Loading...
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 */
69struct 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
97struct p9_rdma_req;
98
99/**
100 * struct p9_rdma_context - Keeps track of in-process WR
101 *
102 * @cqe: completion queue entry
103 * @busa: Bus address to unmap when the WR completes
104 * @req: Keeps track of requests (send)
105 * @rc: Keepts track of replies (receive)
106 */
107struct p9_rdma_context {
108 struct ib_cqe cqe;
109 dma_addr_t busa;
110 union {
111 struct p9_req_t *req;
112 struct p9_fcall rc;
113 };
114};
115
116/**
117 * struct p9_rdma_opts - Collection of mount options
118 * @port: port of connection
119 * @privport: Whether a privileged port may be used
120 * @sq_depth: The requested depth of the SQ. This really doesn't need
121 * to be any deeper than the number of threads used in the client
122 * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
123 * @timeout: Time to wait in msecs for CM events
124 */
125struct p9_rdma_opts {
126 short port;
127 bool privport;
128 int sq_depth;
129 int rq_depth;
130 long timeout;
131};
132
133/*
134 * Option Parsing (code inspired by NFS code)
135 */
136enum {
137 /* Options that take integer arguments */
138 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
139 /* Options that take no argument */
140 Opt_privport,
141 Opt_err,
142};
143
144static match_table_t tokens = {
145 {Opt_port, "port=%u"},
146 {Opt_sq_depth, "sq=%u"},
147 {Opt_rq_depth, "rq=%u"},
148 {Opt_timeout, "timeout=%u"},
149 {Opt_privport, "privport"},
150 {Opt_err, NULL},
151};
152
153static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
154{
155 struct p9_trans_rdma *rdma = clnt->trans;
156
157 if (rdma->port != P9_PORT)
158 seq_printf(m, ",port=%u", rdma->port);
159 if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
160 seq_printf(m, ",sq=%u", rdma->sq_depth);
161 if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
162 seq_printf(m, ",rq=%u", rdma->rq_depth);
163 if (rdma->timeout != P9_RDMA_TIMEOUT)
164 seq_printf(m, ",timeout=%lu", rdma->timeout);
165 if (rdma->privport)
166 seq_puts(m, ",privport");
167 return 0;
168}
169
170/**
171 * parse_opts - parse mount options into rdma options structure
172 * @params: options string passed from mount
173 * @opts: rdma transport-specific structure to parse options into
174 *
175 * Returns 0 upon success, -ERRNO upon failure
176 */
177static int parse_opts(char *params, struct p9_rdma_opts *opts)
178{
179 char *p;
180 substring_t args[MAX_OPT_ARGS];
181 int option;
182 char *options, *tmp_options;
183
184 opts->port = P9_PORT;
185 opts->sq_depth = P9_RDMA_SQ_DEPTH;
186 opts->rq_depth = P9_RDMA_RQ_DEPTH;
187 opts->timeout = P9_RDMA_TIMEOUT;
188 opts->privport = false;
189
190 if (!params)
191 return 0;
192
193 tmp_options = kstrdup(params, GFP_KERNEL);
194 if (!tmp_options) {
195 p9_debug(P9_DEBUG_ERROR,
196 "failed to allocate copy of option string\n");
197 return -ENOMEM;
198 }
199 options = tmp_options;
200
201 while ((p = strsep(&options, ",")) != NULL) {
202 int token;
203 int r;
204 if (!*p)
205 continue;
206 token = match_token(p, tokens, args);
207 if ((token != Opt_err) && (token != Opt_privport)) {
208 r = match_int(&args[0], &option);
209 if (r < 0) {
210 p9_debug(P9_DEBUG_ERROR,
211 "integer field, but no integer?\n");
212 continue;
213 }
214 }
215 switch (token) {
216 case Opt_port:
217 opts->port = option;
218 break;
219 case Opt_sq_depth:
220 opts->sq_depth = option;
221 break;
222 case Opt_rq_depth:
223 opts->rq_depth = option;
224 break;
225 case Opt_timeout:
226 opts->timeout = option;
227 break;
228 case Opt_privport:
229 opts->privport = true;
230 break;
231 default:
232 continue;
233 }
234 }
235 /* RQ must be at least as large as the SQ */
236 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
237 kfree(tmp_options);
238 return 0;
239}
240
241static int
242p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
243{
244 struct p9_client *c = id->context;
245 struct p9_trans_rdma *rdma = c->trans;
246 switch (event->event) {
247 case RDMA_CM_EVENT_ADDR_RESOLVED:
248 BUG_ON(rdma->state != P9_RDMA_INIT);
249 rdma->state = P9_RDMA_ADDR_RESOLVED;
250 break;
251
252 case RDMA_CM_EVENT_ROUTE_RESOLVED:
253 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
254 rdma->state = P9_RDMA_ROUTE_RESOLVED;
255 break;
256
