xref: /linux/net/9p/trans_rdma.c (revision 7895d662bab8827176d44326d0a7423221287ca9) 
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * linux/fs/9p/trans_rdma.c
4  *
5  * RDMA transport layer based on the trans_fd.c implementation.
6  *
7  *  Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
8  *  Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
9  *  Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
10  *  Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
11  *  Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
12  */
13 
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 
16 #include <linux/in.h>
17 #include <linux/module.h>
18 #include <linux/net.h>
19 #include <linux/ipv6.h>
20 #include <linux/kthread.h>
21 #include <linux/errno.h>
22 #include <linux/kernel.h>
23 #include <linux/un.h>
24 #include <linux/uaccess.h>
25 #include <linux/inet.h>
26 #include <linux/idr.h>
27 #include <linux/file.h>
28 #include <linux/parser.h>
29 #include <linux/semaphore.h>
30 #include <linux/slab.h>
31 #include <linux/seq_file.h>
32 #include <net/9p/9p.h>
33 #include <net/9p/client.h>
34 #include <net/9p/transport.h>
35 #include <rdma/ib_verbs.h>
36 #include <rdma/rdma_cm.h>
37 
38 #define P9_PORT			5640
39 #define P9_RDMA_SQ_DEPTH	32
40 #define P9_RDMA_RQ_DEPTH	32
41 #define P9_RDMA_SEND_SGE	4
42 #define P9_RDMA_RECV_SGE	4
43 #define P9_RDMA_IRD		0
44 #define P9_RDMA_ORD		0
45 #define P9_RDMA_TIMEOUT		30000		/* 30 seconds */
46 #define P9_RDMA_MAXSIZE		(1024*1024)	/* 1MB */
47 
48 /**
49  * struct p9_trans_rdma - RDMA transport instance
50  *
51  * @state: tracks the transport state machine for connection setup and tear down
52  * @cm_id: The RDMA CM ID
53  * @pd: Protection Domain pointer
54  * @qp: Queue Pair pointer
55  * @cq: Completion Queue pointer
56  * @timeout: Number of uSecs to wait for connection management events
57  * @privport: Whether a privileged port may be used
58  * @port: The port to use
59  * @sq_depth: The depth of the Send Queue
60  * @sq_sem: Semaphore for the SQ
61  * @rq_depth: The depth of the Receive Queue.
62  * @rq_sem: Semaphore for the RQ
63  * @excess_rc : Amount of posted Receive Contexts without a pending request.
64  *		See rdma_request()
65  * @addr: The remote peer's address
66  * @req_lock: Protects the active request list
67  * @cm_done: Completion event for connection management tracking
68  */
69 struct p9_trans_rdma {
70 	enum {
71 		P9_RDMA_INIT,
72 		P9_RDMA_ADDR_RESOLVED,
73 		P9_RDMA_ROUTE_RESOLVED,
74 		P9_RDMA_CONNECTED,
75 		P9_RDMA_FLUSHING,
76 		P9_RDMA_CLOSING,
77 		P9_RDMA_CLOSED,
78 	} state;
79 	struct rdma_cm_id *cm_id;
80 	struct ib_pd *pd;
81 	struct ib_qp *qp;
82 	struct ib_cq *cq;
83 	long timeout;
84 	bool privport;
85 	u16 port;
86 	int sq_depth;
87 	struct semaphore sq_sem;
88 	int rq_depth;
89 	struct semaphore rq_sem;
90 	atomic_t excess_rc;
91 	struct sockaddr_in addr;
92 	spinlock_t req_lock;
93 
94 	struct completion cm_done;
95 };
96 
97 struct p9_rdma_req;
98 
99 /**
100  * struct p9_rdma_context - Keeps track of in-process WR
101  *
102  * @busa: Bus address to unmap when the WR completes
103  * @req: Keeps track of requests (send)
104  * @rc: Keepts track of replies (receive)
105  */
106 struct p9_rdma_context {
107 	struct ib_cqe cqe;
108 	dma_addr_t busa;
109 	union {
110 		struct p9_req_t *req;
111 		struct p9_fcall rc;
112 	};
113 };
114 
115 /**
116  * struct p9_rdma_opts - Collection of mount options
117  * @port: port of connection
118  * @sq_depth: The requested depth of the SQ. This really doesn't need
119  * to be any deeper than the number of threads used in the client
120  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
121  * @timeout: Time to wait in msecs for CM events
122  */
123 struct p9_rdma_opts {
124 	short port;
125 	bool privport;
126 	int sq_depth;
