blob: 2c69ddd691a16a0056ed071695fb3ebc6293a917 [file] [log] [blame]
Tom Tuckerfc79d4b2008-10-22 18:47:39 -05001/*
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
Joe Perches5d385152011-11-28 10:40:46 -080029#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
30
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050031#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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090045#include <linux/slab.h>
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050046#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>
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050051
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 (4*4096) /* Min SGE is 4, so we can
61 * safely advertise a maxsize
62 * of 64k */
63
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050064/**
65 * struct p9_trans_rdma - RDMA transport instance
66 *
67 * @state: tracks the transport state machine for connection setup and tear down
68 * @cm_id: The RDMA CM ID
69 * @pd: Protection Domain pointer
70 * @qp: Queue Pair pointer
71 * @cq: Completion Queue pointer
Abhishek Kulkarni0e155972009-07-19 13:41:55 -060072 * @dm_mr: DMA Memory Region pointer
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050073 * @lkey: The local access only memory region key
74 * @timeout: Number of uSecs to wait for connection management events
75 * @sq_depth: The depth of the Send Queue
76 * @sq_sem: Semaphore for the SQ
77 * @rq_depth: The depth of the Receive Queue.
Abhishek Kulkarni0e155972009-07-19 13:41:55 -060078 * @rq_count: Count of requests in the Receive Queue.
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050079 * @addr: The remote peer's address
80 * @req_lock: Protects the active request list
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050081 * @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 atomic_t rq_count;
104 struct sockaddr_in addr;
105 spinlock_t req_lock;
106
107 struct completion cm_done;
108};
109
110/**
111 * p9_rdma_context - Keeps track of in-process WR
112 *
113 * @wc_op: The original WR op for when the CQE completes in error.
114 * @busa: Bus address to unmap when the WR completes
115 * @req: Keeps track of requests (send)
116 * @rc: Keepts track of replies (receive)
117 */
118struct p9_rdma_req;
119struct p9_rdma_context {
120 enum ib_wc_opcode wc_op;
121 dma_addr_t busa;
122 union {
123 struct p9_req_t *req;
124 struct p9_fcall *rc;
125 };
126};
127
128/**
129 * p9_rdma_opts - Collection of mount options
130 * @port: port of connection
131 * @sq_depth: The requested depth of the SQ. This really doesn't need
132 * to be any deeper than the number of threads used in the client
133 * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
134 * @timeout: Time to wait in msecs for CM events
135 */
136struct p9_rdma_opts {
137 short port;
138 int sq_depth;
139 int rq_depth;
140 long timeout;
141};
142
143/*
144 * Option Parsing (code inspired by NFS code)
145 */
146enum {
147 /* Options that take integer arguments */
148 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
149};
150
151static match_table_t tokens = {
152 {Opt_port, "port=%u"},
153 {Opt_sq_depth, "sq=%u"},
154 {Opt_rq_depth, "rq=%u"},
155 {Opt_timeout, "timeout=%u"},
156 {Opt_err, NULL},
157};
158
159/**
Abhishek Kulkarni0e155972009-07-19 13:41:55 -0600160 * parse_opts - parse mount options into rdma options structure
161 * @params: options string passed from mount
162 * @opts: rdma transport-specific structure to parse options into
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500163 *
164 * Returns 0 upon success, -ERRNO upon failure
165 */
166static int parse_opts(char *params, struct p9_rdma_opts *opts)
167{
168 char *p;
169 substring_t args[MAX_OPT_ARGS];
170 int option;
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600171 char *options, *tmp_options;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500172
173 opts->port = P9_PORT;
174 opts->sq_depth = P9_RDMA_SQ_DEPTH;
175 opts->rq_depth = P9_RDMA_RQ_DEPTH;
176 opts->timeout = P9_RDMA_TIMEOUT;
177
178 if (!params)
179 return 0;
180
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600181 tmp_options = kstrdup(params, GFP_KERNEL);
182 if (!tmp_options) {
Joe Perches5d385152011-11-28 10:40:46 -0800183 p9_debug(P9_DEBUG_ERROR,
184 "failed to allocate copy of option string\n");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500185 return -ENOMEM;
186 }
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600187 options = tmp_options;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500188
189 while ((p = strsep(&options, ",")) != NULL) {
190 int token;
191 int r;
192 if (!*p)
193 continue;
194 token = match_token(p, tokens, args);
195 r = match_int(&args[0], &option);
196 if (r < 0) {
Joe Perches5d385152011-11-28 10:40:46 -0800197 p9_debug(P9_DEBUG_ERROR,
198 "integer field, but no integer?\n");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500199 continue;
200 }
201 switch (token) {
202 case Opt_port:
203 opts->port = option;
204 break;
205 case Opt_sq_depth:
206 opts->sq_depth = option;
207 break;
208 case Opt_rq_depth:
209 opts->rq_depth = option;
210 break;
211 case Opt_timeout:
212 opts->timeout = option;
213 break;
214 default:
215 continue;
216 }
217 }
218 /* RQ must be at least as large as the SQ */
219 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600220 kfree(tmp_options);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500221 return 0;
222}
223
224static int
225p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
226{
227 struct p9_client *c = id->context;
228 struct p9_trans_rdma *rdma = c->trans;
229 switch (event->event) {
230 case RDMA_CM_EVENT_ADDR_RESOLVED:
231 BUG_ON(rdma->state != P9_RDMA_INIT);
232 rdma->state = P9_RDMA_ADDR_RESOLVED;
233 break;
234
235 case RDMA_CM_EVENT_ROUTE_RESOLVED:
236 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
237 rdma->state = P9_RDMA_ROUTE_RESOLVED;
238 break;
239
240 case RDMA_CM_EVENT_ESTABLISHED:
241 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
