xref: /openbmc/linux/net/9p/trans_rdma.c (revision 2a598d0b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * RDMA transport layer based on the trans_fd.c implementation.
4  *
5  *  Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
6  *  Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
7  *  Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
8  *  Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
9  *  Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
10  */
11 
12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 
14 #include <linux/in.h>
15 #include <linux/module.h>
16 #include <linux/net.h>
17 #include <linux/ipv6.h>
18 #include <linux/kthread.h>
19 #include <linux/errno.h>
20 #include <linux/kernel.h>
21 #include <linux/un.h>
22 #include <linux/uaccess.h>
23 #include <linux/inet.h>
24 #include <linux/file.h>
25 #include <linux/parser.h>
26 #include <linux/semaphore.h>
27 #include <linux/slab.h>
28 #include <linux/seq_file.h>
29 #include <net/9p/9p.h>
30 #include <net/9p/client.h>
31 #include <net/9p/transport.h>
32 #include <rdma/ib_verbs.h>
33 #include <rdma/rdma_cm.h>
34 
35 #define P9_PORT			5640
36 #define P9_RDMA_SQ_DEPTH	32
37 #define P9_RDMA_RQ_DEPTH	32
38 #define P9_RDMA_SEND_SGE	4
39 #define P9_RDMA_RECV_SGE	4
40 #define P9_RDMA_IRD		0
41 #define P9_RDMA_ORD		0
42 #define P9_RDMA_TIMEOUT		30000		/* 30 seconds */
43 #define P9_RDMA_MAXSIZE		(1024*1024)	/* 1MB */
44 
45 /**
46  * struct p9_trans_rdma - RDMA transport instance
47  *
48  * @state: tracks the transport state machine for connection setup and tear down
49  * @cm_id: The RDMA CM ID
50  * @pd: Protection Domain pointer
51  * @qp: Queue Pair pointer
52  * @cq: Completion Queue pointer
53  * @timeout: Number of uSecs to wait for connection management events
54  * @privport: Whether a privileged port may be used
55  * @port: The port to use
56  * @sq_depth: The depth of the Send Queue
57  * @sq_sem: Semaphore for the SQ
58  * @rq_depth: The depth of the Receive Queue.
59  * @rq_sem: Semaphore for the RQ
60  * @excess_rc : Amount of posted Receive Contexts without a pending request.
61  *		See rdma_request()
62  * @addr: The remote peer's address
63  * @req_lock: Protects the active request list
64  * @cm_done: Completion event for connection management tracking
65  */
66 struct p9_trans_rdma {
67 	enum {
68 		P9_RDMA_INIT,
69 		P9_RDMA_ADDR_RESOLVED,
70 		P9_RDMA_ROUTE_RESOLVED,
71 		P9_RDMA_CONNECTED,
72 		P9_RDMA_FLUSHING,
73 		P9_RDMA_CLOSING,
74 		P9_RDMA_CLOSED,
75 	} state;
76 	struct rdma_cm_id *cm_id;
77 	struct ib_pd *pd;
78 	struct ib_qp *qp;
79 	struct ib_cq *cq;
80 	long timeout;
81 	bool privport;
82 	u16 port;
83 	int sq_depth;
84 	struct semaphore sq_sem;
85 	int rq_depth;
86 	struct semaphore rq_sem;
87 	atomic_t excess_rc;
88 	struct sockaddr_in addr;
89 	spinlock_t req_lock;
90 
91 	struct completion cm_done;
92 };
93 
94 struct p9_rdma_req;
95 
96 /**
97  * struct p9_rdma_context - Keeps track of in-process WR
98  *
99  * @cqe: completion queue entry
100  * @busa: Bus address to unmap when the WR completes
101  * @req: Keeps track of requests (send)
102  * @rc: Keepts track of replies (receive)
103  */
104 struct p9_rdma_context {
105 	struct ib_cqe cqe;
106 	dma_addr_t busa;
107 	union {
108 		struct p9_req_t *req;
109 		struct p9_fcall rc;
110 	};
111 };
112 
113 /**
114  * struct p9_rdma_opts - Collection of mount options
115  * @port: port of connection
116  * @privport: Whether a privileged port may be used
117  * @sq_depth: The requested depth of the SQ. This really doesn't need
118  * to be any deeper than the number of threads used in the client
119  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
120  * @timeout: Time to wait in msecs for CM events
121  */
122 struct p9_rdma_opts {
123 	short port;
124 	bool privport;
125 	int sq_depth;
126 	int rq_depth;
127 	long timeout;
128 };
129 
130 /*
131  * Option Parsing (code inspired by NFS code)
132  */
133 enum {
134 	/* Options that take integer arguments */