127 	int rq_depth;
128 	long timeout;
129 };
130 
131 /*
132  * Option Parsing (code inspired by NFS code)
133  */
134 enum {
135 	/* Options that take integer arguments */
136 	Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
137 	/* Options that take no argument */
138 	Opt_privport,
139 	Opt_err,
140 };
141 
142 static match_table_t tokens = {
143 	{Opt_port, "port=%u"},
144 	{Opt_sq_depth, "sq=%u"},
145 	{Opt_rq_depth, "rq=%u"},
146 	{Opt_timeout, "timeout=%u"},
147 	{Opt_privport, "privport"},
148 	{Opt_err, NULL},
149 };
150 
151 static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
152 {
153 	struct p9_trans_rdma *rdma = clnt->trans;
154 
155 	if (rdma->port != P9_PORT)
156 		seq_printf(m, ",port=%u", rdma->port);
157 	if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
158 		seq_printf(m, ",sq=%u", rdma->sq_depth);
159 	if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
160 		seq_printf(m, ",rq=%u", rdma->rq_depth);
161 	if (rdma->timeout != P9_RDMA_TIMEOUT)
162 		seq_printf(m, ",timeout=%lu", rdma->timeout);
163 	if (rdma->privport)
164 		seq_puts(m, ",privport");
165 	return 0;
166 }
167 
168 /**
169  * parse_opts - parse mount options into rdma options structure
170  * @params: options string passed from mount
171  * @opts: rdma transport-specific structure to parse options into
172  *
173  * Returns 0 upon success, -ERRNO upon failure
174  */
175 static int parse_opts(char *params, struct p9_rdma_opts *opts)
176 {
177 	char *p;
178 	substring_t args[MAX_OPT_ARGS];
179 	int option;
180 	char *options, *tmp_options;
181 
182 	opts->port = P9_PORT;
183 	opts->sq_depth = P9_RDMA_SQ_DEPTH;
184 	opts->rq_depth = P9_RDMA_RQ_DEPTH;
185 	opts->timeout = P9_RDMA_TIMEOUT;
186 	opts->privport = false;
187 
188 	if (!params)
189 		return 0;
190 
191 	tmp_options = kstrdup(params, GFP_KERNEL);
192 	if (!tmp_options) {
193 		p9_debug(P9_DEBUG_ERROR,
194 			 "failed to allocate copy of option string\n");
195 		return -ENOMEM;
196 	}
197 	options = tmp_options;
198 
199 	while ((p = strsep(&options, ",")) != NULL) {
200 		int token;
201 		int r;
202 		if (!*p)
203 			continue;
204 		token = match_token(p, tokens, args);
205 		if ((token != Opt_err) && (token != Opt_privport)) {
206 			r = match_int(&args[0], &option);
207 			if (r < 0) {
208 				p9_debug(P9_DEBUG_ERROR,
209 					 "integer field, but no integer?\n");
210 				continue;
211 			}
212 		}
213 		switch (token) {
214 		case Opt_port:
215 			opts->port = option;
216 			break;
217 		case Opt_sq_depth:
218 			opts->sq_depth = option;
219 			break;
220 		case Opt_rq_depth:
221 			opts->rq_depth = option;
222 			break;
223 		case Opt_timeout:
224 			opts->timeout = option;
225 			break;
226 		case Opt_privport:
227 			opts->privport = true;
228 			break;
229 		default:
230 			continue;
231 		}
232 	}
233 	/* RQ must be at least as large as the SQ */
234 	opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
235 	kfree(tmp_options);
236 	return 0;
237 }
238 
239 static int
240 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
241 {
242 	struct p9_client *c = id->context;
243 	struct p9_trans_rdma *rdma = c->trans;
244 	switch (event->event) {
245 	case RDMA_CM_EVENT_ADDR_RESOLVED:
246 		BUG_ON(rdma->state != P9_RDMA_INIT);
247 		rdma->state = P9_RDMA_ADDR_RESOLVED;
248 		break;
249 
250 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
251 		BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
252 		rdma->state = P9_RDMA_ROUTE_RESOLVED;
253 		break;
254 
255 	case RDMA_CM_EVENT_ESTABLISHED:
256 		BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
257 		rdma->state = P9_RDMA_CONNECTED;
258 		break;
259 
260 	case RDMA_CM_EVENT_DISCONNECTED:
261 		if (rdma)
262 			rdma->state = P9_RDMA_CLOSED;
263 		c->status = Disconnected;
264 		break;
265 
266 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
