv = sess->srv;
166 	struct rtrs_srv_op *id;
167 	int i;
168 
169 	sess->ops_ids = kcalloc(srv->queue_depth, sizeof(*sess->ops_ids),
170 				GFP_KERNEL);
171 	if (!sess->ops_ids)
172 		goto err;
173 
174 	for (i = 0; i < srv->queue_depth; ++i) {
175 		id = kzalloc(sizeof(*id), GFP_KERNEL);
176 		if (!id)
177 			goto err;
178 
179 		sess->ops_ids[i] = id;
180 	}
181 	init_waitqueue_head(&sess->ids_waitq);
182 	atomic_set(&sess->ids_inflight, 0);
183 
184 	return 0;
185 
186 err:
187 	rtrs_srv_free_ops_ids(sess);
188 	return -ENOMEM;
189 }
190 
191 static inline void rtrs_srv_get_ops_ids(struct rtrs_srv_sess *sess)
192 {
193 	atomic_inc(&sess->ids_inflight);
194 }
195 
196 static inline void rtrs_srv_put_ops_ids(struct rtrs_srv_sess *sess)
197 {
198 	if (atomic_dec_and_test(&sess->ids_inflight))
199 		wake_up(&sess->ids_waitq);
200 }
201 
202 static void rtrs_srv_wait_ops_ids(struct rtrs_srv_sess *sess)
203 {
204 	wait_event(sess->ids_waitq, !atomic_read(&sess->ids_inflight));
205 }
206 
207 
208 static void rtrs_srv_reg_mr_done(struct ib_cq *cq, struct ib_wc *wc)
209 {
210 	struct rtrs_srv_con *con = cq->cq_context;
211 	struct rtrs_sess *s = con->c.sess;
212 	struct rtrs_srv_sess *sess = to_srv_sess(s);
213 
214 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
215 		rtrs_err(s, "REG MR failed: %s\n",
216 			  ib_wc_status_msg(wc->status));
217 		close_sess(sess);
218 		return;
219 	}
220 }
221 
222 static struct ib_cqe local_reg_cqe = {
223 	.done = rtrs_srv_reg_mr_done
224 };
225 
226 static int rdma_write_sg(struct rtrs_srv_op *id)
227 {
228 	struct rtrs_sess *s = id->con->c.sess;
229 	struct rtrs_srv_sess *sess = to_srv_sess(s);
230 	dma_addr_t dma_addr = sess->dma_addr[id->msg_id];
231 	struct rtrs_srv_mr *srv_mr;
232 	struct rtrs_srv *srv = sess->srv;
233 	struct ib_send_wr inv_wr, imm_wr;
234 	struct ib_rdma_wr *wr = NULL;
235 	enum ib_send_flags flags;
236 	size_t sg_cnt;
237 	int err, offset;
238 	bool need_inval;
239 	u32 rkey = 0;
240 	struct ib_reg_wr rwr;
241 	struct ib_sge *plist;
242 	struct ib_sge list;
243 
244 	sg_cnt = le16_to_cpu(id->rd_msg->sg_cnt);
245 	need_inval = le16_to_cpu(id->rd_msg->flags) & RTRS_MSG_NEED_INVAL_F;
246 	if (unlikely(sg_cnt != 1))
247 		return -EINVAL;
248 
249 	offset = 0;
250 
251 	wr		= &id->tx_wr;
252 	plist		= &id->tx_sg;
253 	plist->addr	= dma_addr + offset;
254 	plist->length	= le32_to_cpu(id->rd_msg->desc[0].len);
255 
256 	/* WR will fail with length error
257 	 * if this is 0
258 	 */
259 	if (unlikely(plist->length == 0)) {
260 		rtrs_err(s, "Invalid RDMA-Write sg list length 0\n");
261 		return -EINVAL;
262 	}
263 
264 	plist->lkey = sess->s.dev->ib_pd->local_dma_lkey;
265 	offset += plist->length;
266 
267 	wr->wr.sg_list	= plist;
268 	wr->wr.num_sge	= 1;
269 	wr->remote_addr	= le64_to_cpu(id->rd_msg->desc[0].addr);
270 	wr->rkey	= le32_to_cpu(id->rd_msg->desc[0].key);
271 	if (rkey == 0)
272 		rkey = wr->rkey;
273 	else
274 		/* Only one key is actually used */
275 		WARN_ON_ONCE(rkey != wr->rkey);
276 
277 	wr->wr.opcode = IB_WR_RDMA_WRITE;
278 	wr->wr.ex.imm_data = 0;
279 	wr->wr.send_flags  = 0;
280 
281 	if (need_inval && always_invalidate) {
282 		wr->wr.next = &rwr.wr;
283 		rwr.wr.next = &inv_wr;
284 		inv_wr.next = &imm_wr;
285 	} else if (always_invalidate) {
286 		wr->wr.next = &rwr.wr;
287 		rwr.wr.next = &imm_wr;
288 	} else if (need_inval) {
289 		wr->wr.next = &inv_wr;
290 		inv_wr.next = &imm_wr;
291 	} else {
292 		wr->wr.next = &imm_wr;
293 	}
294 	/*
295 	 * From time to time we have to post signaled sends,
296 	 * or send queue will fill up and only QP reset can help.
297 	 */
298 	flags = (atomic_inc_return(&id->con->wr_cnt) % srv->queue_depth) ?
299 		0 : IB_SEND_SIGNALED;
300 
301 	if (need_inval) {
302 		inv_wr.sg_list = NULL;
303 		inv_wr.num_sge = 0;
304 		inv_wr.opcode = IB_WR_SEND_WITH_INV;
305 		inv_wr.send_flags = 0;
306 		inv_wr.ex.invalidate_rkey = rkey;
307 	}
308 
309 	imm_wr.next = NULL;
310 	if (always_invalidate) {
311 		struct rtrs_msg_rkey_rsp *msg;
312 
313 		srv_mr = &sess->mrs[id->msg_id];
314 		rwr.wr.opcode = IB_WR_REG_MR;
315 		rwr.wr.num_sge = 0;
316 		rwr.mr = srv_mr->mr;
317 		rwr.wr.send_flags = 0;
318 		rwr.key = srv_mr->mr->rkey;
319 		rwr.access = (IB_ACCESS_LOCAL_WRITE |
320 			      IB_ACCESS_REMOTE_WRITE);
321 		msg = srv_mr->iu->buf;
322 		msg->buf_id = cpu_to_le16(id->msg_id);
323 		msg->type = cpu_to_le16(RTRS_MSG_RKEY_RSP);
324 		msg->rkey = cpu_to_le32(srv_mr->mr->rkey);
325 
326 		list.addr   = srv_mr->iu->dma_addr;
327 		list.length = sizeof(*msg);
328 		list.lkey   = sess->s.dev->ib_pd->local_dma_lkey;
329 		imm_wr.sg_list = &list;
330 		imm_wr.num_sge = 1;
331 		imm_wr.opcode = IB_WR_SEND_WITH_IMM;
332 		ib_dma_sync_single_for_device(sess->s.dev->ib_dev,
333 					      srv_mr->iu->dma_addr,
334 					      srv_mr->iu->size, DMA_TO_DEVICE);
335 	} else {
336 		imm_wr.sg_list = NULL;
337 		imm_wr.num_sge = 0;
338 		imm_wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM;
339 	}
340 	imm_wr.send_flags = flags;
341 	imm_wr.ex.imm_data = cpu_to_be32(rtrs_to_io_rsp_imm(id->msg_id,
342 							     0, need_inval));
343 
344 	imm_wr.wr_cqe   = &io_comp_cqe;
345 	ib_dma_sync_single_for_device(sess->s.dev->ib_dev, dma_addr,
346 				      offset, DMA_BIDIRECTIONAL);
347 
348 	err = ib_post_send(id->con->c.qp, &id->tx_wr.wr, NULL);
349 	if (unlikely(err))
350 		rtrs_err(s,
351 			  "Posting RDMA-Write-Request to QP failed, err: %d\n",
352 			  err);
353 
354 	return err;
355 }
356 
357 /**
358  * send_io_resp_imm() - respond to client with empty IMM on failed READ/WRITE
359  *                      requests or on successful WRITE request.
360  * @con:	the connection to send back result
361  * @id:		the id associated with the IO
362  * @errno:	the error number of the IO.
363  *
364  * Return 0 on success, errno otherwise.
365  */
366 static int send_io_resp_imm(struct rtrs_srv_con *con, struct rtrs_srv_op *id,
367 			    int errno)
368 {
369 	struct rtrs_sess *s = con->c.sess;
370 	struct rtrs_srv_sess *sess = to_srv_sess(s);
371 	struct ib_send_wr inv_wr, imm_wr, *wr = NULL;
372 	struct ib_reg_wr rwr;
373 	struct rtrs_srv *srv = sess->srv;
374 	struct rtrs_srv_mr *srv_mr;
375 	bool need_inval = false;
376 	enum ib_send_flags flags;
377 	u32 imm;
378 	int err;
379 
380 	if (id->dir == READ) {
381 		struct rtrs_msg_rdma_read *rd_msg = id->rd_msg;
382 		size_t sg_cnt;
383 
384 		need_inval = le16_to_cpu(rd_msg->flags) &
385 				RTRS_MSG_NEED_INVAL_F;
386 		sg_cnt = le16_to_cpu(rd_msg->sg_cnt);
387 
388 		if (need_inval) {
389 			if (likely(sg_cnt)) {
390 				inv_wr.sg_list = NULL;
391 				inv_wr.num_sge = 0;
392 				inv_wr.opcode = IB_WR_SEND_WITH_INV;
393 				inv_wr.send_flags = 0;
394 				/* Only one key is actually used */
395 				inv_wr.ex.invalidate_rkey =
396 					le32_to_cpu(rd_msg->desc[0].key);
397 			} else {
398 				WARN_ON_ONCE(1);
399 				need_inval = false;
400 			}
401 		}
402 	}
403 
404 	if (need_inval && always_invalidate) {
405 		wr = &inv_wr;
406 		inv_wr.next = &rwr.wr;
407 		rwr.wr.next = &imm_wr;
408 	} else if (always_invalidate) {
409 		wr = &rwr.wr;
410 		rwr.wr.next = &imm_wr;
411 	} else if (need_inval) {
412 		wr = &inv_wr;
413 		inv_wr.next = &imm_wr;
414 	} else {
415 		wr = &imm_wr;
416 	}
417 	/*
418 	 * From time to time we have to post signalled sends,
419 	 * or send queue will fill up and only QP reset can help.