135 	Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
136 	/* Options that take no argument */
137 	Opt_privport,
138 	Opt_err,
139 };
140 
141 static match_table_t tokens = {
142 	{Opt_port, "port=%u"},
143 	{Opt_sq_depth, "sq=%u"},
144 	{Opt_rq_depth, "rq=%u"},
145 	{Opt_timeout, "timeout=%u"},
146 	{Opt_privport, "privport"},
147 	{Opt_err, NULL},
148 };
149 
150 static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
151 {
152 	struct p9_trans_rdma *rdma = clnt->trans;
153 
154 	if (rdma->port != P9_PORT)
155 		seq_printf(m, ",port=%u", rdma->port);
156 	if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
157 		seq_printf(m, ",sq=%u", rdma->sq_depth);
158 	if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
159 		seq_printf(m, ",rq=%u", rdma->rq_depth);
160 	if (rdma->timeout != P9_RDMA_TIMEOUT)
161 		seq_printf(m, ",timeout=%lu", rdma->timeout);
162 	if (rdma->privport)
163 		seq_puts(m, ",privport");
164 	return 0;
165 }
166 
167 /**
168  * parse_opts - parse mount options into rdma options structure
169  * @params: options string passed from mount
170  * @opts: rdma transport-specific structure to parse options into
171  *
172  * Returns 0 upon success, -ERRNO upon failure
173  */
174 static int parse_opts(char *params, struct p9_rdma_opts *opts)
175 {
176 	char *p;
177 	substring_t args[MAX_OPT_ARGS];
178 	int option;
179 	char *options, *tmp_options;
180 
181 	opts->port = P9_PORT;
182 	opts->sq_depth = P9_RDMA_SQ_DEPTH;
183 	opts->rq_depth = P9_RDMA_RQ_DEPTH;
184 	opts->timeout = P9_RDMA_TIMEOUT;
185 	opts->privport = false;
186 
187 	if (!params)
188 		return 0;
189 
190 	tmp_options = kstrdup(params, GFP_KERNEL);
191 	if (!tmp_options) {
192 		p9_debug(P9_DEBUG_ERROR,
193 			 "failed to allocate copy of option string\n");
194 		return -ENOMEM;
195 	}
196 	options = tmp_options;
197 
198 	while ((p = strsep(&options, ",")) != NULL) {
199 		int token;
200 		int r;
201 		if (!*p)
202 			continue;
203 		token = match_token(p, tokens, args);
204 		if ((token != Opt_err) && (token != Opt_privport)) {
205 			r = match_int(&args[0], &option);
206 			if (r < 0) {
207 				p9_debug(P9_DEBUG_ERROR,
208 					 "integer field, but no integer?\n");
209 				continue;
210 			}
211 		}
212 		switch (token) {
213 		case Opt_port:
214 			opts->port = option;
215 			break;
216 		case Opt_sq_depth:
217 			opts->sq_depth = option;
218 			break;
219 		case Opt_rq_depth:
220 			opts->rq_depth = option;
221 			break;
222 		case Opt_timeout:
223 			opts->timeout = option;
224 			break;
225 		case Opt_privport:
226 			opts->privport = true;
227 			break;
228 		default:
229 			continue;
230 		}
231 	}
232 	/* RQ must be at least as large as the SQ */
233 	opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
234 	kfree(tmp_options);
235 	return 0;
236 }
237 
238 static int
239 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
240 {
241 	struct p9_client *c = id->context;
242 	struct p9_trans_rdma *rdma = c->trans;
243 	switch (event->event) {
244 	case RDMA_CM_EVENT_ADDR_RESOLVED:
245 		BUG_ON(rdma->state != P9_RDMA_INIT);
246 		rdma->state = P9_RDMA_ADDR_RESOLVED;
247 		break;
248 
249 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
250 		BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
251 		rdma->state = P9_RDMA_ROUTE_RESOLVED;
252 		break;
253 
254 	case RDMA_CM_EVENT_ESTABLISHED:
255 		BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
256 		rdma->state = P9_RDMA_CONNECTED;
257 		break;
258 
259 	case RDMA_CM_EVENT_DISCONNECTED:
260 		if (rdma)
261 			rdma->state = P9_RDMA_CLOSED;
262 		c->status = Disconnected;
263 		break;
264 
265 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
266 		break;
267 
268 	case RDMA_CM_EVENT_ADDR_CHANGE:
269 	case RDMA_CM_EVENT_ROUTE_ERROR:
270 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
271 	case RDMA_CM_EVENT_MULTICAST_JOIN:
272 	case RDMA_CM_EVENT_MULTICAST_ERROR:
273 	case RDMA_CM_EVENT_REJECTED:
274 	case RDMA_CM_EVENT_CONNECT_REQUEST:
275 	case RDMA_CM_EVENT_CONNECT_RESPONSE:
276 	case RDMA_CM_EVENT_CONNECT_ERROR:
277 	case RDMA_CM_EVENT_ADDR_ERROR:
278 	case RDMA_CM_EVENT_UNREACHABLE:
279 		c->status = Disconnected;
280 		rdma_disconnect(rdma->cm_id);
281 		break;
282 	default:
283 		BUG();
284 	}
285 	complete(&rdma->cm_done);
286 	return 0;
287 }
288 
289 static void
290 recv_done(struct ib_cq *cq, struct ib_wc *wc)
291 {
292 	struct p9_client *client = cq->cq_context;
293 	struct p9_trans_rdma *rdma = client->trans;
294 	struct p9_rdma_context *c =
295 		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
296 	struct p9_req_t *req;
297 	int err = 0;
298 	int16_t tag;
299 
300 	req = NULL;
301 	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
302 							 DMA_FROM_DEVICE);
303 
304 	if (wc->status != IB_WC_SUCCESS)
305 		goto err_out;
306 
307 	c->rc.size = wc->byte_len;
308 	err = p9_parse_header(&c->rc, NULL, NULL, &tag, 1);
309 	if (err)
310 		goto err_out;
311 
312 	req = p9_tag_lookup(client, tag);
313 	if (!req)
314 		goto err_out;
315 
316 	/* Check that we have not yet received a reply for this request.
317 	 */
318 	if (unlikely(req->rc.sdata)) {
319 		pr_err("Duplicate reply for request %d", tag);
320 		goto err_out;
321 	}
322 
323 	req->rc.size = c->rc.size;
324 	req->rc.sdata = c->rc.sdata;
325 	p9_client_cb(client, req, REQ_STATUS_RCVD);
326 
327  out:
328 	up(&rdma->rq_sem);
329 	kfree(c);
330 	return;
331 
332  err_out:
333 	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
334 			req, err, wc->status);
335 	rdma->state = P9_RDMA_FLUSHING;
336 	client->status = Disconnected;
337 	goto out;
338 }
339 
340 static void
341 send_done(struct ib_cq *cq, struct ib_wc *wc)
342 {
343 	struct p9_client *client = cq->cq_context;
344 	struct p9_trans_rdma *rdma = client->trans;
345 	struct p9_rdma_context *c =
346 		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
347 
348 	ib_dma_unmap_single(rdma->cm_id->device,
349 			    c->busa, c->req->tc.size,
350 			    DMA_TO_DEVICE);
351 	up(&rdma->sq_sem);
352 	p9_req_put(client, c->req);
353 	kfree(c);
354 }
355 
356 static void qp_event_handler(struct ib_event *event, void *context)
357 {
358 	p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
359 		 event->event, context);
360 }
361 
362 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
363 {
364 	if (!rdma)
365 		return;
366 
367 	if (rdma->qp && !IS_ERR(rdma->qp))
368 		ib_destroy_qp(rdma->qp);
369 
370 	if (rdma->pd && !IS_ERR(rdma->pd))
371 		ib_dealloc_pd(rdma->pd);
372 
373 	if (rdma->cq && !IS_ERR(rdma->cq))
374 		ib_free_cq(rdma->cq);
375 
376 	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
377 		rdma_destroy_id(rdma->cm_id);
378 
379 	kfree(rdma);
380 }
381 
382 static int
383 post_recv(struct p9_client *client, struct p9_rdma_context *c)
384 {
385 	struct p9_trans_rdma *rdma = client->trans;
386 	struct ib_recv_wr wr;
387 	struct ib_sge sge;
388 	int ret;
389 
390 	c->busa = ib_dma_map_single(rdma->cm_id->device,
391 				    c->rc.sdata, client->msize,
392 				    DMA_FROM_DEVICE);
393 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
394 		goto error;
395 
396 	c->cqe.done = recv_done;
397 
398 	sge.addr = c->busa;
399 	sge.length = client->msize;
400 	sge.lkey = rdma->pd->local_dma_lkey;
401 
402 	wr.next = NULL;
403 	wr.wr_cqe = &c->cqe;
404 	wr.sg_list = &sge;
405 	wr.num_sge = 1;
406 
407 	ret = ib_post_recv(rdma->qp, &wr, NULL);
408 	if (ret)
409 		ib_dma_unmap_single(rdma->cm_id->device, c->busa,
410 				    client->msize, DMA_FROM_DEVICE);
411 	return ret;
412 
413  error:
414 	p9_debug(P9_DEBUG_ERROR, "EIO\n");
415 	return -EIO;
416 }
417 
418 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
419 {
420 	struct p9_trans_rdma *rdma = client->trans;
421 	struct ib_send_wr wr;
422 	struct ib_sge sge;
423 	int err = 0;
424 	unsigned long flags;
425 	struct p9_rdma_context *c = NULL;
426 	struct p9_rdma_context *rpl_context = NULL;
427 
428 	/* When an error occurs between posting the recv and the send,
429 	 * there will be a receive context posted without a pending request.
