aboutsummaryrefslogtreecommitdiffstats
path: root/net/9p
diff options
context:
space:
mode:
authorTom Tucker <tom@opengridcomputing.com>2008-10-22 19:47:39 -0400
committerEric Van Hensbergen <ericvh@opteron.(none)>2008-10-22 19:47:39 -0400
commitfc79d4b104f0eb8c2a7242150eaf8756ced4c344 (patch)
treeb8e9974570ca194acc3ebd24ba7a864bba80c477 /net/9p
parentea2e7996fc892e9becfed9145fdcefd59f697718 (diff)
9p: rdma: RDMA Transport Support for 9P
This patch implements the RDMA transport provider for 9P. It allows mounts to be performed over iWARP and IB capable network interfaces. Signed-off-by: Tom Tucker <tom@opengridcomputing.com> Signed-off-by: Latchesar Ionkov <lionkov@lanl.gov>
Diffstat (limited to 'net/9p')
-rw-r--r--net/9p/Kconfig6
-rw-r--r--net/9p/Makefile4
-rw-r--r--net/9p/trans_rdma.c712
3 files changed, 722 insertions, 0 deletions
diff --git a/net/9p/Kconfig b/net/9p/Kconfig
index ff34c5acc130..c42c0c400bf9 100644
--- a/net/9p/Kconfig
+++ b/net/9p/Kconfig
@@ -20,6 +20,12 @@ config NET_9P_VIRTIO
20 This builds support for a transports between 20 This builds support for a transports between
21 guest partitions and a host partition. 21 guest partitions and a host partition.
22 22
23config NET_9P_RDMA
24 depends on NET_9P && INFINIBAND && EXPERIMENTAL
25 tristate "9P RDMA Transport (Experimental)"
26 help
27 This builds support for a RDMA transport.
28
23config NET_9P_DEBUG 29config NET_9P_DEBUG
24 bool "Debug information" 30 bool "Debug information"
25 depends on NET_9P 31 depends on NET_9P
diff --git a/net/9p/Makefile b/net/9p/Makefile
index 1041b7bd12e2..198a640d53a6 100644
--- a/net/9p/Makefile
+++ b/net/9p/Makefile
@@ -1,5 +1,6 @@
1obj-$(CONFIG_NET_9P) := 9pnet.o 1obj-$(CONFIG_NET_9P) := 9pnet.o
2obj-$(CONFIG_NET_9P_VIRTIO) += 9pnet_virtio.o 2obj-$(CONFIG_NET_9P_VIRTIO) += 9pnet_virtio.o
3obj-$(CONFIG_NET_9P_RDMA) += 9pnet_rdma.o
3 4
49pnet-objs := \ 59pnet-objs := \
5 mod.o \ 6 mod.o \
@@ -11,3 +12,6 @@ obj-$(CONFIG_NET_9P_VIRTIO) += 9pnet_virtio.o
11 12
129pnet_virtio-objs := \ 139pnet_virtio-objs := \
13 trans_virtio.o \ 14 trans_virtio.o \
15
169pnet_rdma-objs := \
17 trans_rdma.o \
diff --git a/net/9p/trans_rdma.c b/net/9p/trans_rdma.c
new file mode 100644
index 000000000000..8d6cc4777aae
--- /dev/null
+++ b/net/9p/trans_rdma.c
@@ -0,0 +1,712 @@
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 <net/9p/9p.h>
44#include <net/9p/client.h>
45#include <net/9p/transport.h>
46#include <rdma/ib_verbs.h>
47#include <rdma/rdma_cm.h>
48#include <rdma/ib_verbs.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#define P9_RDMA_MAX_SGE (P9_RDMA_MAXSIZE >> PAGE_SHIFT)
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 * @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 * @addr: The remote peer's address
77 * @req_lock: Protects the active request list
78 * @send_wait: Wait list when the SQ fills up
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_options - parse mount options into session structure
159 * @options: options string passed from mount
160 * @opts: 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;
170 int ret;
171
172 opts->port = P9_PORT;
173 opts->sq_depth = P9_RDMA_SQ_DEPTH;
174 opts->rq_depth = P9_RDMA_RQ_DEPTH;
175 opts->timeout = P9_RDMA_TIMEOUT;
176
177 if (!params)
178 return 0;
179
180 options = kstrdup(params, GFP_KERNEL);
181 if (!options) {
182 P9_DPRINTK(P9_DEBUG_ERROR,
183 "failed to allocate copy of option string\n");
184 return -ENOMEM;
185 }
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 ret = r;
198 continue;
199 }
200 switch (token) {
201 case Opt_port:
202 opts->port = option;
203 break;
204 case Opt_sq_depth:
205 opts->sq_depth = option;
206 break;
207 case Opt_rq_depth:
208 opts->rq_depth = option;
209 break;
210 case Opt_timeout:
211 opts->timeout = option;
212 break;
213 default:
214 continue;
215 }
216 }
217 /* RQ must be at least as large as the SQ */
218 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
219 kfree(options);
220 return 0;
221}
222
223static int
224p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
225{
226 struct p9_client *c = id->context;
227 struct p9_trans_rdma *rdma = c->trans;
228 switch (event->event) {
229 case RDMA_CM_EVENT_ADDR_RESOLVED:
230 BUG_ON(rdma->state != P9_RDMA_INIT);
231 rdma->state = P9_RDMA_ADDR_RESOLVED;
232 break;
233
234 case RDMA_CM_EVENT_ROUTE_RESOLVED:
235 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
236 rdma->state = P9_RDMA_ROUTE_RESOLVED;
237 break;
238
239 case RDMA_CM_EVENT_ESTABLISHED:
240 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