420 	 */
421 	flags = (atomic_inc_return(&con->wr_cnt) % srv->queue_depth) ?
422 		0 : IB_SEND_SIGNALED;
423 	imm = rtrs_to_io_rsp_imm(id->msg_id, errno, need_inval);
424 	imm_wr.next = NULL;
425 	if (always_invalidate) {
426 		struct ib_sge list;
427 		struct rtrs_msg_rkey_rsp *msg;
428 
429 		srv_mr = &sess->mrs[id->msg_id];
430 		rwr.wr.next = &imm_wr;
431 		rwr.wr.opcode = IB_WR_REG_MR;
432 		rwr.wr.num_sge = 0;
433 		rwr.wr.send_flags = 0;
434 		rwr.mr = srv_mr->mr;
435 		rwr.key = srv_mr->mr->rkey;
436 		rwr.access = (IB_ACCESS_LOCAL_WRITE |
437 			      IB_ACCESS_REMOTE_WRITE);
438 		msg = srv_mr->iu->buf;
439 		msg->buf_id = cpu_to_le16(id->msg_id);
440 		msg->type = cpu_to_le16(RTRS_MSG_RKEY_RSP);
441 		msg->rkey = cpu_to_le32(srv_mr->mr->rkey);
442 
443 		list.addr   = srv_mr->iu->dma_addr;
444 		list.length = sizeof(*msg);
445 		list.lkey   = sess->s.dev->ib_pd->local_dma_lkey;
446 		imm_wr.sg_list = &list;
447 		imm_wr.num_sge = 1;
448 		imm_wr.opcode = IB_WR_SEND_WITH_IMM;
449 		ib_dma_sync_single_for_device(sess->s.dev->ib_dev,
450 					      srv_mr->iu->dma_addr,
451 					      srv_mr->iu->size, DMA_TO_DEVICE);
452 	} else {
453 		imm_wr.sg_list = NULL;
454 		imm_wr.num_sge = 0;
455 		imm_wr.opcode = IB_WR_RDMA_WRITE_WITH_IMM;
456 	}
457 	imm_wr.send_flags = flags;
458 	imm_wr.wr_cqe   = &io_comp_cqe;
459 
460 	imm_wr.ex.imm_data = cpu_to_be32(imm);
461 
462 	err = ib_post_send(id->con->c.qp, wr, NULL);
463 	if (unlikely(err))
464 		rtrs_err_rl(s, "Posting RDMA-Reply to QP failed, err: %d\n",
465 			     err);
466 
467 	return err;
468 }
469 
470 void close_sess(struct rtrs_srv_sess *sess)
471 {
472 	enum rtrs_srv_state old_state;
473 
474 	if (rtrs_srv_change_state_get_old(sess, RTRS_SRV_CLOSING,
475 					   &old_state))
476 		queue_work(rtrs_wq, &sess->close_work);
477 	WARN_ON(sess->state != RTRS_SRV_CLOSING);
478 }
479 
480 static inline const char *rtrs_srv_state_str(enum rtrs_srv_state state)
481 {
482 	switch (state) {
483 	case RTRS_SRV_CONNECTING:
484 		return "RTRS_SRV_CONNECTING";
485 	case RTRS_SRV_CONNECTED:
486 		return "RTRS_SRV_CONNECTED";
487 	case RTRS_SRV_CLOSING:
488 		return "RTRS_SRV_CLOSING";
489 	case RTRS_SRV_CLOSED:
490 		return "RTRS_SRV_CLOSED";
491 	default:
492 		return "UNKNOWN";
493 	}
494 }
495 
496 /**
497  * rtrs_srv_resp_rdma() - Finish an RDMA request
498  *
499  * @id:		Internal RTRS operation identifier
500  * @status:	Response Code sent to the other side for this operation.
501  *		0 = success, <=0 error
502  * Context: any
503  *
504  * Finish a RDMA operation. A message is sent to the client and the
505  * corresponding memory areas will be released.
506  */
507 bool rtrs_srv_resp_rdma(struct rtrs_srv_op *id, int status)
508 {
509 	struct rtrs_srv_sess *sess;
510 	struct rtrs_srv_con *con;
511 	struct rtrs_sess *s;
512 	int err;
513 
514 	if (WARN_ON(!id))
515 		return true;
516 
517 	con = id->con;
518 	s = con->c.sess;
519 	sess = to_srv_sess(s);
520 
521 	id->status = status;
522 
523 	if (unlikely(sess->state != RTRS_SRV_CONNECTED)) {
524 		rtrs_err_rl(s,
525 			     "Sending I/O response failed,  session is disconnected, sess state %s\n",
526 			     rtrs_srv_state_str(sess->state));
527 		goto out;
528 	}
529 	if (always_invalidate) {
530 		struct rtrs_srv_mr *mr = &sess->mrs[id->msg_id];
531 
532 		ib_update_fast_reg_key(mr->mr, ib_inc_rkey(mr->mr->rkey));
533 	}
534 	if (unlikely(atomic_sub_return(1,
535 				       &con->sq_wr_avail) < 0)) {
536 		pr_err("IB send queue full\n");
537 		atomic_add(1, &con->sq_wr_avail);
538 		spin_lock(&con->rsp_wr_wait_lock);
539 		list_add_tail(&id->wait_list, &con->rsp_wr_wait_list);
540 		spin_unlock(&con->rsp_wr_wait_lock);
541 		return false;
542 	}
543 
544 	if (status || id->dir == WRITE || !id->rd_msg->sg_cnt)
545 		err = send_io_resp_imm(con, id, status);
546 	else
547 		err = rdma_write_sg(id);
548 
549 	if (unlikely(err)) {
550 		rtrs_err_rl(s, "IO response failed: %d\n", err);
551 		close_sess(sess);
552 	}
553 out:
554 	rtrs_srv_put_ops_ids(sess);
555 	return true;
556 }
557 EXPORT_SYMBOL(rtrs_srv_resp_rdma);
558 
559 /**
560  * rtrs_srv_set_sess_priv() - Set private pointer in rtrs_srv.
561  * @srv:	Session pointer
562  * @priv:	The private pointer that is associated with the session.
563  */
564 void rtrs_srv_set_sess_priv(struct rtrs_srv *srv, void *priv)
565 {
566 	srv->priv = priv;
567 }
568 EXPORT_SYMBOL(rtrs_srv_set_sess_priv);
569 
570 static void unmap_cont_bufs(struct rtrs_srv_sess *sess)
571 {
572 	int i;
573 
574 	for (i = 0; i < sess->mrs_num; i++) {
575 		struct rtrs_srv_mr *srv_mr;
576 
577 		srv_mr = &sess->mrs[i];
578 		rtrs_iu_free(srv_mr->iu, DMA_TO_DEVICE,
579 			      sess->s.dev->ib_dev, 1);
580 		ib_dereg_mr(srv_mr->mr);
581 		ib_dma_unmap_sg(sess->s.dev->ib_dev, srv_mr->sgt.sgl,
582 				srv_mr->sgt.nents, DMA_BIDIRECTIONAL);
583 		sg_free_table(&srv_mr->sgt);
584 	}
585 	kfree(sess->mrs);
586 }
587 
588 static int map_cont_bufs(struct rtrs_srv_sess *sess)
589 {
590 	struct rtrs_srv *srv = sess->srv;
591 	struct rtrs_sess *ss = &sess->s;
592 	int i, mri, err, mrs_num;
593 	unsigned int chunk_bits;
594 	int chunks_per_mr = 1;
595 
596 	/*
597 	 * Here we map queue_depth chunks to MR.  Firstly we have to
598 	 * figure out how many chunks can we map per MR.
599 	 */
600 	if (always_invalidate) {
601 		/*
602 		 * in order to do invalidate for each chunks of memory, we needs
603 		 * more memory regions.
