xref: /openbmc/linux/net/9p/trans_rdma.c (revision 12eb4683) 
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 <net/9p/9p.h>
47 #include <net/9p/client.h>
48 #include <net/9p/transport.h>
49 #include <rdma/ib_verbs.h>
50 #include <rdma/rdma_cm.h>
51 
52 #define P9_PORT			5640
53 #define P9_RDMA_SQ_DEPTH	32
54 #define P9_RDMA_RQ_DEPTH	32
55 #define P9_RDMA_SEND_SGE	4
56 #define P9_RDMA_RECV_SGE	4
57 #define P9_RDMA_IRD		0
58 #define P9_RDMA_ORD		0
59 #define P9_RDMA_TIMEOUT		30000		/* 30 seconds */
60 #define P9_RDMA_MAXSIZE		(1024*1024)	/* 1MB */
61 
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_sem: Semaphore for the RQ
77  * @excess_rc : Amount of posted Receive Contexts without a pending request.
78  *		See rdma_request()
79  * @addr: The remote peer's address
80  * @req_lock: Protects the active request list
81  * @cm_done: Completion event for connection management tracking
82  */
83 struct p9_trans_rdma {
84 	enum {
85 		P9_RDMA_INIT,
86 		P9_RDMA_ADDR_RESOLVED,
87 		P9_RDMA_ROUTE_RESOLVED,
88 		P9_RDMA_CONNECTED,
89 		P9_RDMA_FLUSHING,
90 		P9_RDMA_CLOSING,
91 		P9_RDMA_CLOSED,
92 	} state;
93 	struct rdma_cm_id *cm_id;
94 	struct ib_pd *pd;
95 	struct ib_qp *qp;
96 	struct ib_cq *cq;
97 	struct ib_mr *dma_mr;
98 	u32 lkey;
99 	long timeout;
100 	int sq_depth;
101 	struct semaphore sq_sem;
102 	int rq_depth;
103 	struct semaphore rq_sem;
104 	atomic_t excess_rc;
105 	struct sockaddr_in addr;
106 	spinlock_t req_lock;
107 
108 	struct completion cm_done;
109 };
110 
111 /**
112  * p9_rdma_context - Keeps track of in-process WR
113  *
114  * @wc_op: The original WR op for when the CQE completes in error.
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 	enum ib_wc_opcode wc_op;
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 	int sq_depth;
140 	int rq_depth;
141 	long timeout;
142 };
143 
144 /*
145  * Option Parsing (code inspired by NFS code)
146  */
147 enum {
148 	/* Options that take integer arguments */
149 	Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
150 };
151 
152 static match_table_t tokens = {
153 	{Opt_port, "port=%u"},
154 	{Opt_sq_depth, "sq=%u"},
155 	{Opt_rq_depth, "rq=%u"},
156 	{Opt_timeout, "timeout=%u"},
157 	{Opt_err, NULL},
158 };
159 
160 /**
161  * parse_opts - parse mount options into rdma options structure
162  * @params: options string passed from mount
163  * @opts: rdma transport-specific structure to parse options into
164  *
165  * Returns 0 upon success, -ERRNO upon failure
166  */
167 static int parse_opts(char *params, struct p9_rdma_opts *opts)
168 {
169 	char *p;
170 	substring_t args[MAX_OPT_ARGS];
171 	int option;
172 	char *options, *tmp_options;
173 
174 	opts->port = P9_PORT;
175 	opts->sq_depth = P9_RDMA_SQ_DEPTH;
176 	opts->rq_depth = P9_RDMA_RQ_DEPTH;
177 	opts->timeout = P9_RDMA_TIMEOUT;
178 
179 	if (!params)
180 		return 0;
181 
182 	tmp_options = kstrdup(params, GFP_KERNEL);
183 	if (!tmp_options) {
184 		p9_debug(P9_DEBUG_ERROR,
185 			 "failed to allocate copy of option string\n");