267 		break;
268 
269 	case RDMA_CM_EVENT_ADDR_CHANGE:
270 	case RDMA_CM_EVENT_ROUTE_ERROR:
271 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
272 	case RDMA_CM_EVENT_MULTICAST_JOIN:
273 	case RDMA_CM_EVENT_MULTICAST_ERROR:
274 	case RDMA_CM_EVENT_REJECTED:
275 	case RDMA_CM_EVENT_CONNECT_REQUEST:
276 	case RDMA_CM_EVENT_CONNECT_RESPONSE:
277 	case RDMA_CM_EVENT_CONNECT_ERROR:
278 	case RDMA_CM_EVENT_ADDR_ERROR:
279 	case RDMA_CM_EVENT_UNREACHABLE:
280 		c->status = Disconnected;
281 		rdma_disconnect(rdma->cm_id);
282 		break;
283 	default:
284 		BUG();
285 	}
286 	complete(&rdma->cm_done);
287 	return 0;
288 }
289 
290 static void
291 recv_done(struct ib_cq *cq, struct ib_wc *wc)
292 {
293 	struct p9_client *client = cq->cq_context;
294 	struct p9_trans_rdma *rdma = client->trans;
295 	struct p9_rdma_context *c =
296 		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
297 	struct p9_req_t *req;
298 	int err = 0;
299 	int16_t tag;
300 
301 	req = NULL;
302 	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
303 							 DMA_FROM_DEVICE);
304 
305 	if (wc->status != IB_WC_SUCCESS)
306 		goto err_out;
307 
308 	c->rc.size = wc->byte_len;
309 	err = p9_parse_header(&c->rc, NULL, NULL, &tag, 1);
310 	if (err)
311 		goto err_out;
312 
313 	req = p9_tag_lookup(client, tag);
314 	if (!req)
315 		goto err_out;
316 
317 	/* Check that we have not yet received a reply for this request.
318 	 */
319 	if (unlikely(req->rc.sdata)) {
320 		pr_err("Duplicate reply for request %d", tag);
321 		goto err_out;
322 	}
323 
324 	req->rc.size = c->rc.size;
325 	req->rc.sdata = c->rc.sdata;
326 	p9_client_cb(client, req, REQ_STATUS_RCVD);
327 
328  out:
329 	up(&rdma->rq_sem);
330 	kfree(c);
331 	return;
332 
333  err_out:
334 	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
335 			req, err, wc->status);
336 	rdma->state = P9_RDMA_FLUSHING;
337 	client->status = Disconnected;
338 	goto out;
339 }
340 
341 static void
342 send_done(struct ib_cq *cq, struct ib_wc *wc)
343 {
344 	struct p9_client *client = cq->cq_context;
345 	struct p9_trans_rdma *rdma = client->trans;
346 	struct p9_rdma_context *c =
347 		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
348 
349 	ib_dma_unmap_single(rdma->cm_id->device,
350 			    c->busa, c->req->tc.size,
351 			    DMA_TO_DEVICE);
352 	up(&rdma->sq_sem);
353 	p9_req_put(c->req);
354 	kfree(c);
355 }
356 
357 static void qp_event_handler(struct ib_event *event, void *context)
358 {
359 	p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
360 		 event->event, context);
361 }
362 
363 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
364 {
365 	if (!rdma)
366 		return;
367 
368 	if (rdma->qp && !IS_ERR(rdma->qp))
369 		ib_destroy_qp(rdma->qp);
370 
371 	if (rdma->pd && !IS_ERR(rdma->pd))
372 		ib_dealloc_pd(rdma->pd);
373 
374 	if (rdma->cq && !IS_ERR(rdma->cq))
375 		ib_free_cq(rdma->cq);
376 
377 	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
378 		rdma_destroy_id(rdma->cm_id);
379 
380 	kfree(rdma);
381 }
382 
383 static int
384 post_recv(struct p9_client *client, struct p9_rdma_context *c)
385 {
386 	struct p9_trans_rdma *rdma = client->trans;
387 	struct ib_recv_wr wr;
388 	struct ib_sge sge;
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 	return ib_post_recv(rdma->qp, &wr, NULL);
407 
408  error:
409 	p9_debug(P9_DEBUG_ERROR, "EIO\n");
410 	return -EIO;
411 }
412 
413 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
414 {
415 	struct p9_trans_rdma *rdma = client->trans;
416 	struct ib_send_wr wr;
417 	struct ib_sge sge;
418 	int err = 0;
419 	unsigned long flags;
420 	struct p9_rdma_context *c = NULL;
421 	struct p9_rdma_context *rpl_context = NULL;
422 
423 	/* When an error occurs between posting the recv and the send,
424 	 * there will be a receive context posted without a pending request.