604 		 */
605 		mrs_num = srv->queue_depth;
606 	} else {
607 		chunks_per_mr =
608 			sess->s.dev->ib_dev->attrs.max_fast_reg_page_list_len;
609 		mrs_num = DIV_ROUND_UP(srv->queue_depth, chunks_per_mr);
610 		chunks_per_mr = DIV_ROUND_UP(srv->queue_depth, mrs_num);
611 	}
612 
613 	sess->mrs = kcalloc(mrs_num, sizeof(*sess->mrs), GFP_KERNEL);
614 	if (!sess->mrs)
615 		return -ENOMEM;
616 
617 	sess->mrs_num = mrs_num;
618 
619 	for (mri = 0; mri < mrs_num; mri++) {
620 		struct rtrs_srv_mr *srv_mr = &sess->mrs[mri];
621 		struct sg_table *sgt = &srv_mr->sgt;
622 		struct scatterlist *s;
623 		struct ib_mr *mr;
624 		int nr, chunks;
625 
626 		chunks = chunks_per_mr * mri;
627 		if (!always_invalidate)
628 			chunks_per_mr = min_t(int, chunks_per_mr,
629 					      srv->queue_depth - chunks);
630 
631 		err = sg_alloc_table(sgt, chunks_per_mr, GFP_KERNEL);
632 		if (err)
633 			goto err;
634 
635 		for_each_sg(sgt->sgl, s, chunks_per_mr, i)
636 			sg_set_page(s, srv->chunks[chunks + i],
637 				    max_chunk_size, 0);
638 
639 		nr = ib_dma_map_sg(sess->s.dev->ib_dev, sgt->sgl,
640 				   sgt->nents, DMA_BIDIRECTIONAL);
641 		if (nr < sgt->nents) {
642 			err = nr < 0 ? nr : -EINVAL;
643 			goto free_sg;
644 		}
645 		mr = ib_alloc_mr(sess->s.dev->ib_pd, IB_MR_TYPE_MEM_REG,
646 				 sgt->nents);
647 		if (IS_ERR(mr)) {
648 			err = PTR_ERR(mr);
649 			goto unmap_sg;
650 		}
651 		nr = ib_map_mr_sg(mr, sgt->sgl, sgt->nents,
652 				  NULL, max_chunk_size);
653 		if (nr < 0 || nr < sgt->nents) {
654 			err = nr < 0 ? nr : -EINVAL;
655 			goto dereg_mr;
656 		}
657 
658 		if (always_invalidate) {
659 			srv_mr->iu = rtrs_iu_alloc(1,
660 					sizeof(struct rtrs_msg_rkey_rsp),
661 					GFP_KERNEL, sess->s.dev->ib_dev,
662 					DMA_TO_DEVICE, rtrs_srv_rdma_done);
663 			if (!srv_mr->iu) {
664 				err = -ENOMEM;
665 				rtrs_err(ss, "rtrs_iu_alloc(), err: %d\n", err);
666 				goto free_iu;
667 			}
668 		}
669 		/* Eventually dma addr for each chunk can be cached */
670 		for_each_sg(sgt->sgl, s, sgt->orig_nents, i)
671 			sess->dma_addr[chunks + i] = sg_dma_address(s);
672 
673 		ib_update_fast_reg_key(mr, ib_inc_rkey(mr->rkey));
674 		srv_mr->mr = mr;
675 
676 		continue;
677 err:
678 		while (mri--) {
679 			srv_mr = &sess->mrs[mri];
680 			sgt = &srv_mr->sgt;
681 			mr = srv_mr->mr;
682 free_iu:
683 			rtrs_iu_free(srv_mr->iu, DMA_TO_DEVICE,
684 				      sess->s.dev->ib_dev, 1);
685 dereg_mr:
686 			ib_dereg_mr(mr);
687 unmap_sg:
688 			ib_dma_unmap_sg(sess->s.dev->ib_dev, sgt->sgl,
689 					sgt->nents, DMA_BIDIRECTIONAL);
690 free_sg:
691 			sg_free_table(sgt);
692 		}
693 		kfree(sess->mrs);
694 
695 		return err;
696 	}
697 
698 	chunk_bits = ilog2(srv->queue_depth - 1) + 1;
699 	sess->mem_bits = (MAX_IMM_PAYL_BITS - chunk_bits);
700 
701 	return 0;
702 }
703 
704 static void rtrs_srv_hb_err_handler(struct rtrs_con *c)
705 {
706 	close_sess(to_srv_sess(c->sess));
707 }
708 
709 static void rtrs_srv_init_hb(struct rtrs_srv_sess *sess)
710 {
711 	rtrs_init_hb(&sess->s, &io_comp_cqe,
712 		      RTRS_HB_INTERVAL_MS,
713 		      RTRS_HB_MISSED_MAX,
714 		      rtrs_srv_hb_err_handler,
715 		      rtrs_wq);
716 }
717 
718 static void rtrs_srv_start_hb(struct rtrs_srv_sess *sess)
719 {
720 	rtrs_start_hb(&sess->s);
721 }
722 
723 static void rtrs_srv_stop_hb(struct rtrs_srv_sess *sess)
724 {
725 	rtrs_stop_hb(&sess->s);
726 }
727 
728 static void rtrs_srv_info_rsp_done(struct ib_cq *cq, struct ib_wc *wc)
729 {
730 	struct rtrs_srv_con *con = cq->cq_context;
731 	struct rtrs_sess *s = con->c.sess;
732 	struct rtrs_srv_sess *sess = to_srv_sess(s);
733 	struct rtrs_iu *iu;
734 
735 	iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
736 	rtrs_iu_free(iu, DMA_TO_DEVICE, sess->s.dev->ib_dev, 1);
737 
738 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
739 		rtrs_err(s, "Sess info response send failed: %s\n",
740 			  ib_wc_status_msg(wc->status));
741 		close_sess(sess);
742 		return;
743 	}
744 	WARN_ON(wc->opcode != IB_WC_SEND);
745 }
746 
747 static void rtrs_srv_sess_up(struct rtrs_srv_sess *sess)
748 {
749 	struct rtrs_srv *srv = sess->srv;
750 	struct rtrs_srv_ctx *ctx = srv->ctx;
751 	int up;
752 
753 	mutex_lock(&srv->paths_ev_mutex);
754 	up = ++srv->paths_up;
755 	if (up == 1)
756 		ctx->ops.link_ev(srv, RTRS_SRV_LINK_EV_CONNECTED, NULL);
757 	mutex_unlock(&srv->paths_ev_mutex);
758 
759 	/* Mark session as established */
760 	sess->established = true;
761 }
762 
763 static void rtrs_srv_sess_down(struct rtrs_srv_sess *sess)
764 {
765 	struct rtrs_srv *srv = sess->srv;
766 	struct rtrs_srv_ctx *ctx = srv->ctx;
767 
768 	if (!sess->established)
769 		return;
770 
771 	sess->established = false;
772 	mutex_lock(&srv->paths_ev_mutex);
773 	WARN_ON(!srv->paths_up);
774 	if (--srv->paths_up == 0)
775 		ctx->ops.link_ev(srv, RTRS_SRV_LINK_EV_DISCONNECTED, srv->priv);
776 	mutex_unlock(&srv->paths_ev_mutex);
777 }
778 
779 static int post_recv_sess(struct rtrs_srv_sess *sess);
780 
781 static int process_info_req(struct rtrs_srv_con *con,
782 			    struct rtrs_msg_info_req *msg)
783 {
784 	struct rtrs_sess *s = con->c.sess;
785 	struct rtrs_srv_sess *sess = to_srv_sess(s);
786 	struct ib_send_wr *reg_wr = NULL;
787 	struct rtrs_msg_info_rsp *rsp;
788 	struct rtrs_iu *tx_iu;
789 	struct ib_reg_wr *rwr;
790 	int mri, err;
791 	size_t tx_sz;
792 
793 	err = post_recv_sess(sess);
794 	if (unlikely(err)) {
795 		rtrs_err(s, "post_recv_sess(), err: %d\n", err);
796 		return err;
797 	}
798 	rwr = kcalloc(sess->mrs_num, sizeof(*rwr), GFP_KERNEL);
799 	if (unlikely(!rwr))
800 		return -ENOMEM;
801 	strlcpy(sess->s.sessname, msg->sessname, sizeof(sess->s.sessname));
802 
803 	tx_sz  = sizeof(*rsp);
804 	tx_sz += sizeof(rsp->desc[0]) * sess->mrs_num;
805 	tx_iu = rtrs_iu_alloc(1, tx_sz, GFP_KERNEL, sess->s.dev->ib_dev,
806 			       DMA_TO_DEVICE, rtrs_srv_info_rsp_done);
807 	if (unlikely(!tx_iu)) {
808 		err = -ENOMEM;
809 		goto rwr_free;
810 	}
811 
812 	rsp = tx_iu->buf;
813 	rsp->type = cpu_to_le16(RTRS_MSG_INFO_RSP);
814 	rsp->sg_cnt = cpu_to_le16(sess->mrs_num);
815 
816 	for (mri = 0; mri < sess->mrs_num; mri++) {
817 		struct ib_mr *mr = sess->mrs[mri].mr;
818 
819 		rsp->desc[mri].addr = cpu_to_le64(mr->iova);
820 		rsp->desc[mri].key  = cpu_to_le32(mr->rkey);
821 		rsp->desc[mri].len  = cpu_to_le32(mr->length);
822 
823 		/*
824 		 * Fill in reg MR request and chain them *backwards*
825 		 */
826 		rwr[mri].wr.next = mri ? &rwr[mri - 1].wr : NULL;
827 		rwr[mri].wr.opcode = IB_WR_REG_MR;
828 		rwr[mri].wr.wr_cqe = &local_reg_cqe;
829 		rwr[mri].wr.num_sge = 0;
830 		rwr[mri].wr.send_flags = mri ? 0 : IB_SEND_SIGNALED;
831 		rwr[mri].mr = mr;
832 		rwr[mri].key = mr->rkey;
833 		rwr[mri].access = (IB_ACCESS_LOCAL_WRITE |
834 				   IB_ACCESS_REMOTE_WRITE);
835 		reg_wr = &rwr[mri].wr;
836 	}
837 
838 	err = rtrs_srv_create_sess_files(sess);
839 	if (unlikely(err))
840 		goto iu_free;
841 	kobject_get(&sess->kobj);
842 	get_device(&sess->srv->dev);
843 	rtrs_srv_change_state(sess, RTRS_SRV_CONNECTED);
844 	rtrs_srv_start_hb(sess);
845 
846 	/*
847 	 * We do not account number of established connections at the current
848 	 * moment, we rely on the client, which should send info request when
849 	 * all connections are successfully established.  Thus, simply notify
850 	 * listener with a proper event if we are the first path.
