xref: /openbmc/linux/net/9p/trans_rdma.c (revision b6dcefde) 
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 #include <linux/in.h>
30 #include <linux/module.h>
31 #include <linux/net.h>
32 #include <linux/ipv6.h>
33 #include <linux/kthread.h>
34 #include <linux/errno.h>
35 #include <linux/kernel.h>
36 #include <linux/un.h>
37 #include <linux/uaccess.h>
38 #include <linux/inet.h>
39 #include <linux/idr.h>
40 #include <linux/file.h>
41 #include <linux/parser.h>
42 #include <linux/semaphore.h>
43 #include <net/9p/9p.h>
44 #include <net/9p/client.h>
45 #include <net/9p/transport.h>
46 #include <rdma/ib_verbs.h>
47 #include <rdma/rdma_cm.h>
48 
49 #define P9_PORT			5640
50 #define P9_RDMA_SQ_DEPTH	32
51 #define P9_RDMA_RQ_DEPTH	32
52 #define P9_RDMA_SEND_SGE	4
53 #define P9_RDMA_RECV_SGE	4
54 #define P9_RDMA_IRD		0
55 #define P9_RDMA_ORD		0
56 #define P9_RDMA_TIMEOUT		30000		/* 30 seconds */
57 #define P9_RDMA_MAXSIZE		(4*4096)	/* Min SGE is 4, so we can
58 						 * safely advertise a maxsize
59 						 * of 64k */
60 
61 #define P9_RDMA_MAX_SGE (P9_RDMA_MAXSIZE >> PAGE_SHIFT)
62 /**
63  * struct p9_trans_rdma - RDMA transport instance
64  *
65  * @state: tracks the transport state machine for connection setup and tear down
66  * @cm_id: The RDMA CM ID
67  * @pd: Protection Domain pointer
68  * @qp: Queue Pair pointer
69  * @cq: Completion Queue pointer
70  * @dm_mr: DMA Memory Region pointer
71  * @lkey: The local access only memory region key
72  * @timeout: Number of uSecs to wait for connection management events
73  * @sq_depth: The depth of the Send Queue
74  * @sq_sem: Semaphore for the SQ
75  * @rq_depth: The depth of the Receive Queue.
76  * @rq_count: Count of requests in the Receive Queue.
77  * @addr: The remote peer's address
78  * @req_lock: Protects the active request list
79  * @cm_done: Completion event for connection management tracking
80  */
81 struct p9_trans_rdma {
82 	enum {
83 		P9_RDMA_INIT,
84 		P9_RDMA_ADDR_RESOLVED,
85 		P9_RDMA_ROUTE_RESOLVED,
86 		P9_RDMA_CONNECTED,
87 		P9_RDMA_FLUSHING,
88 		P9_RDMA_CLOSING,
89 		P9_RDMA_CLOSED,
90 	} state;
91 	struct rdma_cm_id *cm_id;
92 	struct ib_pd *pd;
93 	struct ib_qp *qp;
94 	struct ib_cq *cq;
95 	struct ib_mr *dma_mr;
96 	u32 lkey;
97 	long timeout;
98 	int sq_depth;
99 	struct semaphore sq_sem;
100 	int rq_depth;
101 	atomic_t rq_count;
102 	struct sockaddr_in addr;
103 	spinlock_t req_lock;
104 
105 	struct completion cm_done;
106 };
107 
108 /**
109  * p9_rdma_context - Keeps track of in-process WR
110  *
111  * @wc_op: The original WR op for when the CQE completes in error.