430 	 * Since there is no way to "un-post" it, we remember it and skip
431 	 * post_recv() for the next request.
432 	 * So here,
433 	 * see if we are this `next request' and need to absorb an excess rc.
434 	 * If yes, then drop and free our own, and do not recv_post().
435 	 **/
436 	if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
437 		if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
438 			/* Got one! */
439 			p9_fcall_fini(&req->rc);
440 			req->rc.sdata = NULL;
441 			goto dont_need_post_recv;
442 		} else {
443 			/* We raced and lost. */
444 			atomic_inc(&rdma->excess_rc);
445 		}
446 	}
447 
448 	/* Allocate an fcall for the reply */
449 	rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
450 	if (!rpl_context) {
451 		err = -ENOMEM;
452 		goto recv_error;
453 	}
454 	rpl_context->rc.sdata = req->rc.sdata;
455 
456 	/*
457 	 * Post a receive buffer for this request. We need to ensure
458 	 * there is a reply buffer available for every outstanding
459 	 * request. A flushed request can result in no reply for an
460 	 * outstanding request, so we must keep a count to avoid
461 	 * overflowing the RQ.
462 	 */
463 	if (down_interruptible(&rdma->rq_sem)) {
464 		err = -EINTR;
465 		goto recv_error;
466 	}
467 
468 	err = post_recv(client, rpl_context);
469 	if (err) {
470 		p9_debug(P9_DEBUG_ERROR, "POST RECV failed: %d\n", err);
471 		goto recv_error;
472 	}
473 	/* remove posted receive buffer from request structure */
474 	req->rc.sdata = NULL;
475 
476 dont_need_post_recv:
477 	/* Post the request */
478 	c = kmalloc(sizeof *c, GFP_NOFS);
479 	if (!c) {
480 		err = -ENOMEM;
481 		goto send_error;
482 	}
483 	c->req = req;
484 
485 	c->busa = ib_dma_map_single(rdma->cm_id->device,
486 				    c->req->tc.sdata, c->req->tc.size,
487 				    DMA_TO_DEVICE);
488 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
489 		err = -EIO;
490 		goto send_error;
491 	}
492 
493 	c->cqe.done = send_done;
494 
495 	sge.addr = c->busa;
496 	sge.length = c->req->tc.size;
497 	sge.lkey = rdma->pd->local_dma_lkey;
498 
499 	wr.next = NULL;
500 	wr.wr_cqe = &c->cqe;
501 	wr.opcode = IB_WR_SEND;
502 	wr.send_flags = IB_SEND_SIGNALED;
503 	wr.sg_list = &sge;
504 	wr.num_sge = 1;
505 
506 	if (down_interruptible(&rdma->sq_sem)) {
507 		err = -EINTR;
508 		goto dma_unmap;
509 	}
510 
511 	/* Mark request as `sent' *before* we actually send it,
512 	 * because doing if after could erase the REQ_STATUS_RCVD
513 	 * status in case of a very fast reply.