851 	 */
852 	rtrs_srv_sess_up(sess);
853 
854 	ib_dma_sync_single_for_device(sess->s.dev->ib_dev, tx_iu->dma_addr,
855 				      tx_iu->size, DMA_TO_DEVICE);
856 
857 	/* Send info response */
858 	err = rtrs_iu_post_send(&con->c, tx_iu, tx_sz, reg_wr);
859 	if (unlikely(err)) {
860 		rtrs_err(s, "rtrs_iu_post_send(), err: %d\n", err);
861 iu_free:
862 		rtrs_iu_free(tx_iu, DMA_TO_DEVICE, sess->s.dev->ib_dev, 1);
863 	}
864 rwr_free:
865 	kfree(rwr);
866 
867 	return err;
868 }
869 
870 static void rtrs_srv_info_req_done(struct ib_cq *cq, struct ib_wc *wc)
871 {
872 	struct rtrs_srv_con *con = cq->cq_context;
873 	struct rtrs_sess *s = con->c.sess;
874 	struct rtrs_srv_sess *sess = to_srv_sess(s);
875 	struct rtrs_msg_info_req *msg;
876 	struct rtrs_iu *iu;
877 	int err;
878 
879 	WARN_ON(con->c.cid);
880 
881 	iu = container_of(wc->wr_cqe, struct rtrs_iu, cqe);
882 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
883 		rtrs_err(s, "Sess info request receive failed: %s\n",
884 			  ib_wc_status_msg(wc->status));
885 		goto close;
886 	}
887 	WARN_ON(wc->opcode != IB_WC_RECV);
888 
889 	if (unlikely(wc->byte_len < sizeof(*msg))) {
890 		rtrs_err(s, "Sess info request is malformed: size %d\n",
891 			  wc->byte_len);
892 		goto close;
893 	}
894 	ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, iu->dma_addr,
895 				   iu->size, DMA_FROM_DEVICE);
896 	msg = iu->buf;
897 	if (unlikely(le16_to_cpu(msg->type) != RTRS_MSG_INFO_REQ)) {
898 		rtrs_err(s, "Sess info request is malformed: type %d\n",
899 			  le16_to_cpu(msg->type));
900 		goto close;
901 	}
902 	err = process_info_req(con, msg);
903 	if (unlikely(err))
904 		goto close;
905 
906 out:
907 	rtrs_iu_free(iu, DMA_FROM_DEVICE, sess->s.dev->ib_dev, 1);
908 	return;
909 close:
910 	close_sess(sess);
911 	goto out;
912 }
913 
914 static int post_recv_info_req(struct rtrs_srv_con *con)
915 {
916 	struct rtrs_sess *s = con->c.sess;
917 	struct rtrs_srv_sess *sess = to_srv_sess(s);
918 	struct rtrs_iu *rx_iu;
919 	int err;
920 
921 	rx_iu = rtrs_iu_alloc(1, sizeof(struct rtrs_msg_info_req),
922 			       GFP_KERNEL, sess->s.dev->ib_dev,
923 			       DMA_FROM_DEVICE, rtrs_srv_info_req_done);
924 	if (unlikely(!rx_iu))
925 		return -ENOMEM;
926 	/* Prepare for getting info response */
927 	err = rtrs_iu_post_recv(&con->c, rx_iu);
928 	if (unlikely(err)) {
929 		rtrs_err(s, "rtrs_iu_post_recv(), err: %d\n", err);
930 		rtrs_iu_free(rx_iu, DMA_FROM_DEVICE, sess->s.dev->ib_dev, 1);
931 		return err;
932 	}
933 
934 	return 0;
935 }
936 
937 static int post_recv_io(struct rtrs_srv_con *con, size_t q_size)
938 {
939 	int i, err;
940 
941 	for (i = 0; i < q_size; i++) {
942 		err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
943 		if (unlikely(err))
944 			return err;
945 	}
946 
947 	return 0;
948 }
949 
950 static int post_recv_sess(struct rtrs_srv_sess *sess)
951 {
952 	struct rtrs_srv *srv = sess->srv;
953 	struct rtrs_sess *s = &sess->s;
954 	size_t q_size;
955 	int err, cid;
956 
957 	for (cid = 0; cid < sess->s.con_num; cid++) {
958 		if (cid == 0)
959 			q_size = SERVICE_CON_QUEUE_DEPTH;
960 		else
961 			q_size = srv->queue_depth;
962 
963 		err = post_recv_io(to_srv_con(sess->s.con[cid]), q_size);
964 		if (unlikely(err)) {
965 			rtrs_err(s, "post_recv_io(), err: %d\n", err);
966 			return err;
967 		}
968 	}
969 
970 	return 0;
971 }
972 
973 static void process_read(struct rtrs_srv_con *con,
974 			 struct rtrs_msg_rdma_read *msg,
975 			 u32 buf_id, u32 off)
976 {
977 	struct rtrs_sess *s = con->c.sess;
978 	struct rtrs_srv_sess *sess = to_srv_sess(s);
979 	struct rtrs_srv *srv = sess->srv;
980 	struct rtrs_srv_ctx *ctx = srv->ctx;
981 	struct rtrs_srv_op *id;
982 
983 	size_t usr_len, data_len;
984 	void *data;
985 	int ret;
986 
987 	if (unlikely(sess->state != RTRS_SRV_CONNECTED)) {
988 		rtrs_err_rl(s,
989 			     "Processing read request failed,  session is disconnected, sess state %s\n",
990 			     rtrs_srv_state_str(sess->state));
991 		return;
992 	}
993 	if (unlikely(msg->sg_cnt != 1 && msg->sg_cnt != 0)) {
994 		rtrs_err_rl(s,
995 			    "Processing read request failed, invalid message\n");
996 		return;
997 	}
998 	rtrs_srv_get_ops_ids(sess);
999 	rtrs_srv_update_rdma_stats(sess->stats, off, READ);
1000 	id = sess->ops_ids[buf_id];
1001 	id->con		= con;
1002 	id->dir		= READ;
1003 	id->msg_id	= buf_id;
1004 	id->rd_msg	= msg;
1005 	usr_len = le16_to_cpu(msg->usr_len);
1006 	data_len = off - usr_len;
1007 	data = page_address(srv->chunks[buf_id]);
1008 	ret = ctx->ops.rdma_ev(srv, srv->priv, id, READ, data, data_len,
1009 			   data + data_len, usr_len);
1010 
1011 	if (unlikely(ret)) {
1012 		rtrs_err_rl(s,
1013 			     "Processing read request failed, user module cb reported for msg_id %d, err: %d\n",
1014 			     buf_id, ret);
1015 		goto send_err_msg;
1016 	}
1017 
1018 	return;
1019 
1020 send_err_msg:
1021 	ret = send_io_resp_imm(con, id, ret);
1022 	if (ret < 0) {
1023 		rtrs_err_rl(s,
1024 			     "Sending err msg for failed RDMA-Write-Req failed, msg_id %d, err: %d\n",
1025 			     buf_id, ret);
1026 		close_sess(sess);
1027 	}
1028 	rtrs_srv_put_ops_ids(sess);
1029 }
1030 
1031 static void process_write(struct rtrs_srv_con *con,
1032 			  struct rtrs_msg_rdma_write *req,
1033 			  u32 buf_id, u32 off)
1034 {
1035 	struct rtrs_sess *s = con->c.sess;
1036 	struct rtrs_srv_sess *sess = to_srv_sess(s);
1037 	struct rtrs_srv *srv = sess->srv;
1038 	struct rtrs_srv_ctx *ctx = srv->ctx;
1039 	struct rtrs_srv_op *id;
1040 
1041 	size_t data_len, usr_len;
1042 	void *data;
1043 	int ret;
1044 
1045 	if (unlikely(sess->state != RTRS_SRV_CONNECTED)) {
1046 		rtrs_err_rl(s,
1047 			     "Processing write request failed,  session is disconnected, sess state %s\n",
1048 			     rtrs_srv_state_str(sess->state));
1049 		return;
1050 	}
1051 	rtrs_srv_get_ops_ids(sess);
1052 	rtrs_srv_update_rdma_stats(sess->stats, off, WRITE);
1053 	id = sess->ops_ids[buf_id];
1054 	id->con    = con;
1055 	id->dir    = WRITE;
1056 	id->msg_id = buf_id;
1057 
1058 	usr_len = le16_to_cpu(req->usr_len);
1059 	data_len = off - usr_len;
1060 	data = page_address(srv->chunks[buf_id]);
1061 	ret = ctx->ops.rdma_ev(srv, srv->priv, id, WRITE, data, data_len,
1062 			   data + data_len, usr_len);
1063 	if (unlikely(ret)) {
1064 		rtrs_err_rl(s,
1065 			     "Processing write request failed, user module callback reports err: %d\n",
1066 			     ret);
1067 		goto send_err_msg;
1068 	}
1069 
1070 	return;
1071 
1072 send_err_msg:
1073 	ret = send_io_resp_imm(con, id, ret);
1074 	if (ret < 0) {
1075 		rtrs_err_rl(s,
1076 			     "Processing write request failed, sending I/O response failed, msg_id %d, err: %d\n",
1077 			     buf_id, ret);
1078 		close_sess(sess);
1079 	}
1080 	rtrs_srv_put_ops_ids(sess);
1081 }
1082 
1083 static void process_io_req(struct rtrs_srv_con *con, void *msg,
1084 			   u32 id, u32 off)
1085 {
1086 	struct rtrs_sess *s = con->c.sess;
1087 	struct rtrs_srv_sess *sess = to_srv_sess(s);
1088 	struct rtrs_msg_rdma_hdr *hdr;
1089 	unsigned int type;
1090 
1091 	ib_dma_sync_single_for_cpu(sess->s.dev->ib_dev, sess->dma_addr[id],
1092 				   max_chunk_size, DMA_BIDIRECTIONAL);
1093 	hdr = msg;
1094 	type = le16_to_cpu(hdr->type);
1095 
1096 	switch (type) {
1097 	case RTRS_MSG_WRITE:
1098 		process_write(con, msg, id, off);
1099 		break;
1100 	case RTRS_MSG_READ:
1101 		process_read(con, msg, id, off);
1102 		break;
1103 	default:
1104 		rtrs_err(s,
1105 			  "Processing I/O request failed, unknown message type received: 0x%02x\n",
1106 			  type);
1107 		goto err;
1108 	}
1109 
1110 	return;
1111 
1112 err:
1113 	close_sess(sess);
1114 }
1115 
1116 static void rtrs_srv_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1117 {
1118 	struct rtrs_srv_mr *mr =
1119 		container_of(wc->wr_cqe, typeof(*mr), inv_cqe);
1120 	struct rtrs_srv_con *con = cq->cq_context;
1121 	struct rtrs_sess *s = con->c.sess;
1122 	struct rtrs_srv_sess *sess = to_srv_sess(s);
1123 	struct rtrs_srv *srv = sess->srv;