186 		return -ENOMEM;
187 	}
188 	options = tmp_options;
189 
190 	while ((p = strsep(&options, ",")) != NULL) {
191 		int token;
192 		int r;
193 		if (!*p)
194 			continue;
195 		token = match_token(p, tokens, args);
196 		r = match_int(&args[0], &option);
197 		if (r < 0) {
198 			p9_debug(P9_DEBUG_ERROR,
199 				 "integer field, but no integer?\n");
200 			continue;
201 		}
202 		switch (token) {
203 		case Opt_port:
204 			opts->port = option;
205 			break;
206 		case Opt_sq_depth:
207 			opts->sq_depth = option;
208 			break;
209 		case Opt_rq_depth:
210 			opts->rq_depth = option;
211 			break;
212 		case Opt_timeout:
213 			opts->timeout = option;
214 			break;
215 		default:
216 			continue;
217 		}
218 	}
219 	/* RQ must be at least as large as the SQ */
220 	opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
221 	kfree(tmp_options);
222 	return 0;
223 }
224 
225 static int
226 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
227 {
228 	struct p9_client *c = id->context;
229 	struct p9_trans_rdma *rdma = c->trans;
230 	switch (event->event) {
231 	case RDMA_CM_EVENT_ADDR_RESOLVED:
232 		BUG_ON(rdma->state != P9_RDMA_INIT);
233 		rdma->state = P9_RDMA_ADDR_RESOLVED;
234 		break;
235 
236 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
237 		BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
238 		rdma->state = P9_RDMA_ROUTE_RESOLVED;
239 		break;
240 
241 	case RDMA_CM_EVENT_ESTABLISHED:
242 		BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
243 		rdma->state = P9_RDMA_CONNECTED;
244 		break;
245 
246 	case RDMA_CM_EVENT_DISCONNECTED:
247 		if (rdma)
248 			rdma->state = P9_RDMA_CLOSED;
249 		if (c)
250 			c->status = Disconnected;
251 		break;
252 
253 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
254 		break;
255 
256 	case RDMA_CM_EVENT_ADDR_CHANGE:
257 	case RDMA_CM_EVENT_ROUTE_ERROR:
258 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
259 	case RDMA_CM_EVENT_MULTICAST_JOIN:
260 	case RDMA_CM_EVENT_MULTICAST_ERROR:
261 	case RDMA_CM_EVENT_REJECTED:
262 	case RDMA_CM_EVENT_CONNECT_REQUEST:
263 	case RDMA_CM_EVENT_CONNECT_RESPONSE:
264 	case RDMA_CM_EVENT_CONNECT_ERROR:
265 	case RDMA_CM_EVENT_ADDR_ERROR:
266 	case RDMA_CM_EVENT_UNREACHABLE:
267 		c->status = Disconnected;
268 		rdma_disconnect(rdma->cm_id);
269 		break;
270 	default:
271 		BUG();
272 	}
273 	complete(&rdma->cm_done);
274 	return 0;
275 }
276 
277 static void
278 handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
279 	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
280 {
281 	struct p9_req_t *req;
282 	int err = 0;
283 	int16_t tag;
284 
285 	req = NULL;
286 	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
287 							 DMA_FROM_DEVICE);
288 
289 	if (status != IB_WC_SUCCESS)
290 		goto err_out;
291 
292 	err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
293 	if (err)
294 		goto err_out;
295 
296 	req = p9_tag_lookup(client, tag);
297 	if (!req)
298 		goto err_out;
299 
300 	/* Check that we have not yet received a reply for this request.