257 case RDMA_CM_EVENT_ESTABLISHED:
258 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
259 rdma->state = P9_RDMA_CONNECTED;
260 break;
261
262 case RDMA_CM_EVENT_DISCONNECTED:
263 if (rdma)
264 rdma->state = P9_RDMA_CLOSED;
265 c->status = Disconnected;
266 break;
267
268 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
269 break;
270
271 case RDMA_CM_EVENT_ADDR_CHANGE:
272 case RDMA_CM_EVENT_ROUTE_ERROR:
273 case RDMA_CM_EVENT_DEVICE_REMOVAL:
274 case RDMA_CM_EVENT_MULTICAST_JOIN:
275 case RDMA_CM_EVENT_MULTICAST_ERROR:
276 case RDMA_CM_EVENT_REJECTED:
277 case RDMA_CM_EVENT_CONNECT_REQUEST:
278 case RDMA_CM_EVENT_CONNECT_RESPONSE:
279 case RDMA_CM_EVENT_CONNECT_ERROR:
280 case RDMA_CM_EVENT_ADDR_ERROR:
281 case RDMA_CM_EVENT_UNREACHABLE:
282 c->status = Disconnected;
283 rdma_disconnect(rdma->cm_id);
284 break;
285 default:
286 BUG();
287 }
288 complete(&rdma->cm_done);
289 return 0;
290}
291
292static void
293recv_done(struct ib_cq *cq, struct ib_wc *wc)
294{
295 struct p9_client *client = cq->cq_context;
296 struct p9_trans_rdma *rdma = client->trans;
297 struct p9_rdma_context *c =
298 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
299 struct p9_req_t *req;
300 int err = 0;
301 int16_t tag;
302
303 req = NULL;
304 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
305 DMA_FROM_DEVICE);
306
307 if (wc->status != IB_WC_SUCCESS)
308 goto err_out;
309
310 c->rc.size = wc->byte_len;
311 err = p9_parse_header(&c->rc, NULL, NULL, &tag, 1);
312 if (err)
313 goto err_out;
314
315 req = p9_tag_lookup(client, tag);
316 if (!req)
317 goto err_out;
318
319 /* Check that we have not yet received a reply for this request.
320 */
321 if (unlikely(req->rc.sdata)) {
322 pr_err("Duplicate reply for request %d", tag);
323 goto err_out;
324 }
325
326 req->rc.size = c->rc.size;
327 req->rc.sdata = c->rc.sdata;
328 p9_client_cb(client, req, REQ_STATUS_RCVD);
329
330 out:
331 up(&rdma->rq_sem);
332 kfree(c);
333 return;
334
335 err_out:
336 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
337 req, err, wc->status);
338 rdma->state = P9_RDMA_FLUSHING;
339 client->status = Disconnected;
340 goto out;
341}
342
343static void
344send_done(struct ib_cq *cq, struct ib_wc *wc)
345{
346 struct p9_client *client = cq->cq_context;
347 struct p9_trans_rdma *rdma = client->trans;
348 struct p9_rdma_context *c =
349 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
350
351 ib_dma_unmap_single(rdma->cm_id->device,
352 c->busa, c->req->tc.size,
353 DMA_TO_DEVICE);
354 up(&rdma->sq_sem);
355 p9_req_put(c->req);
356 kfree(c);
357}
358
359static void qp_event_handler(struct ib_event *event, void *context)
360{
361 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
362 event->event, context);
363}
364
365static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
366{
367 if (!rdma)
368 return;
369
370 if (rdma->qp && !IS_ERR(rdma->qp))
371 ib_destroy_qp(rdma->qp);
372
373 if (rdma->pd && !IS_ERR(rdma->pd))
374 ib_dealloc_pd(rdma->pd);
375
376 if (rdma->cq && !IS_ERR(rdma->cq))
377 ib_free_cq(rdma->cq);
378
379 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
380 rdma_destroy_id(rdma->cm_id);
381
382 kfree(rdma);
383}
384
385static int
386post_recv(struct p9_client *client, struct p9_rdma_context *c)
387{
388 struct p9_trans_rdma *rdma = client->trans;
389 struct ib_recv_wr wr;
390 struct ib_sge sge;
391
392 c->busa = ib_dma_map_single(rdma->cm_id->device,
393 c->rc.sdata, client->msize,
394 DMA_FROM_DEVICE);
395 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
396 goto error;
397
398 c->cqe.done = recv_done;
399
400 sge.addr = c->busa;
401 sge.length = client->msize;
402 sge.lkey = rdma->pd->local_dma_lkey;
403
404 wr.next = NULL;
405 wr.wr_cqe = &c->cqe;
406 wr.sg_list = &sge;
407 wr.num_sge = 1;
408 return ib_post_recv(rdma->qp, &wr, NULL);
409
410 error:
411 p9_debug(P9_DEBUG_ERROR, "EIO\n");
412 return -EIO;
413}
414
415static int rdma_request(struct p9_client *client, struct p9_req_t *req)
416{
417 struct p9_trans_rdma *rdma = client->trans;
418 struct ib_send_wr wr;
419 struct ib_sge sge;
420 int err = 0;
421 unsigned long flags;
422 struct p9_rdma_context *c = NULL;
423 struct p9_rdma_context *rpl_context = NULL;
424
425 /* When an error occurs between posting the recv and the send,
426 * there will be a receive context posted without a pending request.