1124 	u32 msg_id, off;
1125 	void *data;
1126 
1127 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1128 		rtrs_err(s, "Failed IB_WR_LOCAL_INV: %s\n",
1129 			  ib_wc_status_msg(wc->status));
1130 		close_sess(sess);
1131 	}
1132 	msg_id = mr->msg_id;
1133 	off = mr->msg_off;
1134 	data = page_address(srv->chunks[msg_id]) + off;
1135 	process_io_req(con, data, msg_id, off);
1136 }
1137 
1138 static int rtrs_srv_inv_rkey(struct rtrs_srv_con *con,
1139 			      struct rtrs_srv_mr *mr)
1140 {
1141 	struct ib_send_wr wr = {
1142 		.opcode		    = IB_WR_LOCAL_INV,
1143 		.wr_cqe		    = &mr->inv_cqe,
1144 		.send_flags	    = IB_SEND_SIGNALED,
1145 		.ex.invalidate_rkey = mr->mr->rkey,
1146 	};
1147 	mr->inv_cqe.done = rtrs_srv_inv_rkey_done;
1148 
1149 	return ib_post_send(con->c.qp, &wr, NULL);
1150 }
1151 
1152 static void rtrs_rdma_process_wr_wait_list(struct rtrs_srv_con *con)
1153 {
1154 	spin_lock(&con->rsp_wr_wait_lock);
1155 	while (!list_empty(&con->rsp_wr_wait_list)) {
1156 		struct rtrs_srv_op *id;
1157 		int ret;
1158 
1159 		id = list_entry(con->rsp_wr_wait_list.next,
1160 				struct rtrs_srv_op, wait_list);
1161 		list_del(&id->wait_list);
1162 
1163 		spin_unlock(&con->rsp_wr_wait_lock);
1164 		ret = rtrs_srv_resp_rdma(id, id->status);
1165 		spin_lock(&con->rsp_wr_wait_lock);
1166 
1167 		if (!ret) {
1168 			list_add(&id->wait_list, &con->rsp_wr_wait_list);
1169 			break;
1170 		}
1171 	}
1172 	spin_unlock(&con->rsp_wr_wait_lock);
1173 }
1174 
1175 static void rtrs_srv_rdma_done(struct ib_cq *cq, struct ib_wc *wc)
1176 {
1177 	struct rtrs_srv_con *con = cq->cq_context;
1178 	struct rtrs_sess *s = con->c.sess;
1179 	struct rtrs_srv_sess *sess = to_srv_sess(s);
1180 	struct rtrs_srv *srv = sess->srv;
1181 	u32 imm_type, imm_payload;
1182 	int err;
1183 
1184 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1185 		if (wc->status != IB_WC_WR_FLUSH_ERR) {
1186 			rtrs_err(s,
1187 				  "%s (wr_cqe: %p, type: %d, vendor_err: 0x%x, len: %u)\n",
1188 				  ib_wc_status_msg(wc->status), wc->wr_cqe,
1189 				  wc->opcode, wc->vendor_err, wc->byte_len);
1190 			close_sess(sess);
1191 		}
1192 		return;
1193 	}
1194 
1195 	switch (wc->opcode) {
1196 	case IB_WC_RECV_RDMA_WITH_IMM:
1197 		/*
1198 		 * post_recv() RDMA write completions of IO reqs (read/write)
1199 		 * and hb
1200 		 */
1201 		if (WARN_ON(wc->wr_cqe != &io_comp_cqe))
1202 			return;
1203 		err = rtrs_post_recv_empty(&con->c, &io_comp_cqe);
1204 		if (unlikely(err)) {
1205 			rtrs_err(s, "rtrs_post_recv(), err: %d\n", err);
1206 			close_sess(sess);
1207 			break;
1208 		}
1209 		rtrs_from_imm(be32_to_cpu(wc->ex.imm_data),
1210 			       &imm_type, &imm_payload);
1211 		if (likely(imm_type == RTRS_IO_REQ_IMM)) {
1212 			u32 msg_id, off;
1213 			void *data;
1214 
1215 			msg_id = imm_payload >> sess->mem_bits;
1216 			off = imm_payload & ((1 << sess->mem_bits) - 1);
1217 			if (unlikely(msg_id >= srv->queue_depth ||
1218 				     off >= max_chunk_size)) {
1219 				rtrs_err(s, "Wrong msg_id %u, off %u\n",
1220 					  msg_id, off);
1221 				close_sess(sess);
1222 				return;
1223 			}
1224 			if (always_invalidate) {
1225 				struct rtrs_srv_mr *mr = &sess->mrs[msg_id];
1226 
1227 				mr->msg_off = off;
1228 				mr->msg_id = msg_id;
1229 				err = rtrs_srv_inv_rkey(con, mr);
1230 				if (unlikely(err)) {
1231 					rtrs_err(s, "rtrs_post_recv(), err: %d\n",
1232 						  err);
1233 					close_sess(sess);
1234 					break;
1235 				}
1236 			} else {
1237 				data = page_address(srv->chunks[msg_id]) + off;
1238 				process_io_req(con, data, msg_id, off);
1239 			}
1240 		} else if (imm_type == RTRS_HB_MSG_IMM) {
1241 			WARN_ON(con->c.cid);
1242 			rtrs_send_hb_ack(&sess->s);
1243 		} else if (imm_type == RTRS_HB_ACK_IMM) {
1244 			WARN_ON(con->c.cid);
1245 			sess->s.hb_missed_cnt = 0;
1246 		} else {
1247 			rtrs_wrn(s, "Unknown IMM type %u\n", imm_type);
1248 		}
1249 		break;
1250 	case IB_WC_RDMA_WRITE:
1251 	case IB_WC_SEND:
1252 		/*
1253 		 * post_send() RDMA write completions of IO reqs (read/write)
1254 		 * and hb
1255 		 */
1256 		atomic_add(srv->queue_depth, &con->sq_wr_avail);
1257 
1258 		if (unlikely(!list_empty_careful(&con->rsp_wr_wait_list)))
1259 			rtrs_rdma_process_wr_wait_list(con);
1260 
1261 		break;
1262 	default:
1263 		rtrs_wrn(s, "Unexpected WC type: %d\n", wc->opcode);
1264 		return;
1265 	}
1266 }
1267 
1268 /**
1269  * rtrs_srv_get_sess_name() - Get rtrs_srv peer hostname.
1270  * @srv:	Session
1271  * @sessname:	Sessname buffer
1272  * @len:	Length of sessname buffer
1273  */
1274 int rtrs_srv_get_sess_name(struct rtrs_srv *srv, char *sessname, size_t len)
1275 {
1276 	struct rtrs_srv_sess *sess;
1277 	int err = -ENOTCONN;
1278 
1279 	mutex_lock(&srv->paths_mutex);
1280 	list_for_each_entry(sess, &srv->paths_list, s.entry) {
1281 		if (sess->state != RTRS_SRV_CONNECTED)
1282 			continue;
1283 		strlcpy(sessname, sess->s.sessname,
1284 		       min_t(size_t, sizeof(sess->s.sessname), len));
1285 		err = 0;
1286 		break;
1287 	}
1288 	mutex_unlock(&srv->paths_mutex);
1289 
1290 	return err;
1291 }
1292 EXPORT_SYMBOL(rtrs_srv_get_sess_name);
1293 
1294 /**
1295  * rtrs_srv_get_sess_qdepth() - Get rtrs_srv qdepth.
1296  * @srv:	Session
1297  */
1298 int rtrs_srv_get_queue_depth(struct rtrs_srv *srv)
1299 {
1300 	return srv->queue_depth;
1301 }
1302 EXPORT_SYMBOL(rtrs_srv_get_queue_depth);
1303 
1304 static int find_next_bit_ring(struct rtrs_srv_sess *sess)
1305 {
1306 	struct ib_device *ib_dev = sess->s.dev->ib_dev;
1307 	int v;
1308 
1309 	v = cpumask_next(sess->cur_cq_vector, &cq_affinity_mask);
1310 	if (v >= nr_cpu_ids || v >= ib_dev->num_comp_vectors)
1311 		v = cpumask_first(&cq_affinity_mask);
1312 	return v;
1313 }
1314 
1315 static int rtrs_srv_get_next_cq_vector(struct rtrs_srv_sess *sess)
1316 {
1317 	sess->cur_cq_vector = find_next_bit_ring(sess);
1318 
1319 	return sess->cur_cq_vector;
1320 }
1321 
1322 static struct rtrs_srv *__alloc_srv(struct rtrs_srv_ctx *ctx,
1323 				     const uuid_t *paths_uuid)
1324 {
1325 	struct rtrs_srv *srv;
1326 	int i;
1327 
1328 	srv = kzalloc(sizeof(*srv), GFP_KERNEL);
1329 	if  (!srv)
1330 		return NULL;
1331 
1332 	refcount_set(&srv->refcount, 1);
1333 	INIT_LIST_HEAD(&srv->paths_list);
1334 	mutex_init(&srv->paths_mutex);
1335 	mutex_init(&srv->paths_ev_mutex);
1336 	uuid_copy(&srv->paths_uuid, paths_uuid);
1337 	srv->queue_depth = sess_queue_depth;
1338 	srv->ctx = ctx;
1339 
1340 	srv->chunks = kcalloc(srv->queue_depth, sizeof(*srv->chunks),
1341 			      GFP_KERNEL);
1342 	if (!srv->chunks)
1343 		goto err_free_srv;
1344 
1345 	for (i = 0; i < srv->queue_depth; i++) {
1346 		srv->chunks[i] = mempool_alloc(chunk_pool, GFP_KERNEL);
1347 		if (!srv->chunks[i])
1348 			goto err_free_chunks;
1349 	}
1350 	list_add(&srv->ctx_list, &ctx->srv_list);
1351 
1352 	return srv;
1353 
1354 err_free_chunks:
1355 	while (i--)
1356 		mempool_free(srv->chunks[i], chunk_pool);
1357 	kfree(srv->chunks);
1358 
1359 err_free_srv:
1360 	kfree(srv);
1361 
1362 	return NULL;
1363 }
1364 
1365 static void free_srv(struct rtrs_srv *srv)
1366 {
1367 	int i;
1368 
1369 	WARN_ON(refcount_read(&srv->refcount));
1370 	for (i = 0; i < srv->queue_depth; i++)
1371 		mempool_free(srv->chunks[i], chunk_pool);
1372 	kfree(srv->chunks);
1373 	mutex_destroy(&srv->paths_mutex);
1374 	mutex_destroy(&srv->paths_ev_mutex);
1375 	/* last put to release the srv structure */
1376 	put_device(&srv->dev);
1377 }
1378 
1379 static inline struct rtrs_srv *__find_srv_and_get(struct rtrs_srv_ctx *ctx,
1380 						   const uuid_t *paths_uuid)
1381 {
1382 	struct rtrs_srv *srv;
1383 
1384 	list_for_each_entry(srv, &ctx->srv_list, ctx_list) {
1385 		if (uuid_equal(&srv->paths_uuid, paths_uuid) &&
1386 		    refcount_inc_not_zero(&srv->refcount))
1387 			return srv;
1388 	}
1389 
1390 	return NULL;
1391 }
1392 
1393 static struct rtrs_srv *get_or_create_srv(struct rtrs_srv_ctx *ctx,
1394 					   const uuid_t *paths_uuid)
1395 {
1396 	struct rtrs_srv *srv;
