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 */
83struct 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 */
119struct p9_rdma_req;
120struct 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 */
137struct p9_rdma_opts {
138	short port;
139	int sq_depth;
140	int rq_depth;
141	long timeout;
142	int privport;
143};
144
145/*
146 * Option Parsing (code inspired by NFS code)
147 */
148enum {
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
156static 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/**
166 * parse_opts - parse mount options into rdma options structure
167 * @params: options string passed from mount
168 * @opts: rdma transport-specific structure to parse options into
169 *
170 * Returns 0 upon success, -ERRNO upon failure
171 */
172static int parse_opts(char *params, struct p9_rdma_opts *opts)
173{
174	char *p;
175	substring_t args[MAX_OPT_ARGS];
176	int option;
177	char *options, *tmp_options;
178
179	opts->port = P9_PORT;
180	opts->sq_depth = P9_RDMA_SQ_DEPTH;
181	opts->rq_depth = P9_RDMA_RQ_DEPTH;
182	opts->timeout = P9_RDMA_TIMEOUT;
183	opts->privport = 0;
184
185	if (!params)
186		return 0;
187
188	tmp_options = kstrdup(params, GFP_KERNEL);
189	if (!tmp_options) {
190		p9_debug(P9_DEBUG_ERROR,
191			 "failed to allocate copy of option string\n");
192		return -ENOMEM;
193	}
194	options = tmp_options;
195
196	while ((p = strsep(&options, ",")) != NULL) {
197		int token;
198		int r;
199		if (!*p)
200			continue;
201		token = match_token(p, tokens, args);
202		if ((token != Opt_err) && (token != Opt_privport)) {
203			r = match_int(&args[0], &option);
204			if (r < 0) {
205				p9_debug(P9_DEBUG_ERROR,
206					 "integer field, but no integer?\n");
207				continue;
208			}
209		}
210		switch (token) {
211		case Opt_port:
212			opts->port = option;
213			break;
214		case Opt_sq_depth:
215			opts->sq_depth = option;
216			break;
217		case Opt_rq_depth:
218			opts->rq_depth = option;
219			break;
220		case Opt_timeout:
221			opts->timeout = option;
222			break;
223		case Opt_privport:
224			opts->privport = 1;
225			break;
226		default:
227			continue;
228		}
229	}
230	/* RQ must be at least as large as the SQ */
231	opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
232	kfree(tmp_options);
233	return 0;
234}
235
236static int
237p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
238{
239	struct p9_client *c = id->context;
240	struct p9_trans_rdma *rdma = c->trans;
241	switch (event->event) {
242	case RDMA_CM_EVENT_ADDR_RESOLVED:
243		BUG_ON(rdma->state != P9_RDMA_INIT);
244		rdma->state = P9_RDMA_ADDR_RESOLVED;
245		break;
246
247	case RDMA_CM_EVENT_ROUTE_RESOLVED:
248		BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
249		rdma->state = P9_RDMA_ROUTE_RESOLVED;
250		break;
251
252	case RDMA_CM_EVENT_ESTABLISHED:
253		BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
254		rdma->state = P9_RDMA_CONNECTED;
255		break;
256
257	case RDMA_CM_EVENT_DISCONNECTED:
258		if (rdma)
259			rdma->state = P9_RDMA_CLOSED;
260		if (c)
261			c->status = Disconnected;
262		break;
263
264	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
265		break;
266
267	case RDMA_CM_EVENT_ADDR_CHANGE:
268	case RDMA_CM_EVENT_ROUTE_ERROR:
269	case RDMA_CM_EVENT_DEVICE_REMOVAL:
270	case RDMA_CM_EVENT_MULTICAST_JOIN:
271	case RDMA_CM_EVENT_MULTICAST_ERROR:
272	case RDMA_CM_EVENT_REJECTED:
273	case RDMA_CM_EVENT_CONNECT_REQUEST:
274	case RDMA_CM_EVENT_CONNECT_RESPONSE:
275	case RDMA_CM_EVENT_CONNECT_ERROR:
276	case RDMA_CM_EVENT_ADDR_ERROR:
277	case RDMA_CM_EVENT_UNREACHABLE:
278		c->status = Disconnected;
279		rdma_disconnect(rdma->cm_id);
280		break;
281	default:
282		BUG();
283	}
284	complete(&rdma->cm_done);
285	return 0;
286}
287
288static void
289handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
290	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
291{
292	struct p9_req_t *req;
293	int err = 0;
294	int16_t tag;
295
296	req = NULL;
297	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
298							 DMA_FROM_DEVICE);
299
300	if (status != IB_WC_SUCCESS)
301		goto err_out;
302
303	err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
304	if (err)
305		goto err_out;
306
307	req = p9_tag_lookup(client, tag);
308	if (!req)
309		goto err_out;
310
311	/* Check that we have not yet received a reply for this request.