301 	 */
302 	if (unlikely(req->rc)) {
303 		pr_err("Duplicate reply for request %d", tag);
304 		goto err_out;
305 	}
306 
307 	req->rc = c->rc;
308 	req->status = REQ_STATUS_RCVD;
309 	p9_client_cb(client, req);
310 
311 	return;
312 
313  err_out:
314 	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
315 	rdma->state = P9_RDMA_FLUSHING;
316 	client->status = Disconnected;
317 }
318 
319 static void
320 handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
321 	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
322 {
323 	ib_dma_unmap_single(rdma->cm_id->device,
324 			    c->busa, c->req->tc->size,
325 			    DMA_TO_DEVICE);
326 }
327 
328 static void qp_event_handler(struct ib_event *event, void *context)
329 {
330 	p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
331 		 event->event, context);
332 }
333 
334 static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
335 {
336 	struct p9_client *client = cq_context;
337 	struct p9_trans_rdma *rdma = client->trans;
338 	int ret;
339 	struct ib_wc wc;
340 
341 	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
342 	while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
343 		struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
344 
345 		switch (c->wc_op) {
346 		case IB_WC_RECV:
347 			handle_recv(client, rdma, c, wc.status, wc.byte_len);
348 			up(&rdma->rq_sem);
349 			break;
350 
351 		case IB_WC_SEND:
352 			handle_send(client, rdma, c, wc.status, wc.byte_len);
353 			up(&rdma->sq_sem);
354 			break;
355 
356 		default:
357 			pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
358 			       c->wc_op, wc.opcode, wc.status);
359 			break;
360 		}
361 		kfree(c);
362 	}
363 }
364 
365 static void cq_event_handler(struct ib_event *e, void *v)
366 {
367 	p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
368 }
369 
370 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
371 {
372 	if (!rdma)
373 		return;
374 
375 	if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
376 		ib_dereg_mr(rdma->dma_mr);
377 
378 	if (rdma->qp && !IS_ERR(rdma->qp))
379 		ib_destroy_qp(rdma->qp);
380 
381 	if (rdma->pd && !IS_ERR(rdma->pd))
382 		ib_dealloc_pd(rdma->pd);
383 
384 	if (rdma->cq && !IS_ERR(rdma->cq))
385 		ib_destroy_cq(rdma->cq);
386 
387 	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
388 		rdma_destroy_id(rdma->cm_id);
389 
390 	kfree(rdma);
391 }
392 
393 static int
394 post_recv(struct p9_client *client, struct p9_rdma_context *c)
395 {
396 	struct p9_trans_rdma *rdma = client->trans;
397 	struct ib_recv_wr wr, *bad_wr;
398 	struct ib_sge sge;
399 
400 	c->busa = ib_dma_map_single(rdma->cm_id->device,
401 				    c->rc->sdata, client->msize,
402 				    DMA_FROM_DEVICE);
403 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
404 		goto error;
405 
406 	sge.addr = c->busa;
407 	sge.length = client->msize;
408 	sge.lkey = rdma->lkey;
409 
410 	wr.next = NULL;
411 	c->wc_op = IB_WC_RECV;
412 	wr.wr_id = (unsigned long) c;
413 	wr.sg_list = &sge;
414 	wr.num_sge = 1;
415 	return ib_post_recv(rdma->qp, &wr, &bad_wr);
416 
417  error:
418 	p9_debug(P9_DEBUG_ERROR, "EIO\n");
419 	return -EIO;
420 }
421 
422 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
423 {
424 	struct p9_trans_rdma *rdma = client->trans;
425 	struct ib_send_wr wr, *bad_wr;
426 	struct ib_sge sge;
427 	int err = 0;
428 	unsigned long flags;
429 	struct p9_rdma_context *c = NULL;
430 	struct p9_rdma_context *rpl_context = NULL;
431 
432 	/* When an error occurs between posting the recv and the send,
433 	 * there will be a receive context posted without a pending request.
434 	 * Since there is no way to "un-post" it, we remember it and skip
435 	 * post_recv() for the next request.
436 	 * So here,
437 	 * see if we are this `next request' and need to absorb an excess rc.
438 	 * If yes, then drop and free our own, and do not recv_post().
439 	 **/
440 	if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
441 		if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
442 			/* Got one ! */
443 			kfree(req->rc);
444 			req->rc = NULL;
445 			goto dont_need_post_recv;
446 		} else {
447 			/* We raced and lost. */
448 			atomic_inc(&rdma->excess_rc);
449 		}
450 	}
451 
452 	/* Allocate an fcall for the reply */
453 	rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
454 	if (!rpl_context) {
455 		err = -ENOMEM;
456 		goto recv_error;
457 	}
458 	rpl_context->rc = req->rc;
459 
460 	/*
461 	 * Post a receive buffer for this request. We need to ensure
462 	 * there is a reply buffer available for every outstanding
463 	 * request. A flushed request can result in no reply for an
464 	 * outstanding request, so we must keep a count to avoid
465 	 * overflowing the RQ.