1397 
1398 	mutex_lock(&ctx->srv_mutex);
1399 	srv = __find_srv_and_get(ctx, paths_uuid);
1400 	if (!srv)
1401 		srv = __alloc_srv(ctx, paths_uuid);
1402 	mutex_unlock(&ctx->srv_mutex);
1403 
1404 	return srv;
1405 }
1406 
1407 static void put_srv(struct rtrs_srv *srv)
1408 {
1409 	if (refcount_dec_and_test(&srv->refcount)) {
1410 		struct rtrs_srv_ctx *ctx = srv->ctx;
1411 
1412 		WARN_ON(srv->dev.kobj.state_in_sysfs);
1413 
1414 		mutex_lock(&ctx->srv_mutex);
1415 		list_del(&srv->ctx_list);
1416 		mutex_unlock(&ctx->srv_mutex);
1417 		free_srv(srv);
1418 	}
1419 }
1420 
1421 static void __add_path_to_srv(struct rtrs_srv *srv,
1422 			      struct rtrs_srv_sess *sess)
1423 {
1424 	list_add_tail(&sess->s.entry, &srv->paths_list);
1425 	srv->paths_num++;
1426 	WARN_ON(srv->paths_num >= MAX_PATHS_NUM);
1427 }
1428 
1429 static void del_path_from_srv(struct rtrs_srv_sess *sess)
1430 {
1431 	struct rtrs_srv *srv = sess->srv;
1432 
1433 	if (WARN_ON(!srv))
1434 		return;
1435 
1436 	mutex_lock(&srv->paths_mutex);
1437 	list_del(&sess->s.entry);
1438 	WARN_ON(!srv->paths_num);
1439 	srv->paths_num--;
1440 	mutex_unlock(&srv->paths_mutex);
1441 }
1442 
1443 /* return true if addresses are the same, error other wise */
1444 static int sockaddr_cmp(const struct sockaddr *a, const struct sockaddr *b)
1445 {
1446 	switch (a->sa_family) {
1447 	case AF_IB:
1448 		return memcmp(&((struct sockaddr_ib *)a)->sib_addr,
1449 			      &((struct sockaddr_ib *)b)->sib_addr,
1450 			      sizeof(struct ib_addr)) &&
1451 			(b->sa_family == AF_IB);
1452 	case AF_INET:
1453 		return memcmp(&((struct sockaddr_in *)a)->sin_addr,
1454 			      &((struct sockaddr_in *)b)->sin_addr,
1455 			      sizeof(struct in_addr)) &&
1456 			(b->sa_family == AF_INET);
1457 	case AF_INET6:
1458 		return memcmp(&((struct sockaddr_in6 *)a)->sin6_addr,
1459 			      &((struct sockaddr_in6 *)b)->sin6_addr,
1460 			      sizeof(struct in6_addr)) &&
1461 			(b->sa_family == AF_INET6);
1462 	default:
1463 		return -ENOENT;
1464 	}
1465 }
1466 
1467 static bool __is_path_w_addr_exists(struct rtrs_srv *srv,
1468 				    struct rdma_addr *addr)
1469 {
1470 	struct rtrs_srv_sess *sess;
1471 
1472 	list_for_each_entry(sess, &srv->paths_list, s.entry)
1473 		if (!sockaddr_cmp((struct sockaddr *)&sess->s.dst_addr,
1474 				  (struct sockaddr *)&addr->dst_addr) &&
1475 		    !sockaddr_cmp((struct sockaddr *)&sess->s.src_addr,
1476 				  (struct sockaddr *)&addr->src_addr))
1477 			return true;
1478 
1479 	return false;
1480 }
1481 
1482 static void free_sess(struct rtrs_srv_sess *sess)
1483 {
1484 	if (sess->kobj.state_in_sysfs)
1485 		kobject_put(&sess->kobj);
1486 	else
1487 		kfree(sess);
1488 }
1489 
1490 static void rtrs_srv_close_work(struct work_struct *work)
1491 {
1492 	struct rtrs_srv_sess *sess;
1493 	struct rtrs_srv_con *con;
1494 	int i;
1495 
1496 	sess = container_of(work, typeof(*sess), close_work);
1497 
1498 	rtrs_srv_destroy_sess_files(sess);
1499 	rtrs_srv_stop_hb(sess);
1500 
1501 	for (i = 0; i < sess->s.con_num; i++) {
1502 		if (!sess->s.con[i])
1503 			continue;
1504 		con = to_srv_con(sess->s.con[i]);
1505 		rdma_disconnect(con->c.cm_id);
1506 		ib_drain_qp(con->c.qp);
1507 	}
1508 	/* Wait for all inflights */
1509 	rtrs_srv_wait_ops_ids(sess);
1510 
1511 	/* Notify upper layer if we are the last path */
1512 	rtrs_srv_sess_down(sess);
1513 
1514 	unmap_cont_bufs(sess);
1515 	rtrs_srv_free_ops_ids(sess);
1516 
1517 	for (i = 0; i < sess->s.con_num; i++) {
1518 		if (!sess->s.con[i])
1519 			continue;
1520 		con = to_srv_con(sess->s.con[i]);
1521 		rtrs_cq_qp_destroy(&con->c);
1522 		rdma_destroy_id(con->c.cm_id);
1523 		kfree(con);
1524 	}
1525 	rtrs_ib_dev_put(sess->s.dev);
1526 
1527 	del_path_from_srv(sess);
1528 	put_srv(sess->srv);
1529 	sess->srv = NULL;
1530 	rtrs_srv_change_state(sess, RTRS_SRV_CLOSED);
1531 
1532 	kfree(sess->dma_addr);
1533 	kfree(sess->s.con);
1534 	free_sess(sess);
1535 }
1536 
1537 static int rtrs_rdma_do_accept(struct rtrs_srv_sess *sess,
1538 			       struct rdma_cm_id *cm_id)
1539 {
1540 	struct rtrs_srv *srv = sess->srv;
1541 	struct rtrs_msg_conn_rsp msg;
1542 	struct rdma_conn_param param;
1543 	int err;
1544 
1545 	param = (struct rdma_conn_param) {
1546 		.rnr_retry_count = 7,
1547 		.private_data = &msg,
1548 		.private_data_len = sizeof(msg),
1549 	};
1550 
1551 	msg = (struct rtrs_msg_conn_rsp) {
1552 		.magic = cpu_to_le16(RTRS_MAGIC),
1553 		.version = cpu_to_le16(RTRS_PROTO_VER),
1554 		.queue_depth = cpu_to_le16(srv->queue_depth),
1555 		.max_io_size = cpu_to_le32(max_chunk_size - MAX_HDR_SIZE),
1556 		.max_hdr_size = cpu_to_le32(MAX_HDR_SIZE),
1557 	};
1558 
1559 	if (always_invalidate)
1560 		msg.flags = cpu_to_le32(RTRS_MSG_NEW_RKEY_F);
1561 
1562 	err = rdma_accept(cm_id, &param);
1563 	if (err)
1564 		pr_err("rdma_accept(), err: %d\n", err);
1565 
1566 	return err;
1567 }
1568 
1569 static int rtrs_rdma_do_reject(struct rdma_cm_id *cm_id, int errno)
1570 {
1571 	struct rtrs_msg_conn_rsp msg;
1572 	int err;
1573 
1574 	msg = (struct rtrs_msg_conn_rsp) {
1575 		.magic = cpu_to_le16(RTRS_MAGIC),
1576 		.version = cpu_to_le16(RTRS_PROTO_VER),
1577 		.errno = cpu_to_le16(errno),
1578 	};
1579 
1580 	err = rdma_reject(cm_id, &msg, sizeof(msg), IB_CM_REJ_CONSUMER_DEFINED);
1581 	if (err)
1582 		pr_err("rdma_reject(), err: %d\n", err);
1583 
1584 	/* Bounce errno back */
1585 	return errno;
1586 }
1587 
1588 static struct rtrs_srv_sess *
1589 __find_sess(struct rtrs_srv *srv, const uuid_t *sess_uuid)
1590 {
1591 	struct rtrs_srv_sess *sess;
1592 
1593 	list_for_each_entry(sess, &srv->paths_list, s.entry) {
1594 		if (uuid_equal(&sess->s.uuid, sess_uuid))
1595 			return sess;
1596 	}
1597 
1598 	return NULL;
1599 }
1600 
1601 static int create_con(struct rtrs_srv_sess *sess,
1602 		      struct rdma_cm_id *cm_id,
1603 		      unsigned int cid)
1604 {
1605 	struct rtrs_srv *srv = sess->srv;
1606 	struct rtrs_sess *s = &sess->s;
1607 	struct rtrs_srv_con *con;
1608 
1609 	u16 cq_size, wr_queue_size;
1610 	int err, cq_vector;
1611 
1612 	con = kzalloc(sizeof(*con), GFP_KERNEL);
1613 	if (!con) {
1614 		err = -ENOMEM;
1615 		goto err;
1616 	}
1617 
1618 	spin_lock_init(&con->rsp_wr_wait_lock);
1619 	INIT_LIST_HEAD(&con->rsp_wr_wait_list);
1620 	con->c.cm_id = cm_id;
1621 	con->c.sess = &sess->s;
1622 	con->c.cid = cid;
1623 	atomic_set(&con->wr_cnt, 0);
1624 
1625 	if (con->c.cid == 0) {
1626 		/*
1627 		 * All receive and all send (each requiring invalidate)
1628 		 * + 2 for drain and heartbeat
1629 		 */
1630 		wr_queue_size = SERVICE_CON_QUEUE_DEPTH * 3 + 2;
1631 		cq_size = wr_queue_size;
1632 	} else {
1633 		/*
1634 		 * If we have all receive requests posted and
1635 		 * all write requests posted and each read request
1636 		 * requires an invalidate request + drain
1637 		 * and qp gets into error state.
1638 		 */
1639 		cq_size = srv->queue_depth * 3 + 1;
1640 		/*
1641 		 * In theory we might have queue_depth * 32
1642 		 * outstanding requests if an unsafe global key is used
1643 		 * and we have queue_depth read requests each consisting
1644 		 * of 32 different addresses. div 3 for mlx5.
1645 		 */
1646 		wr_queue_size = sess->s.dev->ib_dev->attrs.max_qp_wr / 3;
1647 	}
1648 	atomic_set(&con->sq_wr_avail, wr_queue_size);
1649 	cq_vector = rtrs_srv_get_next_cq_vector(sess);
1650 
1651 	/* TODO: SOFTIRQ can be faster, but be careful with softirq context */
1652 	err = rtrs_cq_qp_create(&sess->s, &con->c, 1, cq_vector, cq_size,
1653 				 wr_queue_size, IB_POLL_WORKQUEUE);
1654 	if (err) {
1655 		rtrs_err(s, "rtrs_cq_qp_create(), err: %d\n", err);
1656 		goto free_con;
1657 	}
1658 	if (con->c.cid == 0) {
1659 		err = post_recv_info_req(con);
1660 		if (err)
1661 			goto free_cqqp;
1662 	}
1663 	WARN_ON(sess->s.con[cid]);
1664 	sess->s.con[cid] = &con->c;
1665 
1666 	/*
1667 	 * Change context from server to current connection.  The other
1668 	 * way is to use cm_id->qp->qp_context, which does not work on OFED.