425 	 * Since there is no way to "un-post" it, we remember it and skip
426 	 * post_recv() for the next request.
427 	 * So here,
428 	 * see if we are this `next request' and need to absorb an excess rc.
429 	 * If yes, then drop and free our own, and do not recv_post().
430 	 **/
431 	if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
432 		if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
433 			/* Got one! */
434 			p9_fcall_fini(&req->rc);
435 			req->rc.sdata = NULL;
436 			goto dont_need_post_recv;
437 		} else {
438 			/* We raced and lost. */
439 			atomic_inc(&rdma->excess_rc);
440 		}
441 	}
442 
443 	/* Allocate an fcall for the reply */
444 	rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
445 	if (!rpl_context) {
446 		err = -ENOMEM;
447 		goto recv_error;
448 	}
449 	rpl_context->rc.sdata = req->rc.sdata;
450 
451 	/*
452 	 * Post a receive buffer for this request. We need to ensure
453 	 * there is a reply buffer available for every outstanding
454 	 * request. A flushed request can result in no reply for an
455 	 * outstanding request, so we must keep a count to avoid
456 	 * overflowing the RQ.
457 	 */
458 	if (down_interruptible(&rdma->rq_sem)) {
459 		err = -EINTR;
460 		goto recv_error;
461 	}
462 
463 	err = post_recv(client, rpl_context);
464 	if (err) {
465 		p9_debug(P9_DEBUG_ERROR, "POST RECV failed: %d\n", err);
466 		goto recv_error;
467 	}
468 	/* remove posted receive buffer from request structure */
469 	req->rc.sdata = NULL;
470 
471 dont_need_post_recv:
472 	/* Post the request */
473 	c = kmalloc(sizeof *c, GFP_NOFS);
474 	if (!c) {
475 		err = -ENOMEM;
476 		goto send_error;
477 	}
478 	c->req = req;
479 
480 	c->busa = ib_dma_map_single(rdma->cm_id->device,
481 				    c->req->tc.sdata, c->req->tc.size,
482 				    DMA_TO_DEVICE);
483 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
484 		err = -EIO;
485 		goto send_error;
486 	}
487 
488 	c->cqe.done = send_done;
489 
490 	sge.addr = c->busa;
491 	sge.length = c->req->tc.size;
492 	sge.lkey = rdma->pd->local_dma_lkey;
493 
494 	wr.next = NULL;
495 	wr.wr_cqe = &c->cqe;
496 	wr.opcode = IB_WR_SEND;
497 	wr.send_flags = IB_SEND_SIGNALED;
498 	wr.sg_list = &sge;
499 	wr.num_sge = 1;
500 
501 	if (down_interruptible(&rdma->sq_sem)) {
502 		err = -EINTR;
503 		goto send_error;
504 	}
505 
506 	/* Mark request as `sent' *before* we actually send it,
507 	 * because doing if after could erase the REQ_STATUS_RCVD
508 	 * status in case of a very fast reply.
509 	 */
510 	req->status = REQ_STATUS_SENT;
511 	err = ib_post_send(rdma->qp, &wr, NULL);
512 	if (err)
513 		goto send_error;
514 
515 	/* Success */
516 	return 0;
517 
518  /* Handle errors that happened during or while preparing the send: */
519  send_error:
520 	req->status = REQ_STATUS_ERROR;
521 	kfree(c);
522 	p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
523 
524 	/* Ach.