466 	 */
467 	if (down_interruptible(&rdma->rq_sem)) {
468 		err = -EINTR;
469 		goto recv_error;
470 	}
471 
472 	err = post_recv(client, rpl_context);
473 	if (err) {
474 		p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
475 		goto recv_error;
476 	}
477 	/* remove posted receive buffer from request structure */
478 	req->rc = NULL;
479 
480 dont_need_post_recv:
481 	/* Post the request */
482 	c = kmalloc(sizeof *c, GFP_NOFS);
483 	if (!c) {
484 		err = -ENOMEM;
485 		goto send_error;
486 	}
487 	c->req = req;
488 
489 	c->busa = ib_dma_map_single(rdma->cm_id->device,
490 				    c->req->tc->sdata, c->req->tc->size,
491 				    DMA_TO_DEVICE);
492 	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
493 		err = -EIO;
494 		goto send_error;
495 	}
496 
497 	sge.addr = c->busa;
498 	sge.length = c->req->tc->size;
499 	sge.lkey = rdma->lkey;
500 
501 	wr.next = NULL;
502 	c->wc_op = IB_WC_SEND;
503 	wr.wr_id = (unsigned long) c;
504 	wr.opcode = IB_WR_SEND;
505 	wr.send_flags = IB_SEND_SIGNALED;
506 	wr.sg_list = &sge;
507 	wr.num_sge = 1;
508 
509 	if (down_interruptible(&rdma->sq_sem)) {
510 		err = -EINTR;
511 		goto send_error;
512 	}
513 
514 	err = ib_post_send(rdma->qp, &wr, &bad_wr);
515 	if (err)
516 		goto send_error;
517 
518 	/* Success */
519 	return 0;
520 
521  /* Handle errors that happened during or while preparing the send: */
522  send_error:
523 	kfree(c);
524 	p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
525 
526 	/* Ach.
527 	 *  We did recv_post(), but not send. We have one recv_post in excess.
528 	 */
529 	atomic_inc(&rdma->excess_rc);
530 	return err;
531 
532  /* Handle errors that happened during or while preparing post_recv(): */
533  recv_error:
534 	kfree(rpl_context);
535 	spin_lock_irqsave(&rdma->req_lock, flags);
536 	if (rdma->state < P9_RDMA_CLOSING) {
537 		rdma->state = P9_RDMA_CLOSING;
538 		spin_unlock_irqrestore(&rdma->req_lock, flags);
539 		rdma_disconnect(rdma->cm_id);
540 	} else
541 		spin_unlock_irqrestore(&rdma->req_lock, flags);
542 	return err;
543 }
544 
545 static void rdma_close(struct p9_client *client)
546 {
547 	struct p9_trans_rdma *rdma;
548 
549 	if (!client)
550 		return;
551 
552 	rdma = client->trans;
553 	if (!rdma)
554 		return;
555 
556 	client->status = Disconnected;
557 	rdma_disconnect(rdma->cm_id);
558 	rdma_destroy_trans(rdma);
559 }
560 
561 /**
562  * alloc_rdma - Allocate and initialize the rdma transport structure
563  * @opts: Mount options structure
564  */
565 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
566 {
567 	struct p9_trans_rdma *rdma;
568 
569 	rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
570 	if (!rdma)
571 		return NULL;
572 
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 /* its not clear to me we can do anything after send has been posted */
586 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
587 {
588 	return 1;
589 }
590 
591 /**
592  * trans_create_rdma - Transport method for creating atransport instance
593  * @client: client instance
594  * @addr: IP address string
595  * @args: Mount options string
596  */
597 static int
598 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
599 {
600 	int err;
601 	struct p9_rdma_opts opts;
602 	struct p9_trans_rdma *rdma;
603 	struct rdma_conn_param conn_param;
604 	struct ib_qp_init_attr qp_attr;
605 	struct ib_device_attr devattr;
606 
607 	/* Parse the transport specific mount options */
608 	err = parse_opts(args, &opts);
609 	if (err < 0)
610 		return err;
611 
612 	/* Create and initialize the RDMA transport structure */
613 	rdma = alloc_rdma(&opts);
614 	if (!rdma)
615 		return -ENOMEM;
616 
617 	/* Create the RDMA CM ID */
618 	rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
619 				     IB_QPT_RC);
620 	if (IS_ERR(rdma->cm_id))
621 		goto error;
622 
