xref: /openbmc/linux/net/9p/trans_rdma.c (revision 8bd1369b)
1 /*
2  * linux/fs/9p/trans_rdma.c
3  *
4  * RDMA transport layer based on the trans_fd.c implementation.
5  *
6  *  Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
7  *  Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
8  *  Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
9  *  Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
10  *  Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
11  *
12  *  This program is free software; you can redistribute it and/or modify
13  *  it under the terms of the GNU General Public License version 2
14  *  as published by the Free Software Foundation.
15  *
16  *  This program is distributed in the hope that it will be useful,
17  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
19  *  GNU General Public License for more details.
20  *
21  *  You should have received a copy of the GNU General Public License
22  *  along with this program; if not, write to:
23  *  Free Software Foundation
24  *  51 Franklin Street, Fifth Floor
25  *  Boston, MA  02111-1301  USA
26  *
27  */
28 
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
30 
31 #include <linux/in.h>
32 #include <linux/module.h>
33 #include <linux/net.h>
34 #include <linux/ipv6.h>
35 #include <linux/kthread.h>
36 #include <linux/errno.h>
37 #include <linux/kernel.h>
38 #include <linux/un.h>
39 #include <linux/uaccess.h>
40 #include <linux/inet.h>
41 #include <linux/idr.h>
42 #include <linux/file.h>
43 #include <linux/parser.h>
44 #include <linux/semaphore.h>
45 #include <linux/slab.h>
46 #include <linux/seq_file.h>
47 #include <net/9p/9p.h>
48 #include <net/9p/client.h>
49 #include <net/9p/transport.h>
50 #include <rdma/ib_verbs.h>
51 #include <rdma/rdma_cm.h>
52 
53 #define P9_PORT			5640
54 #define P9_RDMA_SQ_DEPTH	32
55 #define P9_RDMA_RQ_DEPTH	32
56 #define P9_RDMA_SEND_SGE	4
57 #define P9_RDMA_RECV_SGE	4
58 #define P9_RDMA_IRD		0
59 #define P9_RDMA_ORD		0
60 #define P9_RDMA_TIMEOUT		30000		/* 30 seconds */
61 #define P9_RDMA_MAXSIZE		(1024*1024)	/* 1MB */
62 
63 /**
64  * struct p9_trans_rdma - RDMA transport instance
65  *
66  * @state: tracks the transport state machine for connection setup and tear down
67  * @cm_id: The RDMA CM ID
68  * @pd: Protection Domain pointer
69  * @qp: Queue Pair pointer
70  * @cq: Completion Queue pointer
71  * @timeout: Number of uSecs to wait for connection management events
72  * @privport: Whether a privileged port may be used
73  * @port: The port to use
74  * @sq_depth: The depth of the Send Queue
75  * @sq_sem: Semaphore for the SQ
76  * @rq_depth: The depth of the Receive Queue.
77  * @rq_sem: Semaphore for the RQ
78  * @excess_rc : Amount of posted Receive Contexts without a pending request.