112  * @busa: Bus address to unmap when the WR completes
113  * @req: Keeps track of requests (send)
114  * @rc: Keepts track of replies (receive)
115  */
116 struct p9_rdma_req;
117 struct p9_rdma_context {
118 	enum ib_wc_opcode wc_op;
119 	dma_addr_t busa;
120 	union {
121 		struct p9_req_t *req;
122 		struct p9_fcall *rc;
123 	};
124 };
125 
126 /**
127  * p9_rdma_opts - Collection of mount options
128  * @port: port of connection
129  * @sq_depth: The requested depth of the SQ. This really doesn't need
130  * to be any deeper than the number of threads used in the client
131  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
132  * @timeout: Time to wait in msecs for CM events
133  */
134 struct p9_rdma_opts {
135 	short port;
136 	int sq_depth;
137 	int rq_depth;
138 	long timeout;
139 };
140 
141 /*
142  * Option Parsing (code inspired by NFS code)
143  */
144 enum {
145 	/* Options that take integer arguments */
146 	Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
147 };
148 
149 static match_table_t tokens = {
150 	{Opt_port, "port=%u"},
151 	{Opt_sq_depth, "sq=%u"},
152 	{Opt_rq_depth, "rq=%u"},
153 	{Opt_timeout, "timeout=%u"},
154 	{Opt_err, NULL},
155 };
156 
157 /**
158  * parse_opts - parse mount options into rdma options structure
159  * @params: options string passed from mount
160  * @opts: rdma transport-specific structure to parse options into
161  *
162  * Returns 0 upon success, -ERRNO upon failure
163  */
164 static int parse_opts(char *params, struct p9_rdma_opts *opts)
165 {
166 	char *p;
167 	substring_t args[MAX_OPT_ARGS];
168 	int option;
169 	char *options, *tmp_options;
170 	int ret;
171 
172 	opts->port = P9_PORT;
173 	opts->sq_depth = P9_RDMA_SQ_DEPTH;
174 	opts->rq_depth = P9_RDMA_RQ_DEPTH;
175 	opts->timeout = P9_RDMA_TIMEOUT;
176 
177 	if (!params)
178 		return 0;
179 
180 	tmp_options = kstrdup(params, GFP_KERNEL);
181 	if (!tmp_options) {
182 		P9_DPRINTK(P9_DEBUG_ERROR,
183 			   "failed to allocate copy of option string\n");
184 		return -ENOMEM;
185 	}
186 	options = tmp_options;
187 
188 	while ((p = strsep(&options, ",")) != NULL) {
189 		int token;
190 		int r;
191 		if (!*p)
192 			continue;
193 		token = match_token(p, tokens, args);
194 		r = match_int(&args[0], &option);
195 		if (r < 0) {
196 			P9_DPRINTK(P9_DEBUG_ERROR,
197 				   "integer field, but no integer?\n");
198 			ret = r;
199 			continue;
200 		}
201 		switch (token) {
202 		case Opt_port:
203 			opts->port = option;
204 			break;
205 		case Opt_sq_depth:
206 			opts->sq_depth = option;
207 			break;
208 		case Opt_rq_depth:
209 			opts->rq_depth = option;
210 			break;
211 		case Opt_timeout:
212 			opts->timeout = option;
213 			break;
214 		default:
215 			continue;
216 		}
217 	}
218 	/* RQ must be at least as large as the SQ */
219 	opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
220 	kfree(tmp_options);
221 	return 0;
222 }
223 
224 static int
225 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
226 {
227 	struct p9_client *c = id->context;
228 	struct p9_trans_rdma *rdma = c->trans;
229 	switch (event->event) {
230 	case RDMA_CM_EVENT_ADDR_RESOLVED:
231 		BUG_ON(rdma->state != P9_RDMA_INIT);
232 		rdma->state = P9_RDMA_ADDR_RESOLVED;
233 		break;
234 
235 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
236 		BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
237 		rdma->state = P9_RDMA_ROUTE_RESOLVED;
238 		break;
239 
240 	case RDMA_CM_EVENT_ESTABLISHED:
241 		BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
242 		rdma->state = P9_RDMA_CONNECTED;
243 		break;
244 
245 	case RDMA_CM_EVENT_DISCONNECTED:
246 		if (rdma)
247 			rdma->state = P9_RDMA_CLOSED;
248 		if (c)
249 			c->status = Disconnected;
250 		break;
251 
252 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
253 		break;
254 
255 	case RDMA_CM_EVENT_ADDR_CHANGE:
256 	case RDMA_CM_EVENT_ROUTE_ERROR:
257 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
258 	case RDMA_CM_EVENT_MULTICAST_JOIN:
259 	case RDMA_CM_EVENT_MULTICAST_ERROR:
260 	case RDMA_CM_EVENT_REJECTED:
261 	case RDMA_CM_EVENT_CONNECT_REQUEST:
262 	case RDMA_CM_EVENT_CONNECT_RESPONSE:
263 	case RDMA_CM_EVENT_CONNECT_ERROR:
264 	case RDMA_CM_EVENT_ADDR_ERROR:
265 	case RDMA_CM_EVENT_UNREACHABLE:
266 		c->status = Disconnected;
267 		rdma_disconnect(rdma->cm_id);
268 		break;
269 	default:
270 		BUG();
271 	}
272 	complete(&rdma->cm_done);
273 	return 0;
274 }
275 
276 static void
277 handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
278 	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
279 {
280 	struct p9_req_t *req;
281 	int err = 0;
282 	int16_t tag;
283 
284 	req = NULL;
285 	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
286 							 DMA_FROM_DEVICE);
287 