623 	/* Associate the client with the transport */
624 	client->trans = rdma;
625 
626 	/* Resolve the server's address */
627 	rdma->addr.sin_family = AF_INET;
628 	rdma->addr.sin_addr.s_addr = in_aton(addr);
629 	rdma->addr.sin_port = htons(opts.port);
630 	err = rdma_resolve_addr(rdma->cm_id, NULL,
631 				(struct sockaddr *)&rdma->addr,
632 				rdma->timeout);
633 	if (err)
634 		goto error;
635 	err = wait_for_completion_interruptible(&rdma->cm_done);
636 	if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
637 		goto error;
638 
639 	/* Resolve the route to the server */
640 	err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
641 	if (err)
642 		goto error;
643 	err = wait_for_completion_interruptible(&rdma->cm_done);
644 	if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
645 		goto error;
646 
647 	/* Query the device attributes */
648 	err = ib_query_device(rdma->cm_id->device, &devattr);
649 	if (err)
650 		goto error;
651 
652 	/* Create the Completion Queue */
653 	rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
654 				cq_event_handler, client,
655 				opts.sq_depth + opts.rq_depth + 1, 0);
656 	if (IS_ERR(rdma->cq))
657 		goto error;
658 	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
659 
660 	/* Create the Protection Domain */
661 	rdma->pd = ib_alloc_pd(rdma->cm_id->device);
662 	if (IS_ERR(rdma->pd))
663 		goto error;
664 
665 	/* Cache the DMA lkey in the transport */
666 	rdma->dma_mr = NULL;
667 	if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
668 		rdma->lkey = rdma->cm_id->device->local_dma_lkey;
669 	else {
670 		rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
671 		if (IS_ERR(rdma->dma_mr))
672 			goto error;
673 		rdma->lkey = rdma->dma_mr->lkey;
674 	}
675 
676 	/* Create the Queue Pair */
677 	memset(&qp_attr, 0, sizeof qp_attr);
678 	qp_attr.event_handler = qp_event_handler;
679 	qp_attr.qp_context = client;
680 	qp_attr.cap.max_send_wr = opts.sq_depth;
681 	qp_attr.cap.max_recv_wr = opts.rq_depth;
682 	qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
683 	qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
684 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
685 	qp_attr.qp_type = IB_QPT_RC;
686 	qp_attr.send_cq = rdma->cq;
687 	qp_attr.recv_cq = rdma->cq;
688 	err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
689 	if (err)
690 		goto error;
691 	rdma->qp = rdma->cm_id->qp;
692 
693 	/* Request a connection */
694 	memset(&conn_param, 0, sizeof(conn_param));
695 	conn_param.private_data = NULL;
696 	conn_param.private_data_len = 0;
697 	conn_param.responder_resources = P9_RDMA_IRD;
698 	conn_param.initiator_depth = P9_RDMA_ORD;
699 	err = rdma_connect(rdma->cm_id, &conn_param);
700 	if (err)
701 		goto error;
702 	err = wait_for_completion_interruptible(&rdma->cm_done);
703 	if (err || (rdma->state != P9_RDMA_CONNECTED))
704 		goto error;
705 
706 	client->status = Connected;
707 
708 	return 0;
709 
710 error:
711 	rdma_destroy_trans(rdma);
712 	return -ENOTCONN;
713 }
714 
715 static struct p9_trans_module p9_rdma_trans = {
716 	.name = "rdma",
717 	.maxsize = P9_RDMA_MAXSIZE,
718 	.def = 0,
719 	.owner = THIS_MODULE,
720 	.create = rdma_create_trans,
721 	.close = rdma_close,
722 	.request = rdma_request,
723 	.cancel = rdma_cancel,
724 };
725 
726 /**
727  * p9_trans_rdma_init - Register the 9P RDMA transport driver
728  */
729 static int __init p9_trans_rdma_init(void)
730 {
731 	v9fs_register_trans(&p9_rdma_trans);
732 	return 0;
733 }
734 
735 static void __exit p9_trans_rdma_exit(void)
736 {
737 	v9fs_unregister_trans(&p9_rdma_trans);
738 }
739 
740 module_init(p9_trans_rdma_init);
741 module_exit(p9_trans_rdma_exit);
742 
743 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
744 MODULE_DESCRIPTION("RDMA Transport for 9P");
745 MODULE_LICENSE("Dual BSD/GPL");
746