79  *		See rdma_request()
80  * @addr: The remote peer's address
81  * @req_lock: Protects the active request list
82  * @cm_done: Completion event for connection management tracking
83  */
84 struct p9_trans_rdma {
85 	enum {
86 		P9_RDMA_INIT,
87 		P9_RDMA_ADDR_RESOLVED,
88 		P9_RDMA_ROUTE_RESOLVED,
89 		P9_RDMA_CONNECTED,
90 		P9_RDMA_FLUSHING,
91 		P9_RDMA_CLOSING,
92 		P9_RDMA_CLOSED,
93 	} state;
94 	struct rdma_cm_id *cm_id;
95 	struct ib_pd *pd;
96 	struct ib_qp *qp;
97 	struct ib_cq *cq;
98 	long timeout;
99 	bool privport;
100 	u16 port;
101 	int sq_depth;
102 	struct semaphore sq_sem;
103 	int rq_depth;
104 	struct semaphore rq_sem;
105 	atomic_t excess_rc;
106 	struct sockaddr_in addr;
107 	spinlock_t req_lock;
108 
109 	struct completion cm_done;
110 };
111 
112 /**
113  * p9_rdma_context - Keeps track of in-process WR
114  *
115  * @busa: Bus address to unmap when the WR completes
116  * @req: Keeps track of requests (send)
117  * @rc: Keepts track of replies (receive)
118  */
119 struct p9_rdma_req;
120 struct p9_rdma_context {
121 	struct ib_cqe cqe;
122 	dma_addr_t busa;
123 	union {
124 		struct p9_req_t *req;
125 		struct p9_fcall *rc;
126 	};
127 };
128 
129 /**
130  * p9_rdma_opts - Collection of mount options
131  * @port: port of connection
132  * @sq_depth: The requested depth of the SQ. This really doesn't need
133  * to be any deeper than the number of threads used in the client
134  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
135  * @timeout: Time to wait in msecs for CM events
136  */
137 struct p9_rdma_opts {
138 	short port;
139 	bool privport;
140 	int sq_depth;
141 	int rq_depth;
142 	long timeout;
143 };
144 
145 /*
146  * Option Parsing (code inspired by NFS code)
147  */
148 enum {
149 	/* Options that take integer arguments */
150 	Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
151 	/* Options that take no argument */
152 	Opt_privport,
153 	Opt_err,
154 };
155 
156 static match_table_t tokens = {
157 	{Opt_port, "port=%u"},
158 	{Opt_sq_depth, "sq=%u"},
159 	{Opt_rq_depth, "rq=%u"},
160 	{Opt_timeout, "timeout=%u"},
161 	{Opt_privport, "privport"},
162 	{Opt_err, NULL},
163 };
164 
165 static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
166 {
167 	struct p9_trans_rdma *rdma = clnt->trans;
168 
169 	if (rdma->port != P9_PORT)
170 		seq_printf(m, ",port=%u", rdma->port);
171 	if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
172 		seq_printf(m, ",sq=%u", rdma->sq_depth);
173 	if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
174 		seq_printf(m, ",rq=%u", rdma->rq_depth);
175 	if (rdma->timeout != P9_RDMA_TIMEOUT)
176 		seq_printf(m, ",timeout=%lu", rdma->timeout);
177 	if (rdma->privport)
178 		seq_puts(m, ",privport");
179 	return 0;
180 }
181 
182 /**
183  * parse_opts - parse mount options into rdma options structure
184  * @params: options string passed from mount
185  * @opts: rdma transport-specific structure to parse options into
186  *
187  * Returns 0 upon success, -ERRNO upon failure
188  */
189 static int parse_opts(char *params, struct p9_rdma_opts *opts)
190 {
191 	char *p;
192 	substring_t args[MAX_OPT_ARGS];
193 	int option;
194 	char *options, *tmp_options;
195 
196 	opts->port = P9_PORT;
197 	opts->sq_depth = P9_RDMA_SQ_DEPTH;
198 	opts->rq_depth = P9_RDMA_RQ_DEPTH;
199 	opts->timeout = P9_RDMA_TIMEOUT;
200 	opts->privport = false;
201 
202 	if (!params)
203 		return 0;
204 
205 	tmp_options = kstrdup(params, GFP_KERNEL);
206 	if (!tmp_options) {
207 		p9_debug(P9_DEBUG_ERROR,
208 			 "failed to allocate copy of option string\n");
209 		return -ENOMEM;
210 	}
211 	options = tmp_options;
212 
213 	while ((p = strsep(&options, ",")) != NULL) {
214 		int token;
215 		int r;
216 		if (!*p)
217 			continue;
218 		token = match_token(p, tokens, args);
219 		if ((token != Opt_err) && (token != Opt_privport)) {
220 			r = match_int(&args[0], &option);
221 			if (r < 0) {
222 				p9_debug(P9_DEBUG_ERROR,
223 					 "integer field, but no integer?\n");
224 				continue;
225 			}
226 		}