242 rdma->state = P9_RDMA_CONNECTED;
243 break;
244
245 case RDMA_CM_EVENT_DISCONNECTED:
246 if (rdma)
247 rdma->state = P9_RDMA_CLOSED;
248 if (c)
249 c->status = Disconnected;
250 break;
251
252 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
253 break;
254
255 case RDMA_CM_EVENT_ADDR_CHANGE:
256 case RDMA_CM_EVENT_ROUTE_ERROR:
257 case RDMA_CM_EVENT_DEVICE_REMOVAL:
258 case RDMA_CM_EVENT_MULTICAST_JOIN:
259 case RDMA_CM_EVENT_MULTICAST_ERROR:
260 case RDMA_CM_EVENT_REJECTED:
261 case RDMA_CM_EVENT_CONNECT_REQUEST:
262 case RDMA_CM_EVENT_CONNECT_RESPONSE:
263 case RDMA_CM_EVENT_CONNECT_ERROR:
264 case RDMA_CM_EVENT_ADDR_ERROR:
265 case RDMA_CM_EVENT_UNREACHABLE:
266 c->status = Disconnected;
267 rdma_disconnect(rdma->cm_id);
268 break;
269 default:
270 BUG();
271 }
272 complete(&rdma->cm_done);
273 return 0;
274}
275
276static void
277handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
278 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
279{
280 struct p9_req_t *req;
281 int err = 0;
282 int16_t tag;
283
284 req = NULL;
285 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
286 DMA_FROM_DEVICE);
287
288 if (status != IB_WC_SUCCESS)
289 goto err_out;
290
291 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
292 if (err)
293 goto err_out;
294
295 req = p9_tag_lookup(client, tag);
296 if (!req)
297 goto err_out;
298
299 req->rc = c->rc;
Latchesar Ionkov1bab88b2009-04-05 16:28:59 -0500300 req->status = REQ_STATUS_RCVD;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500301 p9_client_cb(client, req);
302
303 return;
304
305 err_out:
Joe Perches5d385152011-11-28 10:40:46 -0800306 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500307 rdma->state = P9_RDMA_FLUSHING;
308 client->status = Disconnected;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500309}
310
311static void
312handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
313 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
314{
315 ib_dma_unmap_single(rdma->cm_id->device,
316 c->busa, c->req->tc->size,
317 DMA_TO_DEVICE);
318}
319
320static void qp_event_handler(struct ib_event *event, void *context)
321{
Joe Perches5d385152011-11-28 10:40:46 -0800322 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
323 event->event, context);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500324}
325
326static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
327{
328 struct p9_client *client = cq_context;
329 struct p9_trans_rdma *rdma = client->trans;
330 int ret;
331 struct ib_wc wc;
332
333 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
334 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
335 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
336
337 switch (c->wc_op) {
338 case IB_WC_RECV:
339 atomic_dec(&rdma->rq_count);
340 handle_recv(client, rdma, c, wc.status, wc.byte_len);
341 break;
342
343 case IB_WC_SEND:
344 handle_send(client, rdma, c, wc.status, wc.byte_len);
345 up(&rdma->sq_sem);
346 break;
347
348 default:
Joe Perches5d385152011-11-28 10:40:46 -0800349 pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500350 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{
Joe Perches5d385152011-11-28 10:40:46 -0800359 p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500360}
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:
Joe Perches5d385152011-11-28 10:40:46 -0800410 p9_debug(P9_DEBUG_ERROR, "EIO\n");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500411 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 */
Aneesh Kumar K.Veeff66e2011-03-08 16:39:47 +0530425 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000426 if (!rpl_context) {
427 err = -ENOMEM;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500428 goto err_close;
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000429 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500430
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,
Aneesh Kumar K.Veeff66e2011-03-08 16:39:47 +0530438 GFP_NOFS);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500439 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) {
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000447 err = -ENOMEM;
448 goto err_free2;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500449 }
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);
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000460 if (err)
461 goto err_free1;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500462 } 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 */
Aneesh Kumar K.Veeff66e2011-03-08 16:39:47 +0530469 c = kmalloc(sizeof *c, GFP_NOFS);
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000470 if (!c) {
471 err = -ENOMEM;
472 goto err_free1;
473 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500474 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:
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000500 kfree(c);
501 kfree(rpl_context->rc);
502 kfree(rpl_context);
Joe Perches5d385152011-11-28 10:40:46 -0800503 p9_debug(P9_DEBUG_ERROR, "EIO\n");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500504 return -EIO;
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000505 err_free1:
506 kfree(rpl_context->rc);
507 err_free2:
508 kfree(rpl_context);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500509 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
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500538 * @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 */
Sean Heftyb26f9b92010-04-01 17:08:41 +0000592 rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
593 IB_QPT_RC);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500594 if (IS_ERR(rdma->cm_id))
595 goto error;
596
Tom Tucker517ac452008-10-23 16:30:13 -0500597 /* Associate the client with the transport */
598 client->trans = rdma;
599
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500600 /* 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
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500680 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");