288 	if (status != IB_WC_SUCCESS)
289 		goto err_out;
290 
291 	err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
292 	if (err)
293 		goto err_out;
294 
295 	req = p9_tag_lookup(client, tag);
296 	if (!req)
297 		goto err_out;
298 
299 	req->rc = c->rc;
300 	req->status = REQ_STATUS_RCVD;
301 	p9_client_cb(client, req);
302 
303 	return;
304 
305  err_out:
306 	P9_DPRINTK(P9_DEBUG_ERROR, "req %p err %d status %d\n",
307 		   req, err, status);
308 	rdma->state = P9_RDMA_FLUSHING;
309 	client->status = Disconnected;
310 	return;
311 }
312 
313 static void
314 handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
315 	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
316 {
317 	ib_dma_unmap_single(rdma->cm_id->device,
318 			    c->busa, c->req->tc->size,
319 			    DMA_TO_DEVICE);
320 }
321 
322 static void qp_event_handler(struct ib_event *event, void *context)
323 {
324 	P9_DPRINTK(P9_DEBUG_ERROR, "QP event %d context %p\n", event->event,
325 								context);
326 }
327 
328 static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
329 {
330 	struct p9_client *client = cq_context;
331 	struct p9_trans_rdma *rdma = client->trans;
332 	int ret;
333 	struct ib_wc wc;
334 
335 	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
336 	while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
337 		struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
338 
339 		switch (c->wc_op) {
340 		case IB_WC_RECV:
341 			atomic_dec(&rdma->rq_count);
342 			handle_recv(client, rdma, c, wc.status, wc.byte_len);
343 			break;
344 
345 		case IB_WC_SEND:
346 			handle_send(client, rdma, c, wc.status, wc.byte_len);
347 			up(&rdma->sq_sem);
348 			break;
349 
350 		default:
351 			printk(KERN_ERR "9prdma: unexpected completion type, "
352 			       "c->wc_op=%d, wc.opcode=%d, status=%d\n",
353 			       c->wc_op, wc.opcode, wc.status);
354 			break;
355 		}
356 		kfree(c);
357 	}
358 }
359 
360 static void cq_event_handler(struct ib_event *e, void *v)
361 {
362 	P9_DPRINTK(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
363 }
364 
365 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
366 {
367 	if (!rdma)
368 		return;
369 
370 	if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
371 		ib_dereg_mr(rdma->dma_mr);
372 
373 	if (rdma->qp && !IS_ERR(rdma->qp))
374 		ib_destroy_qp(rdma->qp);
375 
376 	if (rdma->pd && !IS_ERR(rdma->pd))
377 		ib_dealloc_pd(rdma->pd);
378 
379 	if (rdma->cq && !IS_ERR(rdma->cq))
380 		ib_destroy_cq(rdma->cq);
381 
382 	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
383 		rdma_destroy_id(rdma->cm_id);
384 
385 	kfree(rdma);
386 }
387 
388 static int
389 post_recv(struct p9_client *client, struct p9_rdma_context *c)
390 {
391 	struct p9_trans_rdma *rdma = client->trans;
392 	struct ib_recv_wr wr, *bad_wr;
393 	struct ib_sge sge;
394 
395 	c->busa = ib_dma_map_single(rdma->cm_id->device,
396 				    c->rc->sdata, client->msize,
397 				    DMA_FROM_DEVICE);
398 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
399 		goto error;
400 
401 	sge.addr = c->busa;
402 	sge.length = client->msize;
403 	sge.lkey = rdma->lkey;
404 
405 	wr.next = NULL;
406 	c->wc_op = IB_WC_RECV;
407 	wr.wr_id = (unsigned long) c;
408 	wr.sg_list = &sge;
409 	wr.num_sge = 1;
410 	return ib_post_recv(rdma->qp, &wr, &bad_wr);
411 
412  error:
413 	P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
414 	return -EIO;
415 }
416 
417 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
418 {
419 	struct p9_trans_rdma *rdma = client->trans;
420 	struct ib_send_wr wr, *bad_wr;
421 	struct ib_sge sge;
422 	int err = 0;
423 	unsigned long flags;
424 	struct p9_rdma_context *c = NULL;
425 	struct p9_rdma_context *rpl_context = NULL;
426 
427 	/* Allocate an fcall for the reply */
428 	rpl_context = kmalloc(sizeof *rpl_context, GFP_KERNEL);
429 	if (!rpl_context)
430 		goto err_close;
431 
432 	/*
433 	 * If the request has a buffer, steal it, otherwise
434 	 * allocate a new one.  Typically, requests should already
435 	 * have receive buffers allocated and just swap them around
436 	 */
437 	if (!req->rc) {
438 		req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
439 								GFP_KERNEL);
440 		if (req->rc) {
441 			req->rc->sdata = (char *) req->rc +
442 						sizeof(struct p9_fcall);
443 			req->rc->capacity = client->msize;
444 		}
445 	}
446 	rpl_context->rc = req->rc;
447 	if (!rpl_context->rc) {
448 		kfree(rpl_context);
449 		goto err_close;
450 	}
451 
452 	/*
453 	 * Post a receive buffer for this request. We need to ensure
454 	 * there is a reply buffer available for every outstanding
455 	 * request. A flushed request can result in no reply for an
456 	 * outstanding request, so we must keep a count to avoid
457 	 * overflowing the RQ.
