xref: /openbmc/linux/net/9p/trans_rdma.c (revision 160b8e75) 
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  * @dm_mr: DMA Memory Region pointer
72  * @lkey: The local access only memory region key
73  * @timeout: Number of uSecs to wait for connection management events
74  * @privport: Whether a privileged port may be used
75  * @port: The port to use
76  * @sq_depth: The depth of the Send Queue
77  * @sq_sem: Semaphore for the SQ
78  * @rq_depth: The depth of the Receive Queue.
79  * @rq_sem: Semaphore for the RQ
80  * @excess_rc : Amount of posted Receive Contexts without a pending request.
81  *		See rdma_request()
82  * @addr: The remote peer's address
83  * @req_lock: Protects the active request list
84  * @cm_done: Completion event for connection management tracking
85  */
86 struct p9_trans_rdma {
87 	enum {
88 		P9_RDMA_INIT,
89 		P9_RDMA_ADDR_RESOLVED,
90 		P9_RDMA_ROUTE_RESOLVED,
91 		P9_RDMA_CONNECTED,
92 		P9_RDMA_FLUSHING,
93 		P9_RDMA_CLOSING,
94 		P9_RDMA_CLOSED,
95 	} state;
96 	struct rdma_cm_id *cm_id;
97 	struct ib_pd *pd;
98 	struct ib_qp *qp;
99 	struct ib_cq *cq;
100 	long timeout;
101 	bool privport;
102 	u16 port;
103 	int sq_depth;
104 	struct semaphore sq_sem;
105 	int rq_depth;
106 	struct semaphore rq_sem;
107 	atomic_t excess_rc;
108 	struct sockaddr_in addr;
109 	spinlock_t req_lock;
110 
111 	struct completion cm_done;
112 };
113 
114 /**
115  * p9_rdma_context - Keeps track of in-process WR
116  *
117  * @busa: Bus address to unmap when the WR completes
118  * @req: Keeps track of requests (send)
119  * @rc: Keepts track of replies (receive)
120  */
121 struct p9_rdma_req;
122 struct p9_rdma_context {
123 	struct ib_cqe cqe;
124 	dma_addr_t busa;
125 	union {
126 		struct p9_req_t *req;
127 		struct p9_fcall *rc;
128 	};
129 };
130 
131 /**
132  * p9_rdma_opts - Collection of mount options
133  * @port: port of connection
134  * @sq_depth: The requested depth of the SQ. This really doesn't need
135  * to be any deeper than the number of threads used in the client
136  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
137  * @timeout: Time to wait in msecs for CM events
138  */
139 struct p9_rdma_opts {
140 	short port;
141 	bool privport;
142 	int sq_depth;
143 	int rq_depth;
144 	long timeout;
145 };
146 
147 /*
148  * Option Parsing (code inspired by NFS code)
149  */
150 enum {
151 	/* Options that take integer arguments */
152 	Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
153 	/* Options that take no argument */
154 	Opt_privport,
155 	Opt_err,
156 };
157 
158 static match_table_t tokens = {
159 	{Opt_port, "port=%u"},
160 	{Opt_sq_depth, "sq=%u"},
161 	{Opt_rq_depth, "rq=%u"},
162 	{Opt_timeout, "timeout=%u"},
163 	{Opt_privport, "privport"},
164 	{Opt_err, NULL},
165 };
166 
167 static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
168 {
169 	struct p9_trans_rdma *rdma = clnt->trans;
170 
171 	if (rdma->port != P9_PORT)
172 		seq_printf(m, ",port=%u", rdma->port);
173 	if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
174 		seq_printf(m, ",sq=%u", rdma->sq_depth);
175 	if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
176 		seq_printf(m, ",rq=%u", rdma->rq_depth);
177 	if (rdma->timeout != P9_RDMA_TIMEOUT)
178 		seq_printf(m, ",timeout=%lu", rdma->timeout);
179 	if (rdma->privport)
180 		seq_puts(m, ",privport");
181 	return 0;
182 }
183 
184 /**
185  * parse_opts - parse mount options into rdma options structure
186  * @params: options string passed from mount
187  * @opts: rdma transport-specific structure to parse options into
188  *
189  * Returns 0 upon success, -ERRNO upon failure
190  */
191 static int parse_opts(char *params, struct p9_rdma_opts *opts)
192 {
193 	char *p;
194 	substring_t args[MAX_OPT_ARGS];