312	 */
313	if (unlikely(req->rc)) {
314		pr_err("Duplicate reply for request %d", tag);
315		goto err_out;
316	}
317
318	req->rc = c->rc;
319	p9_client_cb(client, req, REQ_STATUS_RCVD);
320
321	return;
322
323 err_out:
324	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
325	rdma->state = P9_RDMA_FLUSHING;
326	client->status = Disconnected;
327}
328
329static void
330handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
331	    struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
332{
333	ib_dma_unmap_single(rdma->cm_id->device,
334			    c->busa, c->req->tc->size,
335			    DMA_TO_DEVICE);
336}
337
338static void qp_event_handler(struct ib_event *event, void *context)
339{
340	p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
341		 event->event, context);
342}
343
344static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
345{
346	struct p9_client *client = cq_context;
347	struct p9_trans_rdma *rdma = client->trans;
348	int ret;
349	struct ib_wc wc;
350
351	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
352	while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
353		struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
354
355		switch (c->wc_op) {
356		case IB_WC_RECV:
357			handle_recv(client, rdma, c, wc.status, wc.byte_len);
358			up(&rdma->rq_sem);
359			break;
360
361		case IB_WC_SEND:
362			handle_send(client, rdma, c, wc.status, wc.byte_len);
363			up(&rdma->sq_sem);
364			break;
365
366		default:
367			pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
368			       c->wc_op, wc.opcode, wc.status);
369			break;
370		}
371		kfree(c);
372	}
373}
374
375static void cq_event_handler(struct ib_event *e, void *v)
376{
377	p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
378}
379
380static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
381{
382	if (!rdma)
383		return;
384
385	if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
386		ib_dereg_mr(rdma->dma_mr);
387
388	if (rdma->qp && !IS_ERR(rdma->qp))
389		ib_destroy_qp(rdma->qp);
390
391	if (rdma->pd && !IS_ERR(rdma->pd))
392		ib_dealloc_pd(rdma->pd);
393
394	if (rdma->cq && !IS_ERR(rdma->cq))
395		ib_destroy_cq(rdma->cq);
396
397	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
398		rdma_destroy_id(rdma->cm_id);
399
400	kfree(rdma);
401}
402
403static int
404post_recv(struct p9_client *client, struct p9_rdma_context *c)
405{
406	struct p9_trans_rdma *rdma = client->trans;
407	struct ib_recv_wr wr, *bad_wr;
408	struct ib_sge sge;
409
410	c->busa = ib_dma_map_single(rdma->cm_id->device,
411				    c->rc->sdata, client->msize,
412				    DMA_FROM_DEVICE);
413	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
414		goto error;
415
416	sge.addr = c->busa;
417	sge.length = client->msize;
418	sge.lkey = rdma->lkey;
419
420	wr.next = NULL;
421	c->wc_op = IB_WC_RECV;
422	wr.wr_id = (unsigned long) c;
423	wr.sg_list = &sge;
424	wr.num_sge = 1;
425	return ib_post_recv(rdma->qp, &wr, &bad_wr);
426
427 error:
428	p9_debug(P9_DEBUG_ERROR, "EIO\n");
429	return -EIO;
430}
431
432static int rdma_request(struct p9_client *client, struct p9_req_t *req)
433{
434	struct p9_trans_rdma *rdma = client->trans;
435	struct ib_send_wr wr, *bad_wr;
436	struct ib_sge sge;
437	int err = 0;
438	unsigned long flags;
439	struct p9_rdma_context *c = NULL;
440	struct p9_rdma_context *rpl_context = NULL;
441
442	/* When an error occurs between posting the recv and the send,
443	 * there will be a receive context posted without a pending request.