458 	 */
459 	if (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
460 		err = post_recv(client, rpl_context);
461 		if (err) {
462 			kfree(rpl_context->rc);
463 			kfree(rpl_context);
464 			goto err_close;
465 		}
466 	} else
467 		atomic_dec(&rdma->rq_count);
468 
469 	/* remove posted receive buffer from request structure */
470 	req->rc = NULL;
471 
472 	/* Post the request */
473 	c = kmalloc(sizeof *c, GFP_KERNEL);
474 	if (!c)
475 		goto err_close;
476 	c->req = req;
477 
478 	c->busa = ib_dma_map_single(rdma->cm_id->device,
479 				    c->req->tc->sdata, c->req->tc->size,
480 				    DMA_TO_DEVICE);
481 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
482 		goto error;
483 
484 	sge.addr = c->busa;
485 	sge.length = c->req->tc->size;
486 	sge.lkey = rdma->lkey;
487 
488 	wr.next = NULL;
489 	c->wc_op = IB_WC_SEND;
490 	wr.wr_id = (unsigned long) c;
491 	wr.opcode = IB_WR_SEND;
492 	wr.send_flags = IB_SEND_SIGNALED;
493 	wr.sg_list = &sge;
494 	wr.num_sge = 1;
495 
496 	if (down_interruptible(&rdma->sq_sem))
497 		goto error;
498 
499 	return ib_post_send(rdma->qp, &wr, &bad_wr);
500 
501  error:
502 	P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
503 	return -EIO;
504 
505  err_close:
506 	spin_lock_irqsave(&rdma->req_lock, flags);
507 	if (rdma->state < P9_RDMA_CLOSING) {
508 		rdma->state = P9_RDMA_CLOSING;
509 		spin_unlock_irqrestore(&rdma->req_lock, flags);
510 		rdma_disconnect(rdma->cm_id);
511 	} else
512 		spin_unlock_irqrestore(&rdma->req_lock, flags);
513 	return err;
514 }
515 
516 static void rdma_close(struct p9_client *client)
517 {
518 	struct p9_trans_rdma *rdma;
519 
520 	if (!client)
521 		return;
522 
523 	rdma = client->trans;
524 	if (!rdma)
525 		return;
526 
527 	client->status = Disconnected;
528 	rdma_disconnect(rdma->cm_id);
529 	rdma_destroy_trans(rdma);
530 }
531 
532 /**
533  * alloc_rdma - Allocate and initialize the rdma transport structure
534  * @opts: Mount options structure
535  */
536 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
537 {
538 	struct p9_trans_rdma *rdma;
539 
540 	rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
541 	if (!rdma)
542 		return NULL;
543 
544 	rdma->sq_depth = opts->sq_depth;
545 	rdma->rq_depth = opts->rq_depth;
546 	rdma->timeout = opts->timeout;
547 	spin_lock_init(&rdma->req_lock);
548 	init_completion(&rdma->cm_done);
549 	sema_init(&rdma->sq_sem, rdma->sq_depth);
550 	atomic_set(&rdma->rq_count, 0);
551 
552 	return rdma;
553 }
554 
555 /* its not clear to me we can do anything after send has been posted */
556 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
557 {
558 	return 1;
559 }
560 
561 /**
562  * trans_create_rdma - Transport method for creating atransport instance
563  * @client: client instance
564  * @addr: IP address string
565  * @args: Mount options string
566  */
567 static int
568 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
569 {
570 	int err;
571 	struct p9_rdma_opts opts;
572 	struct p9_trans_rdma *rdma;
573 	struct rdma_conn_param conn_param;
574 	struct ib_qp_init_attr qp_attr;
575 	struct ib_device_attr devattr;
576 
577 	/* Parse the transport specific mount options */
578 	err = parse_opts(args, &opts);
579 	if (err < 0)
580 		return err;
581 
582 	/* Create and initialize the RDMA transport structure */
583 	rdma = alloc_rdma(&opts);
584 	if (!rdma)
585 		return -ENOMEM;
586 
587 	/* Create the RDMA CM ID */
588 	rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP);
589 	if (IS_ERR(rdma->cm_id))
590 		goto error;
591 
592 	/* Associate the client with the transport */
593 	client->trans = rdma;
594 
595 	/* Resolve the server's address */
596 	rdma->addr.sin_family = AF_INET;