427 * Since there is no way to "un-post" it, we remember it and skip
428 * post_recv() for the next request.
429 * So here,
430 * see if we are this `next request' and need to absorb an excess rc.
431 * If yes, then drop and free our own, and do not recv_post().
432 **/
433 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
434 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
435 /* Got one! */
436 p9_fcall_fini(&req->rc);
437 req->rc.sdata = NULL;
438 goto dont_need_post_recv;
439 } else {
440 /* We raced and lost. */
441 atomic_inc(&rdma->excess_rc);
442 }
443 }
444
445 /* Allocate an fcall for the reply */
446 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
447 if (!rpl_context) {
448 err = -ENOMEM;
449 goto recv_error;
450 }
451 rpl_context->rc.sdata = req->rc.sdata;
452
453 /*
454 * Post a receive buffer for this request. We need to ensure
455 * there is a reply buffer available for every outstanding
456 * request. A flushed request can result in no reply for an
457 * outstanding request, so we must keep a count to avoid
458 * overflowing the RQ.
459 */
460 if (down_interruptible(&rdma->rq_sem)) {
461 err = -EINTR;
462 goto recv_error;
463 }
464
465 err = post_recv(client, rpl_context);
466 if (err) {
467 p9_debug(P9_DEBUG_ERROR, "POST RECV failed: %d\n", err);
468 goto recv_error;
469 }
470 /* remove posted receive buffer from request structure */
471 req->rc.sdata = NULL;
472
473dont_need_post_recv:
474 /* Post the request */
475 c = kmalloc(sizeof *c, GFP_NOFS);
476 if (!c) {
477 err = -ENOMEM;
478 goto send_error;
479 }
480 c->req = req;
481
482 c->busa = ib_dma_map_single(rdma->cm_id->device,
483 c->req->tc.sdata, c->req->tc.size,
484 DMA_TO_DEVICE);
485 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
486 err = -EIO;
487 goto send_error;
488 }
489
490 c->cqe.done = send_done;
491
492 sge.addr = c->busa;
493 sge.length = c->req->tc.size;
494 sge.lkey = rdma->pd->local_dma_lkey;
495
496 wr.next = NULL;
497 wr.wr_cqe = &c->cqe;
498 wr.opcode = IB_WR_SEND;
499 wr.send_flags = IB_SEND_SIGNALED;
500 wr.sg_list = &sge;
501 wr.num_sge = 1;
502
503 if (down_interruptible(&rdma->sq_sem)) {
504 err = -EINTR;
505 goto send_error;
506 }
507
508 /* Mark request as `sent' *before* we actually send it,
509 * because doing if after could erase the REQ_STATUS_RCVD
510 * status in case of a very fast reply.
511 */
512 req->status = REQ_STATUS_SENT;
513 err = ib_post_send(rdma->qp, &wr, NULL);
514 if (err)
515 goto send_error;
516
517 /* Success */
518 return 0;
519
520 /* Handle errors that happened during or while preparing the send: */
521 send_error:
522 req->status = REQ_STATUS_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 (err != -EINTR && 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
545static 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 */
565static 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->port = opts->port;
574 rdma->privport = opts->privport;
575 rdma->sq_depth = opts->sq_depth;
576 rdma->rq_depth = opts->rq_depth;
577 rdma->timeout = opts->timeout;
578 spin_lock_init(&rdma->req_lock);
579 init_completion(&rdma->cm_done);
580 sema_init(&rdma->sq_sem, rdma->sq_depth);
581 sema_init(&rdma->rq_sem, rdma->rq_depth);
582 atomic_set(&rdma->excess_rc, 0);
583
584 return rdma;
585}
586
587static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
588{
589 /* Nothing to do here.