227 		switch (token) {
228 		case Opt_port:
229 			opts->port = option;
230 			break;
231 		case Opt_sq_depth:
232 			opts->sq_depth = option;
233 			break;
234 		case Opt_rq_depth:
235 			opts->rq_depth = option;
236 			break;
237 		case Opt_timeout:
238 			opts->timeout = option;
239 			break;
240 		case Opt_privport:
241 			opts->privport = true;
242 			break;
243 		default:
244 			continue;
245 		}
246 	}
247 	/* RQ must be at least as large as the SQ */
248 	opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
249 	kfree(tmp_options);
250 	return 0;
251 }
252 
253 static int
254 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
255 {
256 	struct p9_client *c = id->context;
257 	struct p9_trans_rdma *rdma = c->trans;
258 	switch (event->event) {
259 	case RDMA_CM_EVENT_ADDR_RESOLVED:
260 		BUG_ON(rdma->state != P9_RDMA_INIT);
261 		rdma->state = P9_RDMA_ADDR_RESOLVED;
262 		break;
263 
264 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
265 		BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
266 		rdma->state = P9_RDMA_ROUTE_RESOLVED;
267 		break;
268 
269 	case RDMA_CM_EVENT_ESTABLISHED:
270 		BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
271 		rdma->state = P9_RDMA_CONNECTED;
272 		break;
273 
274 	case RDMA_CM_EVENT_DISCONNECTED:
275 		if (rdma)
276 			rdma->state = P9_RDMA_CLOSED;
277 		if (c)
278 			c->status = Disconnected;
279 		break;
280 
281 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
282 		break;
283 
284 	case RDMA_CM_EVENT_ADDR_CHANGE:
285 	case RDMA_CM_EVENT_ROUTE_ERROR:
286 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
287 	case RDMA_CM_EVENT_MULTICAST_JOIN:
288 	case RDMA_CM_EVENT_MULTICAST_ERROR:
289 	case RDMA_CM_EVENT_REJECTED:
290 	case RDMA_CM_EVENT_CONNECT_REQUEST:
291 	case RDMA_CM_EVENT_CONNECT_RESPONSE:
292 	case RDMA_CM_EVENT_CONNECT_ERROR:
293 	case RDMA_CM_EVENT_ADDR_ERROR:
294 	case RDMA_CM_EVENT_UNREACHABLE:
295 		c->status = Disconnected;
296 		rdma_disconnect(rdma->cm_id);
297 		break;
298 	default:
299 		BUG();
300 	}
301 	complete(&rdma->cm_done);
302 	return 0;
303 }
304 
305 static void
306 recv_done(struct ib_cq *cq, struct ib_wc *wc)
307 {
308 	struct p9_client *client = cq->cq_context;
309 	struct p9_trans_rdma *rdma = client->trans;
310 	struct p9_rdma_context *c =
311 		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
312 	struct p9_req_t *req;
313 	int err = 0;
314 	int16_t tag;
315 
316 	req = NULL;
317 	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
318 							 DMA_FROM_DEVICE);
319 
320 	if (wc->status != IB_WC_SUCCESS)
321 		goto err_out;
322 
323 	err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
324 	if (err)
325 		goto err_out;
326 
327 	req = p9_tag_lookup(client, tag);
328 	if (!req)
329 		goto err_out;
330 
331 	/* Check that we have not yet received a reply for this request.
332 	 */
333 	if (unlikely(req->rc)) {
334 		pr_err("Duplicate reply for request %d", tag);
335 		goto err_out;
336 	}
337 
338 	req->rc = c->rc;
339 	p9_client_cb(client, req, REQ_STATUS_RCVD);
340 
341  out:
342 	up(&rdma->rq_sem);
343 	kfree(c);
344 	return;
345 
346  err_out:
347 	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
348 			req, err, wc->status);
349 	rdma->state = P9_RDMA_FLUSHING;
350 	client->status = Disconnected;
351 	goto out;
352 }
353 
354 static void
355 send_done(struct ib_cq *cq, struct ib_wc *wc)
356 {
357 	struct p9_client *client = cq->cq_context;
358 	struct p9_trans_rdma *rdma = client->trans;
359 	struct p9_rdma_context *c =
360 		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
361 
362 	ib_dma_unmap_single(rdma->cm_id->device,
363 			    c->busa, c->req->tc->size,
364 			    DMA_TO_DEVICE);
365 	up(&rdma->sq_sem);
366 	kfree(c);
367 }
368 
369 static void qp_event_handler(struct ib_event *event, void *context)
370 {
371 	p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
372 		 event->event, context);
373 }
374 
375 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
376 {
377 	if (!rdma)
378 		return;
379 
