2ab4574183c93f8c403fa79346cb2a3935ce5d02
[muen/linux.git] / net / 9p / trans_rdma.c
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  * @timeout: Number of uSecs to wait for connection management events
72  * @privport: Whether a privileged port may be used
73  * @port: The port to use
74  * @sq_depth: The depth of the Send Queue
75  * @sq_sem: Semaphore for the SQ
76  * @rq_depth: The depth of the Receive Queue.
77  * @rq_sem: Semaphore for the RQ
78  * @excess_rc : Amount of posted Receive Contexts without a pending request.
79  *              See rdma_request()
80  * @addr: The remote peer's address
81  * @req_lock: Protects the active request list
82  * @cm_done: Completion event for connection management tracking
83  */
84 struct p9_trans_rdma {
85         enum {
86                 P9_RDMA_INIT,
87                 P9_RDMA_ADDR_RESOLVED,
88                 P9_RDMA_ROUTE_RESOLVED,
89                 P9_RDMA_CONNECTED,
90                 P9_RDMA_FLUSHING,
91                 P9_RDMA_CLOSING,
92                 P9_RDMA_CLOSED,
93         } state;
94         struct rdma_cm_id *cm_id;
95         struct ib_pd *pd;
96         struct ib_qp *qp;
97         struct ib_cq *cq;
98         long timeout;
99         bool privport;
100         u16 port;
101         int sq_depth;
102         struct semaphore sq_sem;
103         int rq_depth;
104         struct semaphore rq_sem;
105         atomic_t excess_rc;
106         struct sockaddr_in addr;
107         spinlock_t req_lock;
108
109         struct completion cm_done;
110 };
111
112 /**
113  * p9_rdma_context - Keeps track of in-process WR
114  *
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         struct ib_cqe cqe;
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         bool privport;
140         int sq_depth;
141         int rq_depth;
142         long timeout;
143 };
144
145 /*
146  * Option Parsing (code inspired by NFS code)
147  */
148 enum {
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
156 static 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 static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
166 {
167         struct p9_trans_rdma *rdma = clnt->trans;
168
169         if (rdma->port != P9_PORT)
170                 seq_printf(m, ",port=%u", rdma->port);
171         if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
172                 seq_printf(m, ",sq=%u", rdma->sq_depth);
173         if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
174                 seq_printf(m, ",rq=%u", rdma->rq_depth);
175         if (rdma->timeout != P9_RDMA_TIMEOUT)
176                 seq_printf(m, ",timeout=%lu", rdma->timeout);
177         if (rdma->privport)
178                 seq_puts(m, ",privport");
179         return 0;
180 }
181
182 /**
183  * parse_opts - parse mount options into rdma options structure
184  * @params: options string passed from mount
185  * @opts: rdma transport-specific structure to parse options into
186  *
187  * Returns 0 upon success, -ERRNO upon failure
188  */
189 static int parse_opts(char *params, struct p9_rdma_opts *opts)
190 {
191         char *p;
192         substring_t args[MAX_OPT_ARGS];
193         int option;
194         char *options, *tmp_options;
195
196         opts->port = P9_PORT;
197         opts->sq_depth = P9_RDMA_SQ_DEPTH;
198         opts->rq_depth = P9_RDMA_RQ_DEPTH;
199         opts->timeout = P9_RDMA_TIMEOUT;
200         opts->privport = false;
201
202         if (!params)
203                 return 0;
204
205         tmp_options = kstrdup(params, GFP_KERNEL);
206         if (!tmp_options) {
207                 p9_debug(P9_DEBUG_ERROR,
208                          "failed to allocate copy of option string\n");
209                 return -ENOMEM;
210         }
211         options = tmp_options;
212
213         while ((p = strsep(&options, ",")) != NULL) {
214                 int token;
215                 int r;
216                 if (!*p)
217                         continue;
218                 token = match_token(p, tokens, args);
219                 if ((token != Opt_err) && (token != Opt_privport)) {
220                         r = match_int(&args[0], &option);