195 	int option;
196 	char *options, *tmp_options;
197 
198 	opts->port = P9_PORT;
199 	opts->sq_depth = P9_RDMA_SQ_DEPTH;
200 	opts->rq_depth = P9_RDMA_RQ_DEPTH;
201 	opts->timeout = P9_RDMA_TIMEOUT;
202 	opts->privport = false;
203 
204 	if (!params)
205 		return 0;
206 
207 	tmp_options = kstrdup(params, GFP_KERNEL);
208 	if (!tmp_options) {
209 		p9_debug(P9_DEBUG_ERROR,
210 			 "failed to allocate copy of option string\n");
211 		return -ENOMEM;
212 	}
213 	options = tmp_options;
214 
215 	while ((p = strsep(&options, ",")) != NULL) {
216 		int token;
217 		int r;
218 		if (!*p)
219 			continue;
220 		token = match_token(p, tokens, args);
221 		if ((token != Opt_err) && (token != Opt_privport)) {
222 			r = match_int(&args[0], &option);
223 			if (r < 0) {
224 				p9_debug(P9_DEBUG_ERROR,
225 					 "integer field, but no integer?\n");
226 				continue;
227 			}
228 		}
229 		switch (token) {
230 		case Opt_port:
231 			opts->port = option;
232 			break;
233 		case Opt_sq_depth:
234 			opts->sq_depth = option;
235 			break;
236 		case Opt_rq_depth:
237 			opts->rq_depth = option;
238 			break;
239 		case Opt_timeout:
240 			opts->timeout = option;
241 			break;
242 		case Opt_privport:
243 			opts->privport = true;
244 			break;
245 		default:
246 			continue;
247 		}
248 	}
249 	/* RQ must be at least as large as the SQ */
250 	opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
251 	kfree(tmp_options);
252 	return 0;
253 }
254 
255 static int
256 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
257 {
258 	struct p9_client *c = id->context;
259 	struct p9_trans_rdma *rdma = c->trans;
260 	switch (event->event) {
261 	case RDMA_CM_EVENT_ADDR_RESOLVED:
262 		BUG_ON(rdma->state != P9_RDMA_INIT);
263 		rdma->state = P9_RDMA_ADDR_RESOLVED;
264 		break;
265 
266 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
267 		BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
268 		rdma->state = P9_RDMA_ROUTE_RESOLVED;
269 		break;
270 
271 	case RDMA_CM_EVENT_ESTABLISHED:
272 		BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
273 		rdma->state = P9_RDMA_CONNECTED;
274 		break;
275 
276 	case RDMA_CM_EVENT_DISCONNECTED:
277 		if (rdma)
278 			rdma->state = P9_RDMA_CLOSED;
279 		if (c)
280 			c->status = Disconnected;
281 		break;
282 
283 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
284 		break;
285 
286 	case RDMA_CM_EVENT_ADDR_CHANGE:
287 	case RDMA_CM_EVENT_ROUTE_ERROR:
288 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
289 	case RDMA_CM_EVENT_MULTICAST_JOIN:
290 	case RDMA_CM_EVENT_MULTICAST_ERROR:
291 	case RDMA_CM_EVENT_REJECTED:
292 	case RDMA_CM_EVENT_CONNECT_REQUEST:
293 	case RDMA_CM_EVENT_CONNECT_RESPONSE:
294 	case RDMA_CM_EVENT_CONNECT_ERROR:
295 	case RDMA_CM_EVENT_ADDR_ERROR:
296 	case RDMA_CM_EVENT_UNREACHABLE:
297 		c->status = Disconnected;
298 		rdma_disconnect(rdma->cm_id);
299 		break;
300 	default:
301 		BUG();
302 	}
303 	complete(&rdma->cm_done);
304 	return 0;
305 }
306 
307 static void
308 recv_done(struct ib_cq *cq, struct ib_wc *wc)
309 {
310 	struct p9_client *client = cq->cq_context;
311 	struct p9_trans_rdma *rdma = client->trans;
312 	struct p9_rdma_context *c =
313 		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
314 	struct p9_req_t *req;
315 	int err = 0;
316 	int16_t tag;
317 
318 	req = NULL;
319 	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
320 							 DMA_FROM_DEVICE);
321 
322 	if (wc->status != IB_WC_SUCCESS)
323 		goto err_out;
324 
325 	err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
326 	if (err)
327 		goto err_out;
328 
329 	req = p9_tag_lookup(client, tag);
330 	if (!req)
331 		goto err_out;
332 
333 	/* Check that we have not yet received a reply for this request.