590 * We will take care of it (if we have to) in rdma_cancelled()
591 */
592 return 1;
593}
594
595/* A request has been fully flushed without a reply.
596 * That means we have posted one buffer in excess.
597 */
598static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
599{
600 struct p9_trans_rdma *rdma = client->trans;
601 atomic_inc(&rdma->excess_rc);
602 return 0;
603}
604
605static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
606{
607 struct sockaddr_in cl = {
608 .sin_family = AF_INET,
609 .sin_addr.s_addr = htonl(INADDR_ANY),
610 };
611 int port, err = -EINVAL;
612
613 for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
614 cl.sin_port = htons((ushort)port);
615 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
616 if (err != -EADDRINUSE)
617 break;
618 }
619 return err;
620}
621
622/**
623 * rdma_create_trans - Transport method for creating a transport instance
624 * @client: client instance
625 * @addr: IP address string
626 * @args: Mount options string
627 */
628static int
629rdma_create_trans(struct p9_client *client, const char *addr, char *args)
630{
631 int err;
632 struct p9_rdma_opts opts;
633 struct p9_trans_rdma *rdma;
634 struct rdma_conn_param conn_param;
635 struct ib_qp_init_attr qp_attr;
636
637 if (addr == NULL)
638 return -EINVAL;
639
640 /* Parse the transport specific mount options */
641 err = parse_opts(args, &opts);
642 if (err < 0)
643 return err;
644
645 /* Create and initialize the RDMA transport structure */
646 rdma = alloc_rdma(&opts);
647 if (!rdma)
648 return -ENOMEM;
649
650 /* Create the RDMA CM ID */
651 rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
652 RDMA_PS_TCP, IB_QPT_RC);
653 if (IS_ERR(rdma->cm_id))
654 goto error;
655
656 /* Associate the client with the transport */
657 client->trans = rdma;
658
659 /* Bind to a privileged port if we need to */
660 if (opts.privport) {
661 err = p9_rdma_bind_privport(rdma);
662 if (err < 0) {
663 pr_err("%s (%d): problem binding to privport: %d\n",
664 __func__, task_pid_nr(current), -err);
665 goto error;
666 }
667 }
668
669 /* Resolve the server's address */
670 rdma->addr.sin_family = AF_INET;
671 rdma->addr.sin_addr.s_addr = in_aton(addr);
672 rdma->addr.sin_port = htons(opts.port);
673 err = rdma_resolve_addr(rdma->cm_id, NULL,
674 (struct sockaddr *)&rdma->addr,
675 rdma->timeout);
676 if (err)
677 goto error;
678 err = wait_for_completion_interruptible(&rdma->cm_done);
679 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
680 goto error;
681
682 /* Resolve the route to the server */
683 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
684 if (err)
685 goto error;
686 err = wait_for_completion_interruptible(&rdma->cm_done);
687 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
688 goto error;
689
690 /* Create the Completion Queue */
691 rdma->cq = ib_alloc_cq_any(rdma->cm_id->device, client,
692 opts.sq_depth + opts.rq_depth + 1,
693 IB_POLL_SOFTIRQ);
694 if (IS_ERR(rdma->cq))
695 goto error;
696
697 /* Create the Protection Domain */
698 rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
699 if (IS_ERR(rdma->pd))
700 goto error;
701
702 /* Create the Queue Pair */
703 memset(&qp_attr, 0, sizeof qp_attr);
704 qp_attr.event_handler = qp_event_handler;
705 qp_attr.qp_context = client;
706 qp_attr.cap.max_send_wr = opts.sq_depth;
707 qp_attr.cap.max_recv_wr = opts.rq_depth;
708 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
709 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
710 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
711 qp_attr.qp_type = IB_QPT_RC;
712 qp_attr.send_cq = rdma->cq;
713 qp_attr.recv_cq = rdma->cq;
714 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
715 if (err)
716 goto error;
717 rdma->qp = rdma->cm_id->qp;
718
719 /* Request a connection */
720 memset(&conn_param, 0, sizeof(conn_param));
721 conn_param.private_data = NULL;
722 conn_param.private_data_len = 0;
723 conn_param.responder_resources = P9_RDMA_IRD;
724 conn_param.initiator_depth = P9_RDMA_ORD;
725 err = rdma_connect(rdma->cm_id, &conn_param);
726 if (err)
727 goto error;
728 err = wait_for_completion_interruptible(&rdma->cm_done);