1669 	 */
1670 	cm_id->context = &con->c;
1671 
1672 	return 0;
1673 
1674 free_cqqp:
1675 	rtrs_cq_qp_destroy(&con->c);
1676 free_con:
1677 	kfree(con);
1678 
1679 err:
1680 	return err;
1681 }
1682 
1683 static struct rtrs_srv_sess *__alloc_sess(struct rtrs_srv *srv,
1684 					   struct rdma_cm_id *cm_id,
1685 					   unsigned int con_num,
1686 					   unsigned int recon_cnt,
1687 					   const uuid_t *uuid)
1688 {
1689 	struct rtrs_srv_sess *sess;
1690 	int err = -ENOMEM;
1691 
1692 	if (srv->paths_num >= MAX_PATHS_NUM) {
1693 		err = -ECONNRESET;
1694 		goto err;
1695 	}
1696 	if (__is_path_w_addr_exists(srv, &cm_id->route.addr)) {
1697 		err = -EEXIST;
1698 		pr_err("Path with same addr exists\n");
1699 		goto err;
1700 	}
1701 	sess = kzalloc(sizeof(*sess), GFP_KERNEL);
1702 	if (!sess)
1703 		goto err;
1704 
1705 	sess->stats = kzalloc(sizeof(*sess->stats), GFP_KERNEL);
1706 	if (!sess->stats)
1707 		goto err_free_sess;
1708 
1709 	sess->stats->sess = sess;
1710 
1711 	sess->dma_addr = kcalloc(srv->queue_depth, sizeof(*sess->dma_addr),
1712 				 GFP_KERNEL);
1713 	if (!sess->dma_addr)
1714 		goto err_free_stats;
1715 
1716 	sess->s.con = kcalloc(con_num, sizeof(*sess->s.con), GFP_KERNEL);
1717 	if (!sess->s.con)
1718 		goto err_free_dma_addr;
1719 
1720 	sess->state = RTRS_SRV_CONNECTING;
1721 	sess->srv = srv;
1722 	sess->cur_cq_vector = -1;
1723 	sess->s.dst_addr = cm_id->route.addr.dst_addr;
1724 	sess->s.src_addr = cm_id->route.addr.src_addr;
1725 	sess->s.con_num = con_num;
1726 	sess->s.recon_cnt = recon_cnt;
1727 	uuid_copy(&sess->s.uuid, uuid);
1728 	spin_lock_init(&sess->state_lock);
1729 	INIT_WORK(&sess->close_work, rtrs_srv_close_work);
1730 	rtrs_srv_init_hb(sess);
1731 
1732 	sess->s.dev = rtrs_ib_dev_find_or_add(cm_id->device, &dev_pd);
1733 	if (!sess->s.dev) {
1734 		err = -ENOMEM;
1735 		goto err_free_con;
1736 	}
1737 	err = map_cont_bufs(sess);
1738 	if (err)
1739 		goto err_put_dev;
1740 
1741 	err = rtrs_srv_alloc_ops_ids(sess);
1742 	if (err)
1743 		goto err_unmap_bufs;
1744 
1745 	__add_path_to_srv(srv, sess);
1746 
1747 	return sess;
1748 
1749 err_unmap_bufs:
1750 	unmap_cont_bufs(sess);
1751 err_put_dev:
1752 	rtrs_ib_dev_put(sess->s.dev);
1753 err_free_con:
1754 	kfree(sess->s.con);
1755 err_free_dma_addr:
1756 	kfree(sess->dma_addr);
1757 err_free_stats:
1758 	kfree(sess->stats);
1759 err_free_sess:
1760 	kfree(sess);
1761 err:
1762 	return ERR_PTR(err);
1763 }
1764 
1765 static int rtrs_rdma_connect(struct rdma_cm_id *cm_id,
1766 			      const struct rtrs_msg_conn_req *msg,
1767 			      size_t len)
1768 {
1769 	struct rtrs_srv_ctx *ctx = cm_id->context;
1770 	struct rtrs_srv_sess *sess;
1771 	struct rtrs_srv *srv;
1772 
1773 	u16 version, con_num, cid;
1774 	u16 recon_cnt;
1775 	int err;
1776 
1777 	if (len < sizeof(*msg)) {
1778 		pr_err("Invalid RTRS connection request\n");
1779 		goto reject_w_econnreset;
1780 	}
1781 	if (le16_to_cpu(msg->magic) != RTRS_MAGIC) {
1782 		pr_err("Invalid RTRS magic\n");
1783 		goto reject_w_econnreset;
1784 	}
1785 	version = le16_to_cpu(msg->version);
1786 	if (version >> 8 != RTRS_PROTO_VER_MAJOR) {
1787 		pr_err("Unsupported major RTRS version: %d, expected %d\n",
1788 		       version >> 8, RTRS_PROTO_VER_MAJOR);
1789 		goto reject_w_econnreset;
1790 	}
1791 	con_num = le16_to_cpu(msg->cid_num);
1792 	if (con_num > 4096) {
1793 		/* Sanity check */
1794 		pr_err("Too many connections requested: %d\n", con_num);
1795 		goto reject_w_econnreset;
1796 	}
1797 	cid = le16_to_cpu(msg->cid);
1798 	if (cid >= con_num) {
1799 		/* Sanity check */
1800 		pr_err("Incorrect cid: %d >= %d\n", cid, con_num);
1801 		goto reject_w_econnreset;
1802 	}
1803 	recon_cnt = le16_to_cpu(msg->recon_cnt);
1804 	srv = get_or_create_srv(ctx, &msg->paths_uuid);
1805 	if (!srv) {
1806 		err = -ENOMEM;
1807 		goto reject_w_err;
1808 	}
1809 	mutex_lock(&srv->paths_mutex);
1810 	sess = __find_sess(srv, &msg->sess_uuid);
1811 	if (sess) {
1812 		struct rtrs_sess *s = &sess->s;
1813 
1814 		/* Session already holds a reference */
1815 		put_srv(srv);
1816 
1817 		if (sess->state != RTRS_SRV_CONNECTING) {
1818 			rtrs_err(s, "Session in wrong state: %s\n",
1819 				  rtrs_srv_state_str(sess->state));
1820 			mutex_unlock(&srv->paths_mutex);
1821 			goto reject_w_econnreset;
1822 		}
1823 		/*
1824 		 * Sanity checks
1825 		 */
1826 		if (con_num != s->con_num || cid >= s->con_num) {
1827 			rtrs_err(s, "Incorrect request: %d, %d\n",
1828 				  cid, con_num);
1829 			mutex_unlock(&srv->paths_mutex);
1830 			goto reject_w_econnreset;
1831 		}
1832 		if (s->con[cid]) {
1833 			rtrs_err(s, "Connection already exists: %d\n",
1834 				  cid);
1835 			mutex_unlock(&srv->paths_mutex);
1836 			goto reject_w_econnreset;
1837 		}
1838 	} else {
1839 		sess = __alloc_sess(srv, cm_id, con_num, recon_cnt,
1840 				    &msg->sess_uuid);
1841 		if (IS_ERR(sess)) {
1842 			mutex_unlock(&srv->paths_mutex);
1843 			put_srv(srv);
1844 			err = PTR_ERR(sess);
1845 			goto reject_w_err;
1846 		}
1847 	}
1848 	err = create_con(sess, cm_id, cid);
1849 	if (err) {
1850 		(void)rtrs_rdma_do_reject(cm_id, err);
1851 		/*
1852 		 * Since session has other connections we follow normal way
1853 		 * through workqueue, but still return an error to tell cma.c
1854 		 * to call rdma_destroy_id() for current connection.
1855 		 */
1856 		goto close_and_return_err;
1857 	}
1858 	err = rtrs_rdma_do_accept(sess, cm_id);
1859 	if (err) {
1860 		(void)rtrs_rdma_do_reject(cm_id, err);
1861 		/*
1862 		 * Since current connection was successfully added to the
1863 		 * session we follow normal way through workqueue to close the
1864 		 * session, thus return 0 to tell cma.c we call
1865 		 * rdma_destroy_id() ourselves.