514 	 */
515 	WRITE_ONCE(req->status, REQ_STATUS_SENT);
516 	err = ib_post_send(rdma->qp, &wr, NULL);
517 	if (err)
518 		goto dma_unmap;
519 
520 	/* Success */
521 	return 0;
522 
523 dma_unmap:
524 	ib_dma_unmap_single(rdma->cm_id->device, c->busa,
525 			    c->req->tc.size, DMA_TO_DEVICE);
526  /* Handle errors that happened during or while preparing the send: */
527  send_error:
528 	WRITE_ONCE(req->status, REQ_STATUS_ERROR);
529 	kfree(c);
530 	p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
531 
532 	/* Ach.
533 	 *  We did recv_post(), but not send. We have one recv_post in excess.
534 	 */
535 	atomic_inc(&rdma->excess_rc);
536 	return err;
537 
538  /* Handle errors that happened during or while preparing post_recv(): */
539  recv_error:
540 	kfree(rpl_context);
541 	spin_lock_irqsave(&rdma->req_lock, flags);
542 	if (err != -EINTR && rdma->state < P9_RDMA_CLOSING) {
543 		rdma->state = P9_RDMA_CLOSING;
544 		spin_unlock_irqrestore(&rdma->req_lock, flags);
545 		rdma_disconnect(rdma->cm_id);
546 	} else
547 		spin_unlock_irqrestore(&rdma->req_lock, flags);
548 	return err;
549 }
550 
551 static void rdma_close(struct p9_client *client)
552 {
553 	struct p9_trans_rdma *rdma;
554 
555 	if (!client)
556 		return;
557 
558 	rdma = client->trans;
559 	if (!rdma)
560 		return;
561 
562 	client->status = Disconnected;
563 	rdma_disconnect(rdma->cm_id);
564 	rdma_destroy_trans(rdma);
565 }
566 
567 /**
568  * alloc_rdma - Allocate and initialize the rdma transport structure
569  * @opts: Mount options structure
570  */
571 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
572 {
573 	struct p9_trans_rdma *rdma;
574 
575 	rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
576 	if (!rdma)
577 		return NULL;
578 
579 	rdma->port = opts->port;
580 	rdma->privport = opts->privport;
581 	rdma->sq_depth = opts->sq_depth;
582 	rdma->rq_depth = opts->rq_depth;
583 	rdma->timeout = opts->timeout;
584 	spin_lock_init(&rdma->req_lock);
585 	init_completion(&rdma->cm_done);
586 	sema_init(&rdma->sq_sem, rdma->sq_depth);
587 	sema_init(&rdma->rq_sem, rdma->rq_depth);
588 	atomic_set(&rdma->excess_rc, 0);
589 
590 	return rdma;
591 }
592 
593 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
594 {
595 	/* Nothing to do here.
596 	 * We will take care of it (if we have to) in rdma_cancelled()
597 	 */
598 	return 1;
599 }
600 
601 /* A request has been fully flushed without a reply.
602  * That means we have posted one buffer in excess.
603  */
604 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
605 {
606 	struct p9_trans_rdma *rdma = client->trans;
607 	atomic_inc(&rdma->excess_rc);
608 	return 0;
609 }
610 
611 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
612 {
613 	struct sockaddr_in cl = {
614 		.sin_family = AF_INET,
615 		.sin_addr.s_addr = htonl(INADDR_ANY),
616 	};
617 	int port, err = -EINVAL;
618 
619 	for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
620 		cl.sin_port = htons((ushort)port);
621 		err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
622 		if (err != -EADDRINUSE)
623 			break;
624 	}
625 	return err;
626 }
627 
628 /**
629  * rdma_create_trans - Transport method for creating a transport instance
630  * @client: client instance
631  * @addr: IP address string
632  * @args: Mount options string
633  */
634 static int
635 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
636 {
637 	int err;
638 	struct p9_rdma_opts opts;
639 	struct p9_trans_rdma *rdma;
640 	struct rdma_conn_param conn_param;
641 	struct ib_qp_init_attr qp_attr;
642 
643 	if (addr == NULL)
644 		return -EINVAL;
645 
646 	/* Parse the transport specific mount options */
647 	err = parse_opts(args, &opts);
648 	if (err < 0)
649 		return err;
650 
651 	/* Create and initialize the RDMA transport structure */
652 	rdma = alloc_rdma(&opts);
653 	if (!rdma)
654 		return -ENOMEM;
655 