729 if (err || (rdma->state != P9_RDMA_CONNECTED))
730 goto error;
731
732 client->status = Connected;
733
734 return 0;
735
736error:
737 rdma_destroy_trans(rdma);
738 return -ENOTCONN;
739}
740
741static struct p9_trans_module p9_rdma_trans = {
742 .name = "rdma",
743 .maxsize = P9_RDMA_MAXSIZE,
744 .def = 0,
745 .owner = THIS_MODULE,
746 .create = rdma_create_trans,
747 .close = rdma_close,
748 .request = rdma_request,
749 .cancel = rdma_cancel,
750 .cancelled = rdma_cancelled,
751 .show_options = p9_rdma_show_options,
752};
753
754/**
755 * p9_trans_rdma_init - Register the 9P RDMA transport driver
756 */
757static int __init p9_trans_rdma_init(void)
758{
759 v9fs_register_trans(&p9_rdma_trans);
760 return 0;
761}
762
763static void __exit p9_trans_rdma_exit(void)
764{
765 v9fs_unregister_trans(&p9_rdma_trans);
766}
767
768module_init(p9_trans_rdma_init);
769module_exit(p9_trans_rdma_exit);
770
771MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
772MODULE_DESCRIPTION("RDMA Transport for 9P");
773MODULE_LICENSE("Dual BSD/GPL");
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 <linux/slab.h>
44#include <net/9p/9p.h>
45#include <net/9p/client.h>
46#include <net/9p/transport.h>
47#include <rdma/ib_verbs.h>
48#include <rdma/rdma_cm.h>
49
50#define P9_PORT 5640
51#define P9_RDMA_SQ_DEPTH 32
52#define P9_RDMA_RQ_DEPTH 32
53#define P9_RDMA_SEND_SGE 4
54#define P9_RDMA_RECV_SGE 4
55#define P9_RDMA_IRD 0
56#define P9_RDMA_ORD 0
57#define P9_RDMA_TIMEOUT 30000 /* 30 seconds */
58#define P9_RDMA_MAXSIZE (4*4096) /* Min SGE is 4, so we can
59 * safely advertise a maxsize
60 * of 64k */
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_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 */
81struct 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 */
116struct p9_rdma_req;
117struct 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 */
134struct 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 */
144enum {
145 /* Options that take integer arguments */
146 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
147};
148
149static 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 */
164static 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
171 opts->port = P9_PORT;
172 opts->sq_depth = P9_RDMA_SQ_DEPTH;
173 opts->rq_depth = P9_RDMA_RQ_DEPTH;
174 opts->timeout = P9_RDMA_TIMEOUT;
175
176 if (!params)
177 return 0;
178
179 tmp_options = kstrdup(params, GFP_KERNEL);
180 if (!tmp_options) {
181 P9_DPRINTK(P9_DEBUG_ERROR,
182 "failed to allocate copy of option string\n");
183 return -ENOMEM;
184 }
185 options = tmp_options;
186
187 while ((p = strsep(&options, ",")) != NULL) {
188 int token;
189 int r;
190 if (!*p)
191 continue;
192 token = match_token(p, tokens, args);
193 r = match_int(&args[0], &option);
194 if (r < 0) {
195 P9_DPRINTK(P9_DEBUG_ERROR,
196 "integer field, but no integer?\n");
197 continue;
198 }
199 switch (token) {
200 case Opt_port:
201 opts->port = option;
202 break;
203 case Opt_sq_depth:
204 opts->sq_depth = option;
205 break;
206 case Opt_rq_depth:
207 opts->rq_depth = option;
208 break;
209 case Opt_timeout:
210 opts->timeout = option;
211 break;
212 default:
213 continue;
214 }
215 }
216 /* RQ must be at least as large as the SQ */
217 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
218 kfree(tmp_options);
219 return 0;
220}
221
222static int
223p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
224{
225 struct p9_client *c = id->context;
226 struct p9_trans_rdma *rdma = c->trans;
227 switch (event->event) {
228 case RDMA_CM_EVENT_ADDR_RESOLVED:
229 BUG_ON(rdma->state != P9_RDMA_INIT);
230 rdma->state = P9_RDMA_ADDR_RESOLVED;
231 break;
232
233 case RDMA_CM_EVENT_ROUTE_RESOLVED:
234 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
235 rdma->state = P9_RDMA_ROUTE_RESOLVED;
236 break;
237
238 case RDMA_CM_EVENT_ESTABLISHED:
239 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
240 rdma->state = P9_RDMA_CONNECTED;
241 break;
242
243 case RDMA_CM_EVENT_DISCONNECTED:
244 if (rdma)
245 rdma->state = P9_RDMA_CLOSED;
246 if (c)
247 c->status = Disconnected;
248 break;
249