334 	 */
335 	if (unlikely(req->rc)) {
336 		pr_err("Duplicate reply for request %d", tag);
337 		goto err_out;
338 	}
339 
340 	req->rc = c->rc;
341 	p9_client_cb(client, req, REQ_STATUS_RCVD);
342 
343  out:
344 	up(&rdma->rq_sem);
345 	kfree(c);
346 	return;
347 
348  err_out:
349 	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
350 			req, err, wc->status);
351 	rdma->state = P9_RDMA_FLUSHING;
352 	client->status = Disconnected;
353 	goto out;
354 }
355 
356 static void
357 send_done(struct ib_cq *cq, struct ib_wc *wc)
358 {
359 	struct p9_client *client = cq->cq_context;
360 	struct p9_trans_rdma *rdma = client->trans;
361 	struct p9_rdma_context *c =
362 		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
363 
364 	ib_dma_unmap_single(rdma->cm_id->device,
365 			    c->busa, c->req->tc->size,
366 			    DMA_TO_DEVICE);
367 	up(&rdma->sq_sem);
368 	kfree(c);
369 }
370 
371 static void qp_event_handler(struct ib_event *event, void *context)
372 {
373 	p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
374 		 event->event, context);
375 }
376 
377 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
378 {
379 	if (!rdma)
380 		return;
381 
382 	if (rdma->qp && !IS_ERR(rdma->qp))
383 		ib_destroy_qp(rdma->qp);
384 
385 	if (rdma->pd && !IS_ERR(rdma->pd))
386 		ib_dealloc_pd(rdma->pd);
387 
388 	if (rdma->cq && !IS_ERR(rdma->cq))
389 		ib_free_cq(rdma->cq);
390 
391 	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
392 		rdma_destroy_id(rdma->cm_id);
393 
394 	kfree(rdma);
395 }
396 
397 static int
398 post_recv(struct p9_client *client, struct p9_rdma_context *c)
399 {
400 	struct p9_trans_rdma *rdma = client->trans;
401 	struct ib_recv_wr wr, *bad_wr;
402 	struct ib_sge sge;
403 
404 	c->busa = ib_dma_map_single(rdma->cm_id->device,
405 				    c->rc->sdata, client->msize,
406 				    DMA_FROM_DEVICE);
407 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
408 		goto error;
409 
410 	c->cqe.done = recv_done;
411 
412 	sge.addr = c->busa;
413 	sge.length = client->msize;
414 	sge.lkey = rdma->pd->local_dma_lkey;
415 
416 	wr.next = NULL;
417 	wr.wr_cqe = &c->cqe;
418 	wr.sg_list = &sge;
419 	wr.num_sge = 1;
420 	return ib_post_recv(rdma->qp, &wr, &bad_wr);
421 
422  error:
423 	p9_debug(P9_DEBUG_ERROR, "EIO\n");
424 	return -EIO;
425 }
426 
427 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
428 {
429 	struct p9_trans_rdma *rdma = client->trans;
430 	struct ib_send_wr wr, *bad_wr;
431 	struct ib_sge sge;
432 	int err = 0;
433 	unsigned long flags;
434 	struct p9_rdma_context *c = NULL;
435 	struct p9_rdma_context *rpl_context = NULL;
436 
437 	/* When an error occurs between posting the recv and the send,
438 	 * there will be a receive context posted without a pending request.