656 	/* Create the RDMA CM ID */
657 	rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
658 				     RDMA_PS_TCP, IB_QPT_RC);
659 	if (IS_ERR(rdma->cm_id))
660 		goto error;
661 
662 	/* Associate the client with the transport */
663 	client->trans = rdma;
664 
665 	/* Bind to a privileged port if we need to */
666 	if (opts.privport) {
667 		err = p9_rdma_bind_privport(rdma);
668 		if (err < 0) {
669 			pr_err("%s (%d): problem binding to privport: %d\n",
670 			       __func__, task_pid_nr(current), -err);
671 			goto error;
672 		}
673 	}
674 
675 	/* Resolve the server's address */
676 	rdma->addr.sin_family = AF_INET;
677 	rdma->addr.sin_addr.s_addr = in_aton(addr);
678 	rdma->addr.sin_port = htons(opts.port);
679 	err = rdma_resolve_addr(rdma->cm_id, NULL,
680 				(struct sockaddr *)&rdma->addr,
681 				rdma->timeout);
682 	if (err)
683 		goto error;
684 	err = wait_for_completion_interruptible(&rdma->cm_done);
685 	if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
686 		goto error;
687 
688 	/* Resolve the route to the server */
689 	err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
690 	if (err)
691 		goto error;
692 	err = wait_for_completion_interruptible(&rdma->cm_done);
693 	if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
694 		goto error;
695 
696 	/* Create the Completion Queue */
697 	rdma->cq = ib_alloc_cq_any(rdma->cm_id->device, client,
698 				   opts.sq_depth + opts.rq_depth + 1,
699 				   IB_POLL_SOFTIRQ);
700 	if (IS_ERR(rdma->cq))
701 		goto error;
702 
703 	/* Create the Protection Domain */
704 	rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
705 	if (IS_ERR(rdma->pd))
706 		goto error;
707 
708 	/* Create the Queue Pair */
709 	memset(&qp_attr, 0, sizeof qp_attr);
710 	qp_attr.event_handler = qp_event_handler;
711 	qp_attr.qp_context = client;
712 	qp_attr.cap.max_send_wr = opts.sq_depth;
713 	qp_attr.cap.max_recv_wr = opts.rq_depth;
714 	qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
715 	qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
716 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
717 	qp_attr.qp_type = IB_QPT_RC;
718 	qp_attr.send_cq = rdma->cq;
719 	qp_attr.recv_cq = rdma->cq;
720 	err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
721 	if (err)
722 		goto error;
723 	rdma->qp = rdma->cm_id->qp;
724 
725 	/* Request a connection */
726 	memset(&conn_param, 0, sizeof(conn_param));
727 	conn_param.private_data = NULL;
728 	conn_param.private_data_len = 0;
729 	conn_param.responder_resources = P9_RDMA_IRD;
730 	conn_param.initiator_depth = P9_RDMA_ORD;
731 	err = rdma_connect(rdma->cm_id, &conn_param);
732 	if (err)
733 		goto error;
734 	err = wait_for_completion_interruptible(&rdma->cm_done);
735 	if (err || (rdma->state != P9_RDMA_CONNECTED))
736 		goto error;
737 
738 	client->status = Connected;
739 
740 	return 0;
741 
742 error:
743 	rdma_destroy_trans(rdma);
744 	return -ENOTCONN;
745 }
746 
747 static struct p9_trans_module p9_rdma_trans = {
748 	.name = "rdma",
749 	.maxsize = P9_RDMA_MAXSIZE,
750 	.pooled_rbuffers = true,
751 	.def = 0,
752 	.owner = THIS_MODULE,
753 	.create = rdma_create_trans,
754 	.close = rdma_close,
755 	.request = rdma_request,
756 	.cancel = rdma_cancel,
757 	.cancelled = rdma_cancelled,
758 	.show_options = p9_rdma_show_options,
759 };
760 
761 /**
762  * p9_trans_rdma_init - Register the 9P RDMA transport driver
763  */
764 static int __init p9_trans_rdma_init(void)
765 {
766 	v9fs_register_trans(&p9_rdma_trans);
767 	return 0;
768 }
769 
770 static void __exit p9_trans_rdma_exit(void)
771 {
772 	v9fs_unregister_trans(&p9_rdma_trans);
773 }
774 
775 module_init(p9_trans_rdma_init);
776 module_exit(p9_trans_rdma_exit);
777 MODULE_ALIAS_9P("rdma");
778 
779 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
780 MODULE_DESCRIPTION("RDMA Transport for 9P");
781 MODULE_LICENSE("Dual BSD/GPL");
782