250 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
251 break;
252
253 case RDMA_CM_EVENT_ADDR_CHANGE:
254 case RDMA_CM_EVENT_ROUTE_ERROR:
255 case RDMA_CM_EVENT_DEVICE_REMOVAL:
256 case RDMA_CM_EVENT_MULTICAST_JOIN:
257 case RDMA_CM_EVENT_MULTICAST_ERROR:
258 case RDMA_CM_EVENT_REJECTED:
259 case RDMA_CM_EVENT_CONNECT_REQUEST:
260 case RDMA_CM_EVENT_CONNECT_RESPONSE:
261 case RDMA_CM_EVENT_CONNECT_ERROR:
262 case RDMA_CM_EVENT_ADDR_ERROR:
263 case RDMA_CM_EVENT_UNREACHABLE:
264 c->status = Disconnected;
265 rdma_disconnect(rdma->cm_id);
266 break;
267 default:
268 BUG();
269 }
270 complete(&rdma->cm_done);
271 return 0;
272}
273
274static void
275handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
276 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
277{
278 struct p9_req_t *req;
279 int err = 0;
280 int16_t tag;
281
282 req = NULL;
283 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
284 DMA_FROM_DEVICE);
285
286 if (status != IB_WC_SUCCESS)
287 goto err_out;
288
289 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
290 if (err)
291 goto err_out;
292
293 req = p9_tag_lookup(client, tag);
294 if (!req)
295 goto err_out;
296
297 req->rc = c->rc;
298 req->status = REQ_STATUS_RCVD;
299 p9_client_cb(client, req);
300
301 return;
302
303 err_out:
304 P9_DPRINTK(P9_DEBUG_ERROR, "req %p err %d status %d\n",
305 req, err, status);
306 rdma->state = P9_RDMA_FLUSHING;
307 client->status = Disconnected;
308}
309
310static void
311handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
312 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
313{
314 ib_dma_unmap_single(rdma->cm_id->device,
315 c->busa, c->req->tc->size,
316 DMA_TO_DEVICE);
317}
318
319static void qp_event_handler(struct ib_event *event, void *context)
320{
321 P9_DPRINTK(P9_DEBUG_ERROR, "QP event %d context %p\n", event->event,
322 context);
323}
324
325static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
326{
327 struct p9_client *client = cq_context;
328 struct p9_trans_rdma *rdma = client->trans;
329 int ret;
330 struct ib_wc wc;
331
332 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
333 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
334 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
335
336 switch (c->wc_op) {
337 case IB_WC_RECV:
338 atomic_dec(&rdma->rq_count);
339 handle_recv(client, rdma, c, wc.status, wc.byte_len);
340 break;
341
342 case IB_WC_SEND:
343 handle_send(client, rdma, c, wc.status, wc.byte_len);
344 up(&rdma->sq_sem);
345 break;
346
347 default:
348 printk(KERN_ERR "9prdma: unexpected completion type, "
349 "c->wc_op=%d, wc.opcode=%d, status=%d\n",
350 c->wc_op, wc.opcode, wc.status);
351 break;
352 }
353 kfree(c);
354 }
355}
356
357static void cq_event_handler(struct ib_event *e, void *v)
358{
359 P9_DPRINTK(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
360}
361
362static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
363{
364 if (!rdma)
365 return;
366
367 if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
368 ib_dereg_mr(rdma->dma_mr);
369
370 if (rdma->qp && !IS_ERR(rdma->qp))
371 ib_destroy_qp(rdma->qp);
372
373 if (rdma->pd && !IS_ERR(rdma->pd))
374 ib_dealloc_pd(rdma->pd);
375
376 if (rdma->cq && !IS_ERR(rdma->cq))
377 ib_destroy_cq(rdma->cq);
378
379 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
380 rdma_destroy_id(rdma->cm_id);
381
382 kfree(rdma);
383}
384
385static int
386post_recv(struct p9_client *client, struct p9_rdma_context *c)
387{
388 struct p9_trans_rdma *rdma = client->trans;
389 struct ib_recv_wr wr, *bad_wr;
390 struct ib_sge sge;
391
392 c->busa = ib_dma_map_single(rdma->cm_id->device,
393 c->rc->sdata, client->msize,
394 DMA_FROM_DEVICE);
395 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
396 goto error;
397
398 sge.addr = c->busa;
399 sge.length = client->msize;
400 sge.lkey = rdma->lkey;
401
402 wr.next = NULL;
403 c->wc_op = IB_WC_RECV;
404 wr.wr_id = (unsigned long) c;
405 wr.sg_list = &sge;
406 wr.num_sge = 1;
407 return ib_post_recv(rdma->qp, &wr, &bad_wr);
408
409 error:
410 P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
411 return -EIO;
412}