221                         if (r < 0) {
222                                 p9_debug(P9_DEBUG_ERROR,
223                                          "integer field, but no integer?\n");
224                                 continue;
225                         }
226                 }
227                 switch (token) {
228                 case Opt_port:
229                         opts->port = option;
230                         break;
231                 case Opt_sq_depth:
232                         opts->sq_depth = option;
233                         break;
234                 case Opt_rq_depth:
235                         opts->rq_depth = option;
236                         break;
237                 case Opt_timeout:
238                         opts->timeout = option;
239                         break;
240                 case Opt_privport:
241                         opts->privport = true;
242                         break;
243                 default:
244                         continue;
245                 }
246         }
247         /* RQ must be at least as large as the SQ */
248         opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
249         kfree(tmp_options);
250         return 0;
251 }
252
253 static int
254 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
255 {
256         struct p9_client *c = id->context;
257         struct p9_trans_rdma *rdma = c->trans;
258         switch (event->event) {
259         case RDMA_CM_EVENT_ADDR_RESOLVED:
260                 BUG_ON(rdma->state != P9_RDMA_INIT);
261                 rdma->state = P9_RDMA_ADDR_RESOLVED;
262                 break;
263
264         case RDMA_CM_EVENT_ROUTE_RESOLVED:
265                 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
266                 rdma->state = P9_RDMA_ROUTE_RESOLVED;
267                 break;
268
269         case RDMA_CM_EVENT_ESTABLISHED:
270                 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
271                 rdma->state = P9_RDMA_CONNECTED;
272                 break;
273
274         case RDMA_CM_EVENT_DISCONNECTED:
275                 if (rdma)
276                         rdma->state = P9_RDMA_CLOSED;
277                 if (c)
278                         c->status = Disconnected;
279                 break;
280
281         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
282                 break;
283
284         case RDMA_CM_EVENT_ADDR_CHANGE:
285         case RDMA_CM_EVENT_ROUTE_ERROR:
286         case RDMA_CM_EVENT_DEVICE_REMOVAL:
287         case RDMA_CM_EVENT_MULTICAST_JOIN:
288         case RDMA_CM_EVENT_MULTICAST_ERROR:
289         case RDMA_CM_EVENT_REJECTED:
290         case RDMA_CM_EVENT_CONNECT_REQUEST:
291         case RDMA_CM_EVENT_CONNECT_RESPONSE:
292         case RDMA_CM_EVENT_CONNECT_ERROR:
293         case RDMA_CM_EVENT_ADDR_ERROR:
294         case RDMA_CM_EVENT_UNREACHABLE:
295                 c->status = Disconnected;
296                 rdma_disconnect(rdma->cm_id);
297                 break;
298         default:
299                 BUG();
300         }
301         complete(&rdma->cm_done);
302         return 0;
303 }
304
305 static void
306 recv_done(struct ib_cq *cq, struct ib_wc *wc)
307 {
308         struct p9_client *client = cq->cq_context;
309         struct p9_trans_rdma *rdma = client->trans;
310         struct p9_rdma_context *c =
311                 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
312         struct p9_req_t *req;
313         int err = 0;
314         int16_t tag;
315
316         req = NULL;
317         ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
318                                                          DMA_FROM_DEVICE);
319
320         if (wc->status != IB_WC_SUCCESS)
321                 goto err_out;
322
323         c->rc->size = wc->byte_len;
324         err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
325         if (err)
326                 goto err_out;
327
328         req = p9_tag_lookup(client, tag);
329         if (!req)
330                 goto err_out;
331
332         /* Check that we have not yet received a reply for this request.