444	 * Since there is no way to "un-post" it, we remember it and skip
445	 * post_recv() for the next request.
446	 * So here,
447	 * see if we are this `next request' and need to absorb an excess rc.
448	 * If yes, then drop and free our own, and do not recv_post().
449	 **/
450	if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
451		if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
452			/* Got one ! */
453			kfree(req->rc);
454			req->rc = NULL;
455			goto dont_need_post_recv;
456		} else {
457			/* We raced and lost. */
458			atomic_inc(&rdma->excess_rc);
459		}
460	}
461
462	/* Allocate an fcall for the reply */
463	rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
464	if (!rpl_context) {
465		err = -ENOMEM;
466		goto recv_error;
467	}
468	rpl_context->rc = req->rc;
469
470	/*
471	 * Post a receive buffer for this request. We need to ensure
472	 * there is a reply buffer available for every outstanding
473	 * request. A flushed request can result in no reply for an
474	 * outstanding request, so we must keep a count to avoid
475	 * overflowing the RQ.
476	 */
477	if (down_interruptible(&rdma->rq_sem)) {
478		err = -EINTR;
479		goto recv_error;
480	}
481
482	err = post_recv(client, rpl_context);
483	if (err) {
484		p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
485		goto recv_error;
486	}
487	/* remove posted receive buffer from request structure */
488	req->rc = NULL;
489
490dont_need_post_recv:
491	/* Post the request */
492	c = kmalloc(sizeof *c, GFP_NOFS);
493	if (!c) {
494		err = -ENOMEM;
495		goto send_error;
496	}
497	c->req = req;
498
499	c->busa = ib_dma_map_single(rdma->cm_id->device,
500				    c->req->tc->sdata, c->req->tc->size,
501				    DMA_TO_DEVICE);
502	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
503		err = -EIO;
504		goto send_error;
505	}
506
507	sge.addr = c->busa;
508	sge.length = c->req->tc->size;
509	sge.lkey = rdma->lkey;
510
511	wr.next = NULL;
512	c->wc_op = IB_WC_SEND;
513	wr.wr_id = (unsigned long) c;
514	wr.opcode = IB_WR_SEND;
515	wr.send_flags = IB_SEND_SIGNALED;
516	wr.sg_list = &sge;
517	wr.num_sge = 1;
518
519	if (down_interruptible(&rdma->sq_sem)) {
520		err = -EINTR;
521		goto send_error;
522	}
523
524	/* Mark request as `sent' *before* we actually send it,
525	 * because doing if after could erase the REQ_STATUS_RCVD
526	 * status in case of a very fast reply.
527	 */
528	req->status = REQ_STATUS_SENT;
529	err = ib_post_send(rdma->qp, &wr, &bad_wr);
530	if (err)
531		goto send_error;
532
533	/* Success */
534	return 0;
535
536 /* Handle errors that happened during or while preparing the send: */
537 send_error:
538	req->status = REQ_STATUS_ERROR;
539	kfree(c);
540	p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
541
542	/* Ach.
543	 *  We did recv_post(), but not send. We have one recv_post in excess.