413
414static int rdma_request(struct p9_client *client, struct p9_req_t *req)
415{
416 struct p9_trans_rdma *rdma = client->trans;
417 struct ib_send_wr wr, *bad_wr;
418 struct ib_sge sge;
419 int err = 0;
420 unsigned long flags;
421 struct p9_rdma_context *c = NULL;
422 struct p9_rdma_context *rpl_context = NULL;
423
424 /* Allocate an fcall for the reply */
425 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
426 if (!rpl_context) {
427 err = -ENOMEM;
428 goto err_close;
429 }
430
431 /*
432 * If the request has a buffer, steal it, otherwise
433 * allocate a new one. Typically, requests should already
434 * have receive buffers allocated and just swap them around
435 */
436 if (!req->rc) {
437 req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
438 GFP_NOFS);
439 if (req->rc) {
440 req->rc->sdata = (char *) req->rc +
441 sizeof(struct p9_fcall);
442 req->rc->capacity = client->msize;
443 }
444 }
445 rpl_context->rc = req->rc;
446 if (!rpl_context->rc) {
447 err = -ENOMEM;
448 goto err_free2;
449 }
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 (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
459 err = post_recv(client, rpl_context);
460 if (err)
461 goto err_free1;
462 } else
463 atomic_dec(&rdma->rq_count);
464
465 /* remove posted receive buffer from request structure */
466 req->rc = NULL;
467
468 /* Post the request */
469 c = kmalloc(sizeof *c, GFP_NOFS);
470 if (!c) {
471 err = -ENOMEM;
472 goto err_free1;
473 }
474 c->req = req;
475
476 c->busa = ib_dma_map_single(rdma->cm_id->device,
477 c->req->tc->sdata, c->req->tc->size,
478 DMA_TO_DEVICE);
479 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
480 goto error;
481
482 sge.addr = c->busa;
483 sge.length = c->req->tc->size;
484 sge.lkey = rdma->lkey;
485
486 wr.next = NULL;
487 c->wc_op = IB_WC_SEND;
488 wr.wr_id = (unsigned long) c;
489 wr.opcode = IB_WR_SEND;
490 wr.send_flags = IB_SEND_SIGNALED;
491 wr.sg_list = &sge;
492 wr.num_sge = 1;
493
494 if (down_interruptible(&rdma->sq_sem))
495 goto error;
496
497 return ib_post_send(rdma->qp, &wr, &bad_wr);
498
499 error:
500 kfree(c);
501 kfree(rpl_context->rc);
502 kfree(rpl_context);
503 P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
504 return -EIO;
505 err_free1:
506 kfree(rpl_context->rc);
507 err_free2:
508 kfree(rpl_context);
509 err_close:
510 spin_lock_irqsave(&rdma->req_lock, flags);
511 if (rdma->state < P9_RDMA_CLOSING) {
512 rdma->state = P9_RDMA_CLOSING;
513 spin_unlock_irqrestore(&rdma->req_lock, flags);
514 rdma_disconnect(rdma->cm_id);
515 } else
516 spin_unlock_irqrestore(&rdma->req_lock, flags);
517 return err;
518}
519
520static void rdma_close(struct p9_client *client)
521{
522 struct p9_trans_rdma *rdma;
523
524 if (!client)
525 return;
526
527 rdma = client->trans;
528 if (!rdma)
529 return;
530
531 client->status = Disconnected;
532 rdma_disconnect(rdma->cm_id);
533 rdma_destroy_trans(rdma);
534}
535
536/**
537 * alloc_rdma - Allocate and initialize the rdma transport structure
538 * @opts: Mount options structure
539 */
540static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
541{
542 struct p9_trans_rdma *rdma;
543
544 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
545 if (!rdma)
546 return NULL;
547
548 rdma->sq_depth = opts->sq_depth;
549 rdma->rq_depth = opts->rq_depth;
550 rdma->timeout = opts->timeout;
551 spin_lock_init(&rdma->req_lock);
552 init_completion(&rdma->cm_done);
553 sema_init(&rdma->sq_sem, rdma->sq_depth);
554 atomic_set(&rdma->rq_count, 0);
555
556 return rdma;
557}
558
559/* its not clear to me we can do anything after send has been posted */
560static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
561{
562 return 1;
563}
564
565/**
566 * trans_create_rdma - Transport method for creating atransport instance
567 * @client: client instance
568 * @addr: IP address string
569 * @args: Mount options string
570 */
571static int
572rdma_create_trans(struct p9_client *client, const char *addr, char *args)
573{
574 int err;
575 struct p9_rdma_opts opts;
576 struct p9_trans_rdma *rdma;
577 struct rdma_conn_param conn_param;
578 struct ib_qp_init_attr qp_attr;