333          */
334         if (unlikely(req->rc)) {
335                 pr_err("Duplicate reply for request %d", tag);
336                 goto err_out;
337         }
338
339         req->rc = c->rc;
340         p9_client_cb(client, req, REQ_STATUS_RCVD);
341
342  out:
343         up(&rdma->rq_sem);
344         kfree(c);
345         return;
346
347  err_out:
348         p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
349                         req, err, wc->status);
350         rdma->state = P9_RDMA_FLUSHING;
351         client->status = Disconnected;
352         goto out;
353 }
354
355 static void
356 send_done(struct ib_cq *cq, struct ib_wc *wc)
357 {
358         struct p9_client *client = cq->cq_context;
359         struct p9_trans_rdma *rdma = client->trans;
360         struct p9_rdma_context *c =
361                 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
362
363         ib_dma_unmap_single(rdma->cm_id->device,
364                             c->busa, c->req->tc->size,
365                             DMA_TO_DEVICE);
366         up(&rdma->sq_sem);
367         kfree(c);
368 }
369
370 static void qp_event_handler(struct ib_event *event, void *context)
371 {
372         p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
373                  event->event, context);
374 }
375
376 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
377 {
378         if (!rdma)
379                 return;
380
381         if (rdma->qp && !IS_ERR(rdma->qp))
382                 ib_destroy_qp(rdma->qp);
383
384         if (rdma->pd && !IS_ERR(rdma->pd))
385                 ib_dealloc_pd(rdma->pd);
386
387         if (rdma->cq && !IS_ERR(rdma->cq))
388                 ib_free_cq(rdma->cq);
389
390         if (rdma->cm_id && !IS_ERR(rdma->cm_id))
391                 rdma_destroy_id(rdma->cm_id);
392
393         kfree(rdma);
394 }
395
396 static int
397 post_recv(struct p9_client *client, struct p9_rdma_context *c)
398 {
399         struct p9_trans_rdma *rdma = client->trans;
400         struct ib_recv_wr wr, *bad_wr;
401         struct ib_sge sge;
402
403         c->busa = ib_dma_map_single(rdma->cm_id->device,
404                                     c->rc->sdata, client->msize,
405                                     DMA_FROM_DEVICE);
406         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
407                 goto error;
408
409         c->cqe.done = recv_done;
410
411         sge.addr = c->busa;
412         sge.length = client->msize;
413         sge.lkey = rdma->pd->local_dma_lkey;
414
415         wr.next = NULL;
416         wr.wr_cqe = &c->cqe;
417         wr.sg_list = &sge;
418         wr.num_sge = 1;
419         return ib_post_recv(rdma->qp, &wr, &bad_wr);
420
421  error:
422         p9_debug(P9_DEBUG_ERROR, "EIO\n");
423         return -EIO;
424 }
425
426 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
427 {
428         struct p9_trans_rdma *rdma = client->trans;
429         struct ib_send_wr wr, *bad_wr;
430         struct ib_sge sge;
431         int err = 0;
432         unsigned long flags;
433         struct p9_rdma_context *c = NULL;
434         struct p9_rdma_context *rpl_context = NULL;
435
436         /* When an error occurs between posting the recv and the send,
437          * there will be a receive context posted without a pending request.
438          * Since there is no way to "un-post" it, we remember it and skip
439          * post_recv() for the next request.
440          * So here,
441          * see if we are this `next request' and need to absorb an excess rc.
442          * If yes, then drop and free our own, and do not recv_post().
443          **/
444         if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
445                 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
446                         /* Got one ! */
447                         kfree(req->rc);
448                         req->rc = NULL;
449                         goto dont_need_post_recv;
450                 } else {
451                         /* We raced and lost. */
452                         atomic_inc(&rdma->excess_rc);
453                 }
454         }
455
456         /* Allocate an fcall for the reply */
457         rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
458         if (!rpl_context) {
459                 err = -ENOMEM;
460                 goto recv_error;
461         }
462         rpl_context->rc = req->rc;
463
464         /*
465          * Post a receive buffer for this request. We need to ensure
466          * there is a reply buffer available for every outstanding
467          * request. A flushed request can result in no reply for an
468          * outstanding request, so we must keep a count to avoid
469          * overflowing the RQ.
470          */
471         if (down_interruptible(&rdma->rq_sem)) {
472                 err = -EINTR;
473                 goto recv_error;
474         }
475
476         err = post_recv(client, rpl_context);
477         if (err) {
478                 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
479                 goto recv_error;
480         }
481         /* remove posted receive buffer from request structure */
482         req->rc = NULL;
483
484 dont_need_post_recv:
485         /* Post the request */
486         c = kmalloc(sizeof *c, GFP_NOFS);
487         if (!c) {
488                 err = -ENOMEM;
489                 goto send_error;
490         }
491         c->req = req;
492
493         c->busa = ib_dma_map_single(rdma->cm_id->device,
494                                     c->req->tc->sdata, c->req->tc->size,
495                                     DMA_TO_DEVICE);
496         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
497                 err = -EIO;
498                 goto send_error;
499         }
500
501         c->cqe.done = send_done;
502
503         sge.addr = c->busa;
504         sge.length = c->req->tc->size;
505         sge.lkey = rdma->pd->local_dma_lkey;
506
507         wr.next = NULL;
508         wr.wr_cqe = &c->cqe;
509         wr.opcode = IB_WR_SEND;
510         wr.send_flags = IB_SEND_SIGNALED;
511         wr.sg_list = &sge;
512         wr.num_sge = 1;
513
514         if (down_interruptible(&rdma->sq_sem)) {
515                 err = -EINTR;
516                 goto send_error;
517         }
518
519         /* Mark request as `sent' *before* we actually send it,
520          * because doing if after could erase the REQ_STATUS_RCVD
521          * status in case of a very fast reply.