439 	 * Since there is no way to "un-post" it, we remember it and skip
440 	 * post_recv() for the next request.
441 	 * So here,
442 	 * see if we are this `next request' and need to absorb an excess rc.
443 	 * If yes, then drop and free our own, and do not recv_post().
444 	 **/
445 	if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
446 		if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
447 			/* Got one ! */
448 			kfree(req->rc);
449 			req->rc = NULL;
450 			goto dont_need_post_recv;
451 		} else {
452 			/* We raced and lost. */
453 			atomic_inc(&rdma->excess_rc);
454 		}
455 	}
456 
457 	/* Allocate an fcall for the reply */
458 	rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
459 	if (!rpl_context) {
460 		err = -ENOMEM;
461 		goto recv_error;
462 	}
463 	rpl_context->rc = req->rc;
464 
465 	/*
466 	 * Post a receive buffer for this request. We need to ensure
467 	 * there is a reply buffer available for every outstanding
468 	 * request. A flushed request can result in no reply for an
469 	 * outstanding request, so we must keep a count to avoid
470 	 * overflowing the RQ.
471 	 */
472 	if (down_interruptible(&rdma->rq_sem)) {
473 		err = -EINTR;
474 		goto recv_error;
475 	}
476 
477 	err = post_recv(client, rpl_context);
478 	if (err) {
479 		p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
480 		goto recv_error;
481 	}
482 	/* remove posted receive buffer from request structure */
483 	req->rc = NULL;
484 
485 dont_need_post_recv:
486 	/* Post the request */
487 	c = kmalloc(sizeof *c, GFP_NOFS);
488 	if (!c) {
489 		err = -ENOMEM;
490 		goto send_error;
491 	}
492 	c->req = req;
493 
494 	c->busa = ib_dma_map_single(rdma->cm_id->device,
495 				    c->req->tc->sdata, c->req->tc->size,
496 				    DMA_TO_DEVICE);
497 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
498 		err = -EIO;
499 		goto send_error;
500 	}
501 
502 	c->cqe.done = send_done;
503 
504 	sge.addr = c->busa;
505 	sge.length = c->req->tc->size;
506 	sge.lkey = rdma->pd->local_dma_lkey;
507 
508 	wr.next = NULL;
509 	wr.wr_cqe = &c->cqe;
510 	wr.opcode = IB_WR_SEND;
511 	wr.send_flags = IB_SEND_SIGNALED;
512 	wr.sg_list = &sge;
513 	wr.num_sge = 1;
514 
515 	if (down_interruptible(&rdma->sq_sem)) {
516 		err = -EINTR;
517 		goto send_error;
518 	}
519 
520 	/* Mark request as `sent' *before* we actually send it,
521 	 * because doing if after could erase the REQ_STATUS_RCVD
522 	 * status in case of a very fast reply.
523 	 */
524 	req->status = REQ_STATUS_SENT;
525 	err = ib_post_send(rdma->qp, &wr, &bad_wr);
526 	if (err)
527 		goto send_error;
528 
529 	/* Success */
530 	return 0;
531 
532  /* Handle errors that happened during or while preparing the send: */
533  send_error:
534 	req->status = REQ_STATUS_ERROR;
535 	kfree(c);
536 	p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
537 
538 	/* Ach.
539 	 *  We did recv_post(), but not send. We have one recv_post in excess.