597 	rdma->addr.sin_addr.s_addr = in_aton(addr);
598 	rdma->addr.sin_port = htons(opts.port);
599 	err = rdma_resolve_addr(rdma->cm_id, NULL,
600 				(struct sockaddr *)&rdma->addr,
601 				rdma->timeout);
602 	if (err)
603 		goto error;
604 	err = wait_for_completion_interruptible(&rdma->cm_done);
605 	if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
606 		goto error;
607 
608 	/* Resolve the route to the server */
609 	err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
610 	if (err)
611 		goto error;
612 	err = wait_for_completion_interruptible(&rdma->cm_done);
613 	if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
614 		goto error;
615 
616 	/* Query the device attributes */
617 	err = ib_query_device(rdma->cm_id->device, &devattr);
618 	if (err)
619 		goto error;
620 
621 	/* Create the Completion Queue */
622 	rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
623 				cq_event_handler, client,
624 				opts.sq_depth + opts.rq_depth + 1, 0);
625 	if (IS_ERR(rdma->cq))
626 		goto error;
627 	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
628 
629 	/* Create the Protection Domain */
630 	rdma->pd = ib_alloc_pd(rdma->cm_id->device);
631 	if (IS_ERR(rdma->pd))
632 		goto error;
633 
634 	/* Cache the DMA lkey in the transport */
635 	rdma->dma_mr = NULL;
636 	if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
637 		rdma->lkey = rdma->cm_id->device->local_dma_lkey;
638 	else {
639 		rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
640 		if (IS_ERR(rdma->dma_mr))
641 			goto error;
642 		rdma->lkey = rdma->dma_mr->lkey;
643 	}
644 
645 	/* Create the Queue Pair */
646 	memset(&qp_attr, 0, sizeof qp_attr);
647 	qp_attr.event_handler = qp_event_handler;
648 	qp_attr.qp_context = client;
649 	qp_attr.cap.max_send_wr = opts.sq_depth;
650 	qp_attr.cap.max_recv_wr = opts.rq_depth;
651 	qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
652 	qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
653 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
654 	qp_attr.qp_type = IB_QPT_RC;
655 	qp_attr.send_cq = rdma->cq;
656 	qp_attr.recv_cq = rdma->cq;
657 	err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
658 	if (err)
659 		goto error;
660 	rdma->qp = rdma->cm_id->qp;
661 
662 	/* Request a connection */
663 	memset(&conn_param, 0, sizeof(conn_param));
664 	conn_param.private_data = NULL;
665 	conn_param.private_data_len = 0;
666 	conn_param.responder_resources = P9_RDMA_IRD;
667 	conn_param.initiator_depth = P9_RDMA_ORD;
668 	err = rdma_connect(rdma->cm_id, &conn_param);
669 	if (err)
670 		goto error;
671 	err = wait_for_completion_interruptible(&rdma->cm_done);
672 	if (err || (rdma->state != P9_RDMA_CONNECTED))
673 		goto error;
674 
675 	client->status = Connected;
676 
677 	return 0;
678 
679 error:
680 	rdma_destroy_trans(rdma);
681 	return -ENOTCONN;
682 }
683 
684 static struct p9_trans_module p9_rdma_trans = {
685 	.name = "rdma",
686 	.maxsize = P9_RDMA_MAXSIZE,
687 	.def = 0,
688 	.owner = THIS_MODULE,
689 	.create = rdma_create_trans,
690 	.close = rdma_close,
691 	.request = rdma_request,
692 	.cancel = rdma_cancel,
693 };
694 
695 /**
696  * p9_trans_rdma_init - Register the 9P RDMA transport driver
697  */
698 static int __init p9_trans_rdma_init(void)
699 {
700 	v9fs_register_trans(&p9_rdma_trans);
701 	return 0;
702 }
703 
704 static void __exit p9_trans_rdma_exit(void)
705 {
706 	v9fs_unregister_trans(&p9_rdma_trans);
707 }
708 
709 module_init(p9_trans_rdma_init);
710 module_exit(p9_trans_rdma_exit);
711 
712 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
713 MODULE_DESCRIPTION("RDMA Transport for 9P");
714 MODULE_LICENSE("Dual BSD/GPL");
715