579 struct ib_device_attr devattr;
580
581 /* Parse the transport specific mount options */
582 err = parse_opts(args, &opts);
583 if (err < 0)
584 return err;
585
586 /* Create and initialize the RDMA transport structure */
587 rdma = alloc_rdma(&opts);
588 if (!rdma)
589 return -ENOMEM;
590
591 /* Create the RDMA CM ID */
592 rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
593 IB_QPT_RC);
594 if (IS_ERR(rdma->cm_id))
595 goto error;
596
597 /* Associate the client with the transport */
598 client->trans = rdma;
599
600 /* Resolve the server's address */
601 rdma->addr.sin_family = AF_INET;
602 rdma->addr.sin_addr.s_addr = in_aton(addr);
603 rdma->addr.sin_port = htons(opts.port);
604 err = rdma_resolve_addr(rdma->cm_id, NULL,
605 (struct sockaddr *)&rdma->addr,
606 rdma->timeout);
607 if (err)
608 goto error;
609 err = wait_for_completion_interruptible(&rdma->cm_done);
610 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
611 goto error;
612
613 /* Resolve the route to the server */
614 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
615 if (err)
616 goto error;
617 err = wait_for_completion_interruptible(&rdma->cm_done);
618 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
619 goto error;
620
621 /* Query the device attributes */
622 err = ib_query_device(rdma->cm_id->device, &devattr);
623 if (err)
624 goto error;
625
626 /* Create the Completion Queue */
627 rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
628 cq_event_handler, client,
629 opts.sq_depth + opts.rq_depth + 1, 0);
630 if (IS_ERR(rdma->cq))
631 goto error;
632 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
633
634 /* Create the Protection Domain */
635 rdma->pd = ib_alloc_pd(rdma->cm_id->device);
636 if (IS_ERR(rdma->pd))
637 goto error;
638
639 /* Cache the DMA lkey in the transport */
640 rdma->dma_mr = NULL;
641 if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
642 rdma->lkey = rdma->cm_id->device->local_dma_lkey;
643 else {
644 rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
645 if (IS_ERR(rdma->dma_mr))
646 goto error;
647 rdma->lkey = rdma->dma_mr->lkey;
648 }
649
650 /* Create the Queue Pair */
651 memset(&qp_attr, 0, sizeof qp_attr);
652 qp_attr.event_handler = qp_event_handler;
653 qp_attr.qp_context = client;
654 qp_attr.cap.max_send_wr = opts.sq_depth;
655 qp_attr.cap.max_recv_wr = opts.rq_depth;
656 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
657 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
658 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
659 qp_attr.qp_type = IB_QPT_RC;
660 qp_attr.send_cq = rdma->cq;
661 qp_attr.recv_cq = rdma->cq;
662 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
663 if (err)
664 goto error;
665 rdma->qp = rdma->cm_id->qp;
666
667 /* Request a connection */
668 memset(&conn_param, 0, sizeof(conn_param));
669 conn_param.private_data = NULL;
670 conn_param.private_data_len = 0;
671 conn_param.responder_resources = P9_RDMA_IRD;
672 conn_param.initiator_depth = P9_RDMA_ORD;
673 err = rdma_connect(rdma->cm_id, &conn_param);
674 if (err)
675 goto error;
676 err = wait_for_completion_interruptible(&rdma->cm_done);
677 if (err || (rdma->state != P9_RDMA_CONNECTED))
678 goto error;
679
680 client->status = Connected;
681
682 return 0;
683
684error:
685 rdma_destroy_trans(rdma);
686 return -ENOTCONN;
687}
688
689static struct p9_trans_module p9_rdma_trans = {
690 .name = "rdma",
691 .maxsize = P9_RDMA_MAXSIZE,
692 .def = 0,
693 .owner = THIS_MODULE,
694 .create = rdma_create_trans,
695 .close = rdma_close,
696 .request = rdma_request,
697 .cancel = rdma_cancel,
698};
699
700/**
701 * p9_trans_rdma_init - Register the 9P RDMA transport driver
702 */
703static int __init p9_trans_rdma_init(void)
704{
705 v9fs_register_trans(&p9_rdma_trans);
706 return 0;
707}
708
709static void __exit p9_trans_rdma_exit(void)
710{
711 v9fs_unregister_trans(&p9_rdma_trans);
712}
713
714module_init(p9_trans_rdma_init);
715module_exit(p9_trans_rdma_exit);
716
717MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
718MODULE_DESCRIPTION("RDMA Transport for 9P");
719MODULE_LICENSE("Dual BSD/GPL");