540 	 */
541 	atomic_inc(&rdma->excess_rc);
542 	return err;
543 
544  /* Handle errors that happened during or while preparing post_recv(): */
545  recv_error:
546 	kfree(rpl_context);
547 	spin_lock_irqsave(&rdma->req_lock, flags);
548 	if (rdma->state < P9_RDMA_CLOSING) {
549 		rdma->state = P9_RDMA_CLOSING;
550 		spin_unlock_irqrestore(&rdma->req_lock, flags);
551 		rdma_disconnect(rdma->cm_id);
552 	} else
553 		spin_unlock_irqrestore(&rdma->req_lock, flags);
554 	return err;
555 }
556 
557 static void rdma_close(struct p9_client *client)
558 {
559 	struct p9_trans_rdma *rdma;
560 
561 	if (!client)
562 		return;
563 
564 	rdma = client->trans;
565 	if (!rdma)
566 		return;
567 
568 	client->status = Disconnected;
569 	rdma_disconnect(rdma->cm_id);
570 	rdma_destroy_trans(rdma);
571 }
572 
573 /**
574  * alloc_rdma - Allocate and initialize the rdma transport structure
575  * @opts: Mount options structure
576  */
577 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
578 {
579 	struct p9_trans_rdma *rdma;
580 
581 	rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
582 	if (!rdma)
583 		return NULL;
584 
585 	rdma->port = opts->port;
586 	rdma->privport = opts->privport;
587 	rdma->sq_depth = opts->sq_depth;
588 	rdma->rq_depth = opts->rq_depth;
589 	rdma->timeout = opts->timeout;
590 	spin_lock_init(&rdma->req_lock);
591 	init_completion(&rdma->cm_done);
592 	sema_init(&rdma->sq_sem, rdma->sq_depth);
593 	sema_init(&rdma->rq_sem, rdma->rq_depth);
594 	atomic_set(&rdma->excess_rc, 0);
595 
596 	return rdma;
597 }
598 
599 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
600 {
601 	/* Nothing to do here.
602 	 * We will take care of it (if we have to) in rdma_cancelled()
603 	 */
604 	return 1;
605 }
606 
607 /* A request has been fully flushed without a reply.
608  * That means we have posted one buffer in excess.
609  */
610 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
611 {
612 	struct p9_trans_rdma *rdma = client->trans;
613 	atomic_inc(&rdma->excess_rc);
614 	return 0;
615 }
616 
617 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
618 {
619 	struct sockaddr_in cl = {
620 		.sin_family = AF_INET,
621 		.sin_addr.s_addr = htonl(INADDR_ANY),
622 	};
623 	int port, err = -EINVAL;
624 
625 	for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
626 		cl.sin_port = htons((ushort)port);
627 		err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
628 		if (err != -EADDRINUSE)
629 			break;
630 	}
631 	return err;
632 }
633 
634 /**
635  * trans_create_rdma - Transport method for creating atransport instance
636  * @client: client instance
637  * @addr: IP address string
638  * @args: Mount options string
639  */
640 static int
641 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
642 {
643 	int err;
644 	struct p9_rdma_opts opts;
645 	struct p9_trans_rdma *rdma;
646 	struct rdma_conn_param conn_param;
647 	struct ib_qp_init_attr qp_attr;
648 
649 	/* Parse the transport specific mount options */
650 	err = parse_opts(args, &opts);
651 	if (err < 0)
652 		return err;
653 
654 	/* Create and initialize the RDMA transport structure */
655 	rdma = alloc_rdma(&opts);
656 	if (!rdma)
657 		return -ENOMEM;
658 
659 	/* Create the RDMA CM ID */
660 	rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
661 				     RDMA_PS_TCP, IB_QPT_RC);
662 	if (IS_ERR(rdma->cm_id))
663 		goto error;
664 
665 	/* Associate the client with the transport */
666 	client->trans = rdma;
667 
668 	/* Bind to a privileged port if we need to */