522          */
523         req->status = REQ_STATUS_SENT;
524         err = ib_post_send(rdma->qp, &wr, &bad_wr);
525         if (err)
526                 goto send_error;
527
528         /* Success */
529         return 0;
530
531  /* Handle errors that happened during or while preparing the send: */
532  send_error:
533         req->status = REQ_STATUS_ERROR;
534         kfree(c);
535         p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
536
537         /* Ach.
538          *  We did recv_post(), but not send. We have one recv_post in excess.
539          */
540         atomic_inc(&rdma->excess_rc);
541         return err;
542
543  /* Handle errors that happened during or while preparing post_recv(): */
544  recv_error:
545         kfree(rpl_context);
546         spin_lock_irqsave(&rdma->req_lock, flags);
547         if (rdma->state < P9_RDMA_CLOSING) {
548                 rdma->state = P9_RDMA_CLOSING;
549                 spin_unlock_irqrestore(&rdma->req_lock, flags);
550                 rdma_disconnect(rdma->cm_id);
551         } else
552                 spin_unlock_irqrestore(&rdma->req_lock, flags);
553         return err;
554 }
555
556 static void rdma_close(struct p9_client *client)
557 {
558         struct p9_trans_rdma *rdma;
559
560         if (!client)
561                 return;
562
563         rdma = client->trans;
564         if (!rdma)
565                 return;
566
567         client->status = Disconnected;
568         rdma_disconnect(rdma->cm_id);
569         rdma_destroy_trans(rdma);
570 }
571
572 /**
573  * alloc_rdma - Allocate and initialize the rdma transport structure
574  * @opts: Mount options structure
575  */
576 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
577 {
578         struct p9_trans_rdma *rdma;
579
580         rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
581         if (!rdma)
582                 return NULL;
583
584         rdma->port = opts->port;
585         rdma->privport = opts->privport;
586         rdma->sq_depth = opts->sq_depth;
587         rdma->rq_depth = opts->rq_depth;
588         rdma->timeout = opts->timeout;
589         spin_lock_init(&rdma->req_lock);
590         init_completion(&rdma->cm_done);
591         sema_init(&rdma->sq_sem, rdma->sq_depth);
592         sema_init(&rdma->rq_sem, rdma->rq_depth);
593         atomic_set(&rdma->excess_rc, 0);
594
595         return rdma;
596 }
597
598 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
599 {
600         /* Nothing to do here.
601          * We will take care of it (if we have to) in rdma_cancelled()
602          */
603         return 1;
604 }
605
606 /* A request has been fully flushed without a reply.
607  * That means we have posted one buffer in excess.