380 	if (rdma->qp && !IS_ERR(rdma->qp))
381 		ib_destroy_qp(rdma->qp);
382 
383 	if (rdma->pd && !IS_ERR(rdma->pd))
384 		ib_dealloc_pd(rdma->pd);
385 
386 	if (rdma->cq && !IS_ERR(rdma->cq))
387 		ib_free_cq(rdma->cq);
388 
389 	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
390 		rdma_destroy_id(rdma->cm_id);
391 
392 	kfree(rdma);
393 }
394 
395 static int
396 post_recv(struct p9_client *client, struct p9_rdma_context *c)
397 {
398 	struct p9_trans_rdma *rdma = client->trans;
399 	struct ib_recv_wr wr, *bad_wr;
400 	struct ib_sge sge;
401 
402 	c->busa = ib_dma_map_single(rdma->cm_id->device,
403 				    c->rc->sdata, client->msize,
404 				    DMA_FROM_DEVICE);
405 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
406 		goto error;
407 
408 	c->cqe.done = recv_done;
409 
410 	sge.addr = c->busa;
411 	sge.length = client->msize;
412 	sge.lkey = rdma->pd->local_dma_lkey;
413 
414 	wr.next = NULL;
415 	wr.wr_cqe = &c->cqe;
416 	wr.sg_list = &sge;
417 	wr.num_sge = 1;
418 	return ib_post_recv(rdma->qp, &wr, &bad_wr);
419 
420  error:
421 	p9_debug(P9_DEBUG_ERROR, "EIO\n");
422 	return -EIO;
423 }
424 
425 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
426 {
427 	struct p9_trans_rdma *rdma = client->trans;
428 	struct ib_send_wr wr, *bad_wr;
429 	struct ib_sge sge;
430 	int err = 0;
431 	unsigned long flags;
432 	struct p9_rdma_context *c = NULL;
433 	struct p9_rdma_context *rpl_context = NULL;
434 
435 	/* When an error occurs between posting the recv and the send,
436 	 * there will be a receive context posted without a pending request.
437 	 * Since there is no way to "un-post" it, we remember it and skip
438 	 * post_recv() for the next request.
439 	 * So here,
440 	 * see if we are this `next request' and need to absorb an excess rc.
441 	 * If yes, then drop and free our own, and do not recv_post().
442 	 **/
443 	if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
444 		if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
445 			/* Got one ! */
446 			kfree(req->rc);
447 			req->rc = NULL;
448 			goto dont_need_post_recv;
449 		} else {
450 			/* We raced and lost. */
451 			atomic_inc(&rdma->excess_rc);
452 		}
453 	}
454 
455 	/* Allocate an fcall for the reply */
456 	rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
457 	if (!rpl_context) {
458 		err = -ENOMEM;
459 		goto recv_error;
460 	}
461 	rpl_context->rc = req->rc;
462 
463 	/*
464 	 * Post a receive buffer for this request. We need to ensure
465 	 * there is a reply buffer available for every outstanding
466 	 * request. A flushed request can result in no reply for an
467 	 * outstanding request, so we must keep a count to avoid
468 	 * overflowing the RQ.
469 	 */
470 	if (down_interruptible(&rdma->rq_sem)) {
471 		err = -EINTR;
472 		goto recv_error;
473 	}
474 
475 	err = post_recv(client, rpl_context);
476 	if (err) {
477 		p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
478 		goto recv_error;
479 	}
480 	/* remove posted receive buffer from request structure */
481 	req->rc = NULL;
482 
483 dont_need_post_recv:
484 	/* Post the request */
485 	c = kmalloc(sizeof *c, GFP_NOFS);
486 	if (!c) {
487 		err = -ENOMEM;
488 		goto send_error;
489 	}
490 	c->req = req;
491 
492 	c->busa = ib_dma_map_single(rdma->cm_id->device,
493 				    c->req->tc->sdata, c->req->tc->size,
494 				    DMA_TO_DEVICE);
495 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
496 		err = -EIO;
497 		goto send_error;
498 	}
499 
500 	c->cqe.done = send_done;
501 
502 	sge.addr = c->busa;
503 	sge.length = c->req->tc->size;
504 	sge.lkey = rdma->pd->local_dma_lkey;
505 
506 	wr.next = NULL;
507 	wr.wr_cqe = &c->cqe;
508 	wr.opcode = IB_WR_SEND;
509 	wr.send_flags = IB_SEND_SIGNALED;
510 	wr.sg_list = &sge;
511 	wr.num_sge = 1;
512 
513 	if (down_interruptible(&rdma->sq_sem)) {
514 		err = -EINTR;
515 		goto send_error;
516 	}
517 
518 	/* Mark request as `sent' *before* we actually send it,
519 	 * because doing if after could erase the REQ_STATUS_RCVD
520 	 * status in case of a very fast reply.