525 	 *  We did recv_post(), but not send. We have one recv_post in excess.
526 	 */
527 	atomic_inc(&rdma->excess_rc);
528 	return err;
529 
530  /* Handle errors that happened during or while preparing post_recv(): */
531  recv_error:
532 	kfree(rpl_context);
533 	spin_lock_irqsave(&rdma->req_lock, flags);
534 	if (err != -EINTR && rdma->state < P9_RDMA_CLOSING) {
535 		rdma->state = P9_RDMA_CLOSING;
536 		spin_unlock_irqrestore(&rdma->req_lock, flags);
537 		rdma_disconnect(rdma->cm_id);
538 	} else
539 		spin_unlock_irqrestore(&rdma->req_lock, flags);
540 	return err;
541 }
542 
543 static void rdma_close(struct p9_client *client)
544 {
545 	struct p9_trans_rdma *rdma;
546 
547 	if (!client)
548 		return;
549 
550 	rdma = client->trans;
551 	if (!rdma)
552 		return;
553 
554 	client->status = Disconnected;
555 	rdma_disconnect(rdma->cm_id);
556 	rdma_destroy_trans(rdma);
557 }
558 
559 /**
560  * alloc_rdma - Allocate and initialize the rdma transport structure
561  * @opts: Mount options structure
562  */
563 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
564 {
565 	struct p9_trans_rdma *rdma;
566 
567 	rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
568 	if (!rdma)
569 		return NULL;
570 
571 	rdma->port = opts->port;
572 	rdma->privport = opts->privport;
573 	rdma->sq_depth = opts->sq_depth;
574 	rdma->rq_depth = opts->rq_depth;
575 	rdma->timeout = opts->timeout;
576 	spin_lock_init(&rdma->req_lock);
577 	init_completion(&rdma->cm_done);
578 	sema_init(&rdma->sq_sem, rdma->sq_depth);
579 	sema_init(&rdma->rq_sem, rdma->rq_depth);
580 	atomic_set(&rdma->excess_rc, 0);
581 
582 	return rdma;
583 }
584 
585 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
586 {
587 	/* Nothing to do here.
588 	 * We will take care of it (if we have to) in rdma_cancelled()
589 	 */
590 	return 1;
591 }
592 
593 /* A request has been fully flushed without a reply.
594  * That means we have posted one buffer in excess.
595  */
596 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
597 {
598 	struct p9_trans_rdma *rdma = client->trans;
599 	atomic_inc(&rdma->excess_rc);
600 	return 0;
601 }
602 
603 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
604 {
605 	struct sockaddr_in cl = {
606 		.sin_family = AF_INET,
607 		.sin_addr.s_addr = htonl(INADDR_ANY),
608 	};
609 	int port, err = -EINVAL;
610 
611 	for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
612 		cl.sin_port = htons((ushort)port);
613 		err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
614 		if (err != -EADDRINUSE)
615 			break;
616 	}
617 	return err;
618 }
619 
620 /**
621  * rdma_create_trans - Transport method for creating a transport instance
622  * @client: client instance
623  * @addr: IP address string
624  * @args: Mount options string
625  */
626 static int
627 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
628 {
629 	int err;
630 	struct p9_rdma_opts opts;
631 	struct p9_trans_rdma *rdma;
632 	struct rdma_conn_param conn_param;
633 	struct ib_qp_init_attr qp_attr;
634 
635 	if (addr == NULL)
636 		return -EINVAL;
637 
638 	/* Parse the transport specific mount options */
639 	err = parse_opts(args, &opts);
640 	if (err < 0)
641 		return err;
642 
643 	/* Create and initialize the RDMA transport structure */
644 	rdma = alloc_rdma(&opts);
645 	if (!rdma)
646 		return -ENOMEM;
647 
648 	/* Create the RDMA CM ID */
649 	rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
650 				     RDMA_PS_TCP, IB_QPT_RC);
651 	if (IS_ERR(rdma->cm_id))
652 		goto error;
653 
654 	/* Associate the client with the transport */