608  */
609 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
610 {
611         struct p9_trans_rdma *rdma = client->trans;
612         atomic_inc(&rdma->excess_rc);
613         return 0;
614 }
615
616 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
617 {
618         struct sockaddr_in cl = {
619                 .sin_family = AF_INET,
620                 .sin_addr.s_addr = htonl(INADDR_ANY),
621         };
622         int port, err = -EINVAL;
623
624         for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
625                 cl.sin_port = htons((ushort)port);
626                 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
627                 if (err != -EADDRINUSE)
628                         break;
629         }
630         return err;
631 }
632
633 /**
634  * rdma_create_trans - Transport method for creating a transport instance
635  * @client: client instance
636  * @addr: IP address string
637  * @args: Mount options string
638  */
639 static int
640 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
641 {
642         int err;
643         struct p9_rdma_opts opts;
644         struct p9_trans_rdma *rdma;
645         struct rdma_conn_param conn_param;
646         struct ib_qp_init_attr qp_attr;
647
648         if (addr == NULL)
649                 return -EINVAL;
650
651         /* Parse the transport specific mount options */
652         err = parse_opts(args, &opts);
653         if (err < 0)
654                 return err;
655
656         /* Create and initialize the RDMA transport structure */
657         rdma = alloc_rdma(&opts);
658         if (!rdma)
659                 return -ENOMEM;
660
661         /* Create the RDMA CM ID */
662         rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
663                                      RDMA_PS_TCP, IB_QPT_RC);
664         if (IS_ERR(rdma->cm_id))
665                 goto error;
666
667         /* Associate the client with the transport */
668         client->trans = rdma;
669
670         /* Bind to a privileged port if we need to */
671         if (opts.privport) {
672                 err = p9_rdma_bind_privport(rdma);
673                 if (err < 0) {
674                         pr_err("%s (%d): problem binding to privport: %d\n",
675                                __func__, task_pid_nr(current), -err);
676                         goto error;
677                 }
678         }
679
680         /* Resolve the server's address */
681         rdma->addr.sin_family = AF_INET;
682         rdma->addr.sin_addr.s_addr = in_aton(addr);
683         rdma->addr.sin_port = htons(opts.port);
684         err = rdma_resolve_addr(rdma->cm_id, NULL,
685                                 (struct sockaddr *)&rdma->addr,
686                                 rdma->timeout);
687         if (err)
688                 goto error;
689         err = wait_for_completion_interruptible(&rdma->cm_done);
690         if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
691                 goto error;
692
693         /* Resolve the route to the server */
694         err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
695         if (err)
696                 goto error;
697         err = wait_for_completion_interruptible(&rdma->cm_done);
698         if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
699                 goto error;
700
701         /* Create the Completion Queue */
702         rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
703                         opts.sq_depth + opts.rq_depth + 1,
704                         0, IB_POLL_SOFTIRQ);
705         if (IS_ERR(rdma->cq))
706                 goto error;
707
708         /* Create the Protection Domain */
709         rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
710         if (IS_ERR(rdma->pd))
711                 goto error;
712
713         /* Create the Queue Pair */
714         memset(&qp_attr, 0, sizeof qp_attr);
715         qp_attr.event_handler = qp_event_handler;
716         qp_attr.qp_context = client;
717         qp_attr.cap.max_send_wr = opts.sq_depth;
718         qp_attr.cap.max_recv_wr = opts.rq_depth;
719         qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
720         qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
721         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
722         qp_attr.qp_type = IB_QPT_RC;
723         qp_attr.send_cq = rdma->cq;
724         qp_attr.recv_cq = rdma->cq;
725         err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
726         if (err)
727                 goto error;
728         rdma->qp = rdma->cm_id->qp;
729
730         /* Request a connection */
731         memset(&conn_param, 0, sizeof(conn_param));
732         conn_param.private_data = NULL;
733         conn_param.private_data_len = 0;
734         conn_param.responder_resources = P9_RDMA_IRD;
735         conn_param.initiator_depth = P9_RDMA_ORD;
736         err = rdma_connect(rdma->cm_id, &conn_param);
737         if (err)
738                 goto error;
739         err = wait_for_completion_interruptible(&rdma->cm_done);
740         if (err || (rdma->state != P9_RDMA_CONNECTED))
741                 goto error;
742
743         client->status = Connected;
744
745         return 0;
746
747 error:
748         rdma_destroy_trans(rdma);
749         return -ENOTCONN;
750 }
751
752 static struct p9_trans_module p9_rdma_trans = {
753         .name = "rdma",
754         .maxsize = P9_RDMA_MAXSIZE,
755         .def = 0,
756         .owner = THIS_MODULE,
757         .create = rdma_create_trans,
758         .close = rdma_close,
759         .request = rdma_request,
760         .cancel = rdma_cancel,
761         .cancelled = rdma_cancelled,
762         .show_options = p9_rdma_show_options,
763 };
764
765 /**
766  * p9_trans_rdma_init - Register the 9P RDMA transport driver
767  */
768 static int __init p9_trans_rdma_init(void)
769 {
770         v9fs_register_trans(&p9_rdma_trans);
771         return 0;
772 }
773
774 static void __exit p9_trans_rdma_exit(void)
775 {
776         v9fs_unregister_trans(&p9_rdma_trans);
777 }
778
779 module_init(p9_trans_rdma_init);
780 module_exit(p9_trans_rdma_exit);
781
782 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
783 MODULE_DESCRIPTION("RDMA Transport for 9P");
784 MODULE_LICENSE("Dual BSD/GPL");