521 	 */
522 	req->status = REQ_STATUS_SENT;
523 	err = ib_post_send(rdma->qp, &wr, &bad_wr);
524 	if (err)
525 		goto send_error;
526 
527 	/* Success */
528 	return 0;
529 
530  /* Handle errors that happened during or while preparing the send: */
531  send_error:
532 	req->status = REQ_STATUS_ERROR;
533 	kfree(c);
534 	p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
535 
536 	/* Ach.
537 	 *  We did recv_post(), but not send. We have one recv_post in excess.
538 	 */
539 	atomic_inc(&rdma->excess_rc);
540 	return err;
541 
542  /* Handle errors that happened during or while preparing post_recv(): */
543  recv_error:
544 	kfree(rpl_context);
545 	spin_lock_irqsave(&rdma->req_lock, flags);
546 	if (rdma->state < P9_RDMA_CLOSING) {
547 		rdma->state = P9_RDMA_CLOSING;
548 		spin_unlock_irqrestore(&rdma->req_lock, flags);
549 		rdma_disconnect(rdma->cm_id);
550 	} else
551 		spin_unlock_irqrestore(&rdma->req_lock, flags);
552 	return err;
553 }
554 
555 static void rdma_close(struct p9_client *client)
556 {
557 	struct p9_trans_rdma *rdma;
558 
559 	if (!client)
560 		return;
561 
562 	rdma = client->trans;
563 	if (!rdma)
564 		return;
565 
566 	client->status = Disconnected;
567 	rdma_disconnect(rdma->cm_id);
568 	rdma_destroy_trans(rdma);
569 }
570 
571 /**
572  * alloc_rdma - Allocate and initialize the rdma transport structure
573  * @opts: Mount options structure
574  */
575 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
576 {
577 	struct p9_trans_rdma *rdma;
578 
579 	rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
580 	if (!rdma)
581 		return NULL;
582 
583 	rdma->port = opts->port;
584 	rdma->privport = opts->privport;
585 	rdma->sq_depth = opts->sq_depth;
586 	rdma->rq_depth = opts->rq_depth;
587 	rdma->timeout = opts->timeout;
588 	spin_lock_init(&rdma->req_lock);
589 	init_completion(&rdma->cm_done);
590 	sema_init(&rdma->sq_sem, rdma->sq_depth);
591 	sema_init(&rdma->rq_sem, rdma->rq_depth);
592 	atomic_set(&rdma->excess_rc, 0);
593 
594 	return rdma;
595 }
596 
597 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
598 {
599 	/* Nothing to do here.
600 	 * We will take care of it (if we have to) in rdma_cancelled()
601 	 */
602 	return 1;
603 }
604 
605 /* A request has been fully flushed without a reply.
606  * That means we have posted one buffer in excess.