655 	client->trans = rdma;
656 
657 	/* Bind to a privileged port if we need to */
658 	if (opts.privport) {
659 		err = p9_rdma_bind_privport(rdma);
660 		if (err < 0) {
661 			pr_err("%s (%d): problem binding to privport: %d\n",
662 			       __func__, task_pid_nr(current), -err);
663 			goto error;
664 		}
665 	}
666 
667 	/* Resolve the server's address */
668 	rdma->addr.sin_family = AF_INET;
669 	rdma->addr.sin_addr.s_addr = in_aton(addr);
670 	rdma->addr.sin_port = htons(opts.port);
671 	err = rdma_resolve_addr(rdma->cm_id, NULL,
672 				(struct sockaddr *)&rdma->addr,
673 				rdma->timeout);
674 	if (err)
675 		goto error;
676 	err = wait_for_completion_interruptible(&rdma->cm_done);
677 	if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
678 		goto error;
679 
680 	/* Resolve the route to the server */
681 	err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
682 	if (err)
683 		goto error;
684 	err = wait_for_completion_interruptible(&rdma->cm_done);
685 	if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
686 		goto error;
687 
688 	/* Create the Completion Queue */
689 	rdma->cq = ib_alloc_cq_any(rdma->cm_id->device, client,
690 				   opts.sq_depth + opts.rq_depth + 1,
691 				   IB_POLL_SOFTIRQ);
692 	if (IS_ERR(rdma->cq))
693 		goto error;
694 
695 	/* Create the Protection Domain */
696 	rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
697 	if (IS_ERR(rdma->pd))
698 		goto error;
699 
700 	/* Create the Queue Pair */
701 	memset(&qp_attr, 0, sizeof qp_attr);
702 	qp_attr.event_handler = qp_event_handler;
703 	qp_attr.qp_context = client;
704 	qp_attr.cap.max_send_wr = opts.sq_depth;
705 	qp_attr.cap.max_recv_wr = opts.rq_depth;
706 	qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
707 	qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
708 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
709 	qp_attr.qp_type = IB_QPT_RC;
710 	qp_attr.send_cq = rdma->cq;
711 	qp_attr.recv_cq = rdma->cq;
712 	err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
713 	if (err)
714 		goto error;
715 	rdma->qp = rdma->cm_id->qp;
716 
717 	/* Request a connection */
718 	memset(&conn_param, 0, sizeof(conn_param));
719 	conn_param.private_data = NULL;
720 	conn_param.private_data_len = 0;
721 	conn_param.responder_resources = P9_RDMA_IRD;
722 	conn_param.initiator_depth = P9_RDMA_ORD;
723 	err = rdma_connect(rdma->cm_id, &conn_param);
724 	if (err)
725 		goto error;
726 	err = wait_for_completion_interruptible(&rdma->cm_done);
727 	if (err || (rdma->state != P9_RDMA_CONNECTED))
728 		goto error;
729 
730 	client->status = Connected;
731 
732 	return 0;
733 
734 error:
735 	rdma_destroy_trans(rdma);
736 	return -ENOTCONN;
737 }
738 
739 static struct p9_trans_module p9_rdma_trans = {
740 	.name = "rdma",
741 	.maxsize = P9_RDMA_MAXSIZE,
742 	.def = 0,
743 	.owner = THIS_MODULE,
744 	.create = rdma_create_trans,
745 	.close = rdma_close,
746 	.request = rdma_request,
747 	.cancel = rdma_cancel,
748 	.cancelled = rdma_cancelled,
749 	.show_options = p9_rdma_show_options,
750 };
751 
752 /**
753  * p9_trans_rdma_init - Register the 9P RDMA transport driver
754  */
755 static int __init p9_trans_rdma_init(void)
756 {
757 	v9fs_register_trans(&p9_rdma_trans);
758 	return 0;
759 }
760 
761 static void __exit p9_trans_rdma_exit(void)
762 {
763 	v9fs_unregister_trans(&p9_rdma_trans);
764 }
765 
766 module_init(p9_trans_rdma_init);
767 module_exit(p9_trans_rdma_exit);
768 
769 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
770 MODULE_DESCRIPTION("RDMA Transport for 9P");
771 MODULE_LICENSE("Dual BSD/GPL");
772