669 	if (opts.privport) {
670 		err = p9_rdma_bind_privport(rdma);
671 		if (err < 0) {
672 			pr_err("%s (%d): problem binding to privport: %d\n",
673 			       __func__, task_pid_nr(current), -err);
674 			goto error;
675 		}
676 	}
677 
678 	/* Resolve the server's address */
679 	rdma->addr.sin_family = AF_INET;
680 	rdma->addr.sin_addr.s_addr = in_aton(addr);
681 	rdma->addr.sin_port = htons(opts.port);
682 	err = rdma_resolve_addr(rdma->cm_id, NULL,
683 				(struct sockaddr *)&rdma->addr,
684 				rdma->timeout);
685 	if (err)
686 		goto error;
687 	err = wait_for_completion_interruptible(&rdma->cm_done);
688 	if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
689 		goto error;
690 
691 	/* Resolve the route to the server */
692 	err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
693 	if (err)
694 		goto error;
695 	err = wait_for_completion_interruptible(&rdma->cm_done);
696 	if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
697 		goto error;
698 
699 	/* Create the Completion Queue */
700 	rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
701 			opts.sq_depth + opts.rq_depth + 1,
702 			0, IB_POLL_SOFTIRQ);
703 	if (IS_ERR(rdma->cq))
704 		goto error;
705 
706 	/* Create the Protection Domain */
707 	rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
708 	if (IS_ERR(rdma->pd))
709 		goto error;
710 
711 	/* Create the Queue Pair */
712 	memset(&qp_attr, 0, sizeof qp_attr);
713 	qp_attr.event_handler = qp_event_handler;
714 	qp_attr.qp_context = client;
715 	qp_attr.cap.max_send_wr = opts.sq_depth;
716 	qp_attr.cap.max_recv_wr = opts.rq_depth;
717 	qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
718 	qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
719 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
720 	qp_attr.qp_type = IB_QPT_RC;
721 	qp_attr.send_cq = rdma->cq;
722 	qp_attr.recv_cq = rdma->cq;
723 	err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
724 	if (err)
725 		goto error;
726 	rdma->qp = rdma->cm_id->qp;
727 
728 	/* Request a connection */
729 	memset(&conn_param, 0, sizeof(conn_param));
730 	conn_param.private_data = NULL;
731 	conn_param.private_data_len = 0;
732 	conn_param.responder_resources = P9_RDMA_IRD;
733 	conn_param.initiator_depth = P9_RDMA_ORD;
734 	err = rdma_connect(rdma->cm_id, &conn_param);
735 	if (err)
736 		goto error;
737 	err = wait_for_completion_interruptible(&rdma->cm_done);
738 	if (err || (rdma->state != P9_RDMA_CONNECTED))
739 		goto error;
740 
741 	client->status = Connected;
742 
743 	return 0;
744 
745 error:
746 	rdma_destroy_trans(rdma);
747 	return -ENOTCONN;
748 }
749 
750 static struct p9_trans_module p9_rdma_trans = {
751 	.name = "rdma",
752 	.maxsize = P9_RDMA_MAXSIZE,
753 	.def = 0,
754 	.owner = THIS_MODULE,
755 	.create = rdma_create_trans,
756 	.close = rdma_close,
757 	.request = rdma_request,
758 	.cancel = rdma_cancel,
759 	.cancelled = rdma_cancelled,
760 	.show_options = p9_rdma_show_options,
761 };
762 
763 /**
764  * p9_trans_rdma_init - Register the 9P RDMA transport driver
765  */
766 static int __init p9_trans_rdma_init(void)
767 {
768 	v9fs_register_trans(&p9_rdma_trans);
769 	return 0;
770 }
771 
772 static void __exit p9_trans_rdma_exit(void)
773 {
774 	v9fs_unregister_trans(&p9_rdma_trans);
775 }
776 
777 module_init(p9_trans_rdma_init);
778 module_exit(p9_trans_rdma_exit);
779 
780 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
781 MODULE_DESCRIPTION("RDMA Transport for 9P");
782 MODULE_LICENSE("Dual BSD/GPL");
783