544	 */
545	atomic_inc(&rdma->excess_rc);
546	return err;
547
548 /* Handle errors that happened during or while preparing post_recv(): */
549 recv_error:
550	kfree(rpl_context);
551	spin_lock_irqsave(&rdma->req_lock, flags);
552	if (rdma->state < P9_RDMA_CLOSING) {
553		rdma->state = P9_RDMA_CLOSING;
554		spin_unlock_irqrestore(&rdma->req_lock, flags);
555		rdma_disconnect(rdma->cm_id);
556	} else
557		spin_unlock_irqrestore(&rdma->req_lock, flags);
558	return err;
559}
560
561static void rdma_close(struct p9_client *client)
562{
563	struct p9_trans_rdma *rdma;
564
565	if (!client)
566		return;
567
568	rdma = client->trans;
569	if (!rdma)
570		return;
571
572	client->status = Disconnected;
573	rdma_disconnect(rdma->cm_id);
574	rdma_destroy_trans(rdma);
575}
576
577/**
578 * alloc_rdma - Allocate and initialize the rdma transport structure
579 * @opts: Mount options structure
580 */
581static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
582{
583	struct p9_trans_rdma *rdma;
584
585	rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
586	if (!rdma)
587		return NULL;
588
589	rdma->sq_depth = opts->sq_depth;
590	rdma->rq_depth = opts->rq_depth;
591	rdma->timeout = opts->timeout;
592	spin_lock_init(&rdma->req_lock);
593	init_completion(&rdma->cm_done);
594	sema_init(&rdma->sq_sem, rdma->sq_depth);
595	sema_init(&rdma->rq_sem, rdma->rq_depth);
596	atomic_set(&rdma->excess_rc, 0);
597
598	return rdma;
599}
600
601static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
602{
603	/* Nothing to do here.
604	 * We will take care of it (if we have to) in rdma_cancelled()
605	 */
606	return 1;
607}
608
609/* A request has been fully flushed without a reply.
610 * That means we have posted one buffer in excess.
611 */
612static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
613{
614	struct p9_trans_rdma *rdma = client->trans;
615	atomic_inc(&rdma->excess_rc);
616	return 0;
617}
618
619static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
620{
621	struct sockaddr_in cl = {
622		.sin_family = AF_INET,
623		.sin_addr.s_addr = htonl(INADDR_ANY),
624	};
625	int port, err = -EINVAL;
626
627	for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
628		cl.sin_port = htons((ushort)port);
629		err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
630		if (err != -EADDRINUSE)
631			break;
632	}
633	return err;
634}
635
636/**
637 * trans_create_rdma - Transport method for creating atransport instance
638 * @client: client instance
639 * @addr: IP address string
640 * @args: Mount options string
641 */
642static int
643rdma_create_trans(struct p9_client *client, const char *addr, char *args)
644{
645	int err;
646	struct p9_rdma_opts opts;
647	struct p9_trans_rdma *rdma;
648	struct rdma_conn_param conn_param;
649	struct ib_qp_init_attr qp_attr;
650	struct ib_device_attr devattr;
651
652	/* Parse the transport specific mount options */
653	err = parse_opts(args, &opts);
654	if (err < 0)
655		return err;
656
657	/* Create and initialize the RDMA transport structure */
658	rdma = alloc_rdma(&opts);
659	if (!rdma)
660		return -ENOMEM;
661
662	/* Create the RDMA CM ID */
663	rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
664				     IB_QPT_RC);
665	if (IS_ERR(rdma->cm_id))
666		goto error;
667
668	/* Associate the client with the transport */
669	client->trans = rdma;
670
671	/* Bind to a privileged port if we need to */
672	if (opts.privport) {
673		err = p9_rdma_bind_privport(rdma);
674		if (err < 0) {
675			pr_err("%s (%d): problem binding to privport: %d\n",
676			       __func__, task_pid_nr(current), -err);
677			goto error;
678		}
679	}
680
681	/* Resolve the server's address */