607  */
608 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
609 {
610 	struct p9_trans_rdma *rdma = client->trans;
611 	atomic_inc(&rdma->excess_rc);
612 	return 0;
613 }
614 
615 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
616 {
617 	struct sockaddr_in cl = {
618 		.sin_family = AF_INET,
619 		.sin_addr.s_addr = htonl(INADDR_ANY),
620 	};
621 	int port, err = -EINVAL;
622 
623 	for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
624 		cl.sin_port = htons((ushort)port);
625 		err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
626 		if (err != -EADDRINUSE)
627 			break;
628 	}
629 	return err;
630 }
631 
632 /**
633  * rdma_create_trans - Transport method for creating a transport instance
634  * @client: client instance
635  * @addr: IP address string
636  * @args: Mount options string
637  */
638 static int
639 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
640 {
641 	int err;
642 	struct p9_rdma_opts opts;
643 	struct p9_trans_rdma *rdma;
644 	struct rdma_conn_param conn_param;
645 	struct ib_qp_init_attr qp_attr;
646 
647 	/* Parse the transport specific mount options */
648 	err = parse_opts(args, &opts);
649 	if (err < 0)
650 		return err;
651 
652 	/* Create and initialize the RDMA transport structure */
653 	rdma = alloc_rdma(&opts);
654 	if (!rdma)
655 		return -ENOMEM;
656 
657 	/* Create the RDMA CM ID */
658 	rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
659 				     RDMA_PS_TCP, IB_QPT_RC);
660 	if (IS_ERR(rdma->cm_id))
661 		goto error;
662 
663 	/* Associate the client with the transport */
664 	client->trans = rdma;
665 
666 	/* Bind to a privileged port if we need to */
667 	if (opts.privport) {
668 		err = p9_rdma_bind_privport(rdma);
669 		if (err < 0) {
670 			pr_err("%s (%d): problem binding to privport: %d\n",
671 			       __func__, task_pid_nr(current), -err);
672 			goto error;
673 		}
674 	}
675 
676 	/* Resolve the server's address */
677 	rdma->addr.sin_family = AF_INET;
678 	rdma->addr.sin_addr.s_addr = in_aton(addr);
679 	rdma->addr.sin_port = htons(opts.port);
680 	err = rdma_resolve_addr(rdma->cm_id, NULL,
681 				(struct sockaddr *)&rdma->addr,
682 				rdma->timeout);
683 	if (err)
684 		goto error;
685 	err = wait_for_completion_interruptible(&rdma->cm_done);
686 	if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
687 		goto error;
688 
689 	/* Resolve the route to the server */
690 	err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
691 	if (err)
692 		goto error;
693 	err = wait_for_completion_interruptible(&rdma->cm_done);
694 	if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
695 		goto error;
696 
697 	/* Create the Completion Queue */
698 	rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
699 			opts.sq_depth + opts.rq_depth + 1,
700 			0, IB_POLL_SOFTIRQ);
701 	if (IS_ERR(rdma->cq))
702 		goto error;
703 
704 	/* Create the Protection Domain */
705 	rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
706 	if (IS_ERR(rdma->pd))
707 		goto error;
708 
709 	/* Create the Queue Pair */
710 	memset(&qp_attr, 0, sizeof qp_attr);
711 	qp_attr.event_handler = qp_event_handler;
712 	qp_attr.qp_context = client;
713 	qp_attr.cap.max_send_wr = opts.sq_depth;
714 	qp_attr.cap.max_recv_wr = opts.rq_depth;
715 	qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
716 	qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
717 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
718 	qp_attr.qp_type = IB_QPT_RC;
719 	qp_attr.send_cq = rdma->cq;
720 	qp_attr.recv_cq = rdma->cq;
721 	err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
722 	if (err)
723 		goto error;
724 	rdma->qp = rdma->cm_id->qp;
725 
726 	/* Request a connection */
727 	memset(&conn_param, 0, sizeof(conn_param));
728 	conn_param.private_data = NULL;
729 	conn_param.private_data_len = 0;
730 	conn_param.responder_resources = P9_RDMA_IRD;
731 	conn_param.initiator_depth = P9_RDMA_ORD;
732 	err = rdma_connect(rdma->cm_id, &conn_param);
733 	if (err)
734 		goto error;
735 	err = wait_for_completion_interruptible(&rdma->cm_done);
736 	if (err || (rdma->state != P9_RDMA_CONNECTED))
737 		goto error;
738 
739 	client->status = Connected;
740 
741 	return 0;
742 
743 error:
744 	rdma_destroy_trans(rdma);
745 	return -ENOTCONN;
746 }
747 
748 static struct p9_trans_module p9_rdma_trans = {
749 	.name = "rdma",
750 	.maxsize = P9_RDMA_MAXSIZE,
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 
778 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
779 MODULE_DESCRIPTION("RDMA Transport for 9P");
780 MODULE_LICENSE("Dual BSD/GPL");
781