1866 		 */
1867 		err = 0;
1868 		goto close_and_return_err;
1869 	}
1870 	mutex_unlock(&srv->paths_mutex);
1871 
1872 	return 0;
1873 
1874 reject_w_err:
1875 	return rtrs_rdma_do_reject(cm_id, err);
1876 
1877 reject_w_econnreset:
1878 	return rtrs_rdma_do_reject(cm_id, -ECONNRESET);
1879 
1880 close_and_return_err:
1881 	close_sess(sess);
1882 	mutex_unlock(&srv->paths_mutex);
1883 
1884 	return err;
1885 }
1886 
1887 static int rtrs_srv_rdma_cm_handler(struct rdma_cm_id *cm_id,
1888 				     struct rdma_cm_event *ev)
1889 {
1890 	struct rtrs_srv_sess *sess = NULL;
1891 	struct rtrs_sess *s = NULL;
1892 
1893 	if (ev->event != RDMA_CM_EVENT_CONNECT_REQUEST) {
1894 		struct rtrs_con *c = cm_id->context;
1895 
1896 		s = c->sess;
1897 		sess = to_srv_sess(s);
1898 	}
1899 
1900 	switch (ev->event) {
1901 	case RDMA_CM_EVENT_CONNECT_REQUEST:
1902 		/*
1903 		 * In case of error cma.c will destroy cm_id,
1904 		 * see cma_process_remove()
1905 		 */
1906 		return rtrs_rdma_connect(cm_id, ev->param.conn.private_data,
1907 					  ev->param.conn.private_data_len);
1908 	case RDMA_CM_EVENT_ESTABLISHED:
1909 		/* Nothing here */
1910 		break;
1911 	case RDMA_CM_EVENT_REJECTED:
1912 	case RDMA_CM_EVENT_CONNECT_ERROR:
1913 	case RDMA_CM_EVENT_UNREACHABLE:
1914 		rtrs_err(s, "CM error (CM event: %s, err: %d)\n",
1915 			  rdma_event_msg(ev->event), ev->status);
1916 		close_sess(sess);
1917 		break;
1918 	case RDMA_CM_EVENT_DISCONNECTED:
1919 	case RDMA_CM_EVENT_ADDR_CHANGE:
1920 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1921 		close_sess(sess);
1922 		break;
1923 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1924 		close_sess(sess);
1925 		break;
1926 	default:
1927 		pr_err("Ignoring unexpected CM event %s, err %d\n",
1928 		       rdma_event_msg(ev->event), ev->status);
1929 		break;
1930 	}
1931 
1932 	return 0;
1933 }
1934 
1935 static struct rdma_cm_id *rtrs_srv_cm_init(struct rtrs_srv_ctx *ctx,
1936 					    struct sockaddr *addr,
1937 					    enum rdma_ucm_port_space ps)
1938 {
1939 	struct rdma_cm_id *cm_id;
1940 	int ret;
1941 
1942 	cm_id = rdma_create_id(&init_net, rtrs_srv_rdma_cm_handler,
1943 			       ctx, ps, IB_QPT_RC);
1944 	if (IS_ERR(cm_id)) {
1945 		ret = PTR_ERR(cm_id);
1946 		pr_err("Creating id for RDMA connection failed, err: %d\n",
1947 		       ret);
1948 		goto err_out;
1949 	}
1950 	ret = rdma_bind_addr(cm_id, addr);
1951 	if (ret) {
1952 		pr_err("Binding RDMA address failed, err: %d\n", ret);
1953 		goto err_cm;
1954 	}
1955 	ret = rdma_listen(cm_id, 64);
1956 	if (ret) {
1957 		pr_err("Listening on RDMA connection failed, err: %d\n",
1958 		       ret);
1959 		goto err_cm;
1960 	}
1961 
1962 	return cm_id;
1963 
1964 err_cm:
1965 	rdma_destroy_id(cm_id);
1966 err_out:
1967 
1968 	return ERR_PTR(ret);
1969 }
1970 
1971 static int rtrs_srv_rdma_init(struct rtrs_srv_ctx *ctx, u16 port)
1972 {
1973 	struct sockaddr_in6 sin = {
1974 		.sin6_family	= AF_INET6,
1975 		.sin6_addr	= IN6ADDR_ANY_INIT,
1976 		.sin6_port	= htons(port),
1977 	};
1978 	struct sockaddr_ib sib = {
1979 		.sib_family			= AF_IB,
1980 		.sib_sid	= cpu_to_be64(RDMA_IB_IP_PS_IB | port),
1981 		.sib_sid_mask	= cpu_to_be64(0xffffffffffffffffULL),
1982 		.sib_pkey	= cpu_to_be16(0xffff),
1983 	};
1984 	struct rdma_cm_id *cm_ip, *cm_ib;
1985 	int ret;
1986 
1987 	/*
1988 	 * We accept both IPoIB and IB connections, so we need to keep
1989 	 * two cm id's, one for each socket type and port space.
1990 	 * If the cm initialization of one of the id's fails, we abort
1991 	 * everything.
1992 	 */
1993 	cm_ip = rtrs_srv_cm_init(ctx, (struct sockaddr *)&sin, RDMA_PS_TCP);
1994 	if (IS_ERR(cm_ip))
1995 		return PTR_ERR(cm_ip);
1996 
1997 	cm_ib = rtrs_srv_cm_init(ctx, (struct sockaddr *)&sib, RDMA_PS_IB);
1998 	if (IS_ERR(cm_ib)) {
1999 		ret = PTR_ERR(cm_ib);
2000 		goto free_cm_ip;
2001 	}
2002 
2003 	ctx->cm_id_ip = cm_ip;
2004 	ctx->cm_id_ib = cm_ib;
2005 
2006 	return 0;
2007 
2008 free_cm_ip:
2009 	rdma_destroy_id(cm_ip);
2010 
2011 	return ret;
2012 }
2013 
2014 static struct rtrs_srv_ctx *alloc_srv_ctx(struct rtrs_srv_ops *ops)
2015 {
2016 	struct rtrs_srv_ctx *ctx;
2017 
2018 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2019 	if (!ctx)
2020 		return NULL;
2021 
2022 	ctx->ops = *ops;
2023 	mutex_init(&ctx->srv_mutex);
2024 	INIT_LIST_HEAD(&ctx->srv_list);
2025 
2026 	return ctx;
2027 }
2028 
2029 static void free_srv_ctx(struct rtrs_srv_ctx *ctx)
2030 {
2031 	WARN_ON(!list_empty(&ctx->srv_list));
2032 	mutex_destroy(&ctx->srv_mutex);
2033 	kfree(ctx);
2034 }
2035 
2036 /**
2037  * rtrs_srv_open() - open RTRS server context
2038  * @ops:		callback functions
2039  * @port:               port to listen on
2040  *
2041  * Creates server context with specified callbacks.
2042  *
2043  * Return a valid pointer on success otherwise PTR_ERR.
2044  */
2045 struct rtrs_srv_ctx *rtrs_srv_open(struct rtrs_srv_ops *ops, u16 port)
2046 {
2047 	struct rtrs_srv_ctx *ctx;
2048 	int err;
2049 
2050 	ctx = alloc_srv_ctx(ops);
2051 	if (!ctx)
2052 		return ERR_PTR(-ENOMEM);
2053 
2054 	err = rtrs_srv_rdma_init(ctx, port);
2055 	if (err) {
2056 		free_srv_ctx(ctx);
2057 		return ERR_PTR(err);
2058 	}
2059 
2060 	return ctx;
2061 }
2062 EXPORT_SYMBOL(rtrs_srv_open);
2063 
2064 static void close_sessions(struct rtrs_srv *srv)
2065 {
2066 	struct rtrs_srv_sess *sess;
2067 
2068 	mutex_lock(&srv->paths_mutex);
2069 	list_for_each_entry(sess, &srv->paths_list, s.entry)
2070 		close_sess(sess);
2071 	mutex_unlock(&srv->paths_mutex);
2072 }
2073 
2074 static void close_ctx(struct rtrs_srv_ctx *ctx)
2075 {
2076 	struct rtrs_srv *srv;
2077 
2078 	mutex_lock(&ctx->srv_mutex);
2079 	list_for_each_entry(srv, &ctx->srv_list, ctx_list)
2080 		close_sessions(srv);
2081 	mutex_unlock(&ctx->srv_mutex);
2082 	flush_workqueue(rtrs_wq);
2083 }
2084 
2085 /**
2086  * rtrs_srv_close() - close RTRS server context
2087  * @ctx: pointer to server context
2088  *
2089  * Closes RTRS server context with all client sessions.
2090  */
2091 void rtrs_srv_close(struct rtrs_srv_ctx *ctx)
2092 {
2093 	rdma_destroy_id(ctx->cm_id_ip);
2094 	rdma_destroy_id(ctx->cm_id_ib);
2095 	close_ctx(ctx);
2096 	free_srv_ctx(ctx);
2097 }
2098 EXPORT_SYMBOL(rtrs_srv_close);
2099 
2100 static int check_module_params(void)
2101 {
2102 	if (sess_queue_depth < 1 || sess_queue_depth > MAX_SESS_QUEUE_DEPTH) {
2103 		pr_err("Invalid sess_queue_depth value %d, has to be >= %d, <= %d.\n",
2104 		       sess_queue_depth, 1, MAX_SESS_QUEUE_DEPTH);
2105 		return -EINVAL;
2106 	}
2107 	if (max_chunk_size < 4096 || !is_power_of_2(max_chunk_size)) {
2108 		pr_err("Invalid max_chunk_size value %d, has to be >= %d and should be power of two.\n",
2109 		       max_chunk_size, 4096);
2110 		return -EINVAL;
2111 	}
2112 
2113 	/*
2114 	 * Check if IB immediate data size is enough to hold the mem_id and the
2115 	 * offset inside the memory chunk
2116 	 */
2117 	if ((ilog2(sess_queue_depth - 1) + 1) +
2118 	    (ilog2(max_chunk_size - 1) + 1) > MAX_IMM_PAYL_BITS) {
2119 		pr_err("RDMA immediate size (%db) not enough to encode %d buffers of size %dB. Reduce 'sess_queue_depth' or 'max_chunk_size' parameters.\n",
2120 		       MAX_IMM_PAYL_BITS, sess_queue_depth, max_chunk_size);
2121 		return -EINVAL;
2122 	}
2123 
2124 	return 0;
2125 }
2126 
2127 static int __init rtrs_server_init(void)
2128 {
2129 	int err;
2130 
2131 	pr_info("Loading module %s, proto %s: (max_chunk_size: %d (pure IO %ld, headers %ld) , sess_queue_depth: %d, always_invalidate: %d)\n",
2132 		KBUILD_MODNAME, RTRS_PROTO_VER_STRING,
2133 		max_chunk_size, max_chunk_size - MAX_HDR_SIZE, MAX_HDR_SIZE,
2134 		sess_queue_depth, always_invalidate);
2135 
2136 	rtrs_rdma_dev_pd_init(0, &dev_pd);
2137 
2138 	err = check_module_params();
2139 	if (err) {
2140 		pr_err("Failed to load module, invalid module parameters, err: %d\n",
2141 		       err);
2142 		return err;
2143 	}
2144 	chunk_pool = mempool_create_page_pool(sess_queue_depth * CHUNK_POOL_SZ,
2145 					      get_order(max_chunk_size));
2146 	if (!chunk_pool)
2147 		return -ENOMEM;
2148 	rtrs_dev_class = class_create(THIS_MODULE, "rtrs-server");
2149 	if (IS_ERR(rtrs_dev_class)) {
2150 		err = PTR_ERR(rtrs_dev_class);
2151 		goto out_chunk_pool;
2152 	}
2153 	rtrs_wq = alloc_workqueue("rtrs_server_wq", 0, 0);
2154 	if (!rtrs_wq) {
2155 		err = -ENOMEM;
2156 		goto out_dev_class;
2157 	}
2158 
2159 	return 0;
2160 
2161 out_dev_class:
2162 	class_destroy(rtrs_dev_class);
2163 out_chunk_pool:
2164 	mempool_destroy(chunk_pool);
2165 
2166 	return err;
2167 }
2168 
2169 static void __exit rtrs_server_exit(void)
2170 {
2171 	destroy_workqueue(rtrs_wq);
2172 	class_destroy(rtrs_dev_class);
2173 	mempool_destroy(chunk_pool);
2174 	rtrs_rdma_dev_pd_deinit(&dev_pd);
2175 }
2176 
2177 module_init(rtrs_server_init);
2178 module_exit(rtrs_server_exit);
2179