682	rdma->addr.sin_family = AF_INET;
683	rdma->addr.sin_addr.s_addr = in_aton(addr);
684	rdma->addr.sin_port = htons(opts.port);
685	err = rdma_resolve_addr(rdma->cm_id, NULL,
686				(struct sockaddr *)&rdma->addr,
687				rdma->timeout);
688	if (err)
689		goto error;
690	err = wait_for_completion_interruptible(&rdma->cm_done);
691	if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
692		goto error;
693
694	/* Resolve the route to the server */
695	err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
696	if (err)
697		goto error;
698	err = wait_for_completion_interruptible(&rdma->cm_done);
699	if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
700		goto error;
701
702	/* Query the device attributes */
703	err = ib_query_device(rdma->cm_id->device, &devattr);
704	if (err)
705		goto error;
706
707	/* Create the Completion Queue */
708	rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
709				cq_event_handler, client,
710				opts.sq_depth + opts.rq_depth + 1, 0);
711	if (IS_ERR(rdma->cq))
712		goto error;
713	ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
714
715	/* Create the Protection Domain */
716	rdma->pd = ib_alloc_pd(rdma->cm_id->device);
717	if (IS_ERR(rdma->pd))
718		goto error;
719
720	/* Cache the DMA lkey in the transport */
721	rdma->dma_mr = NULL;
722	if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
723		rdma->lkey = rdma->cm_id->device->local_dma_lkey;
724	else {
725		rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
726		if (IS_ERR(rdma->dma_mr))
727			goto error;
728		rdma->lkey = rdma->dma_mr->lkey;
729	}
730
731	/* Create the Queue Pair */
732	memset(&qp_attr, 0, sizeof qp_attr);
733	qp_attr.event_handler = qp_event_handler;
734	qp_attr.qp_context = client;
735	qp_attr.cap.max_send_wr = opts.sq_depth;
736	qp_attr.cap.max_recv_wr = opts.rq_depth;
737	qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
738	qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
739	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
740	qp_attr.qp_type = IB_QPT_RC;
741	qp_attr.send_cq = rdma->cq;
742	qp_attr.recv_cq = rdma->cq;
743	err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
744	if (err)
745		goto error;
746	rdma->qp = rdma->cm_id->qp;
747
748	/* Request a connection */
749	memset(&conn_param, 0, sizeof(conn_param));
750	conn_param.private_data = NULL;
751	conn_param.private_data_len = 0;
752	conn_param.responder_resources = P9_RDMA_IRD;
753	conn_param.initiator_depth = P9_RDMA_ORD;
754	err = rdma_connect(rdma->cm_id, &conn_param);
755	if (err)
756		goto error;
757	err = wait_for_completion_interruptible(&rdma->cm_done);
758	if (err || (rdma->state != P9_RDMA_CONNECTED))
759		goto error;
760
761	client->status = Connected;
762
763	return 0;
764
765error:
766	rdma_destroy_trans(rdma);
767	return -ENOTCONN;
768}
769
770static struct p9_trans_module p9_rdma_trans = {
771	.name = "rdma",
772	.maxsize = P9_RDMA_MAXSIZE,
773	.def = 0,
774	.owner = THIS_MODULE,
775	.create = rdma_create_trans,
776	.close = rdma_close,
777	.request = rdma_request,
778	.cancel = rdma_cancel,
779	.cancelled = rdma_cancelled,
780};
781
782/**
783 * p9_trans_rdma_init - Register the 9P RDMA transport driver
784 */
785static int __init p9_trans_rdma_init(void)
786{
787	v9fs_register_trans(&p9_rdma_trans);
788	return 0;
789}
790
791static void __exit p9_trans_rdma_exit(void)
792{
793	v9fs_unregister_trans(&p9_rdma_trans);
794}
795
796module_init(p9_trans_rdma_init);
797module_exit(p9_trans_rdma_exit);
798
799MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
800MODULE_DESCRIPTION("RDMA Transport for 9P");
801MODULE_LICENSE("Dual BSD/GPL");
802