241 rdma->state = P9_RDMA_CONNECTED;
242 break;
243
244 case RDMA_CM_EVENT_DISCONNECTED:
245 if (rdma)
246 rdma->state = P9_RDMA_CLOSED;
247 if (c)
248 c->status = Disconnected;
249 break;
250
251 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
252 break;
253
254 case RDMA_CM_EVENT_ADDR_CHANGE:
255 case RDMA_CM_EVENT_ROUTE_ERROR:
256 case RDMA_CM_EVENT_DEVICE_REMOVAL:
257 case RDMA_CM_EVENT_MULTICAST_JOIN:
258 case RDMA_CM_EVENT_MULTICAST_ERROR:
259 case RDMA_CM_EVENT_REJECTED:
260 case RDMA_CM_EVENT_CONNECT_REQUEST:
261 case RDMA_CM_EVENT_CONNECT_RESPONSE:
262 case RDMA_CM_EVENT_CONNECT_ERROR:
263 case RDMA_CM_EVENT_ADDR_ERROR:
264 case RDMA_CM_EVENT_UNREACHABLE:
265 c->status = Disconnected;
266 rdma_disconnect(rdma->cm_id);
267 break;
268 default:
269 BUG();
270 }
271 complete(&rdma->cm_done);
272 return 0;
273}
274
275static void
276handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
277 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
278{
279 struct p9_req_t *req;
280 int err = 0;
281 int16_t tag;
282
283 req = NULL;
284 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
285 DMA_FROM_DEVICE);
286
287 if (status != IB_WC_SUCCESS)
288 goto err_out;
289
290 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
291 if (err)
292 goto err_out;
293
294 req = p9_tag_lookup(client, tag);
295 if (!req)
296 goto err_out;
297
298 req->rc = c->rc;
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 return;
309}
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{
322 P9_DPRINTK(P9_DEBUG_ERROR, "QP event %d context %p\n", event->event,
323 context);
324}
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:
349 printk(KERN_ERR "9prdma: unexpected completion type, "
350 "c->wc_op=%d, wc.opcode=%d, status=%d\n",
351 c->wc_op, wc.opcode, wc.status);
352 break;
353 }
354 kfree(c);
355 }
356}
357
358static void cq_event_handler(struct ib_event *e, void *v)
359{
360 P9_DPRINTK(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
361}
362
363static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
364{
365 if (!rdma)
366 return;
367
368 if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
369 ib_dereg_mr(rdma->dma_mr);
370
371 if (rdma->qp && !IS_ERR(rdma->qp))
372 ib_destroy_qp(rdma->qp);
373
374 if (rdma->pd && !IS_ERR(rdma->pd))
375 ib_dealloc_pd(rdma->pd);
376
377 if (rdma->cq && !IS_ERR(rdma->cq))
378 ib_destroy_cq(rdma->cq);
379
380 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
381 rdma_destroy_id(rdma->cm_id);
382
383 kfree(rdma);
384}
385
386static int
387post_recv(struct p9_client *client, struct p9_rdma_context *c)
388{
389 struct p9_trans_rdma *rdma = client->trans;
390 struct ib_recv_wr wr, *bad_wr;
391 struct ib_sge sge;
392
393 c->busa = ib_dma_map_single(rdma->cm_id->device,
394 c->rc->sdata, client->msize,
395 DMA_FROM_DEVICE);
396 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
397 goto error;
398
399 sge.addr = c->busa;
400 sge.length = client->msize;
401 sge.lkey = rdma->lkey;
402
403 wr.next = NULL;
404 c->wc_op = IB_WC_RECV;
405 wr.wr_id = (unsigned long) c;
406 wr.sg_list = &sge;
407 wr.num_sge = 1;
408 return ib_post_recv(rdma->qp, &wr, &bad_wr);
409
410 error:
411 P9_DPRINTK(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, *bad_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 /* Allocate an fcall for the reply */
426 rpl_context = kmalloc(sizeof *rpl_context, GFP_KERNEL);
427 if (!rpl_context)
428 goto err_close;
429
430 /*
431 * If the request has a buffer, steal it, otherwise
432 * allocate a new one. Typically, requests should already
433 * have receive buffers allocated and just swap them around
434 */
435 if (!req->rc) {
436 req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
437 GFP_KERNEL);
438 if (req->rc) {
439 req->rc->sdata = (char *) req->rc +
440 sizeof(struct p9_fcall);
441 req->rc->capacity = client->msize;
442 }
443 }
444 rpl_context->rc = req->rc;
445 if (!rpl_context->rc) {
446 kfree(rpl_context);
447 goto err_close;
448 }
449
450 /*
451 * Post a receive buffer for this request. We need to ensure
452 * there is a reply buffer available for every outstanding
453 * request. A flushed request can result in no reply for an
454 * outstanding request, so we must keep a count to avoid
455 * overflowing the RQ.
456 */
457 if (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
458 err = post_recv(client, rpl_context);
459 if (err) {
460 kfree(rpl_context->rc);
461 kfree(rpl_context);
462 goto err_close;
463 }
464 } else
465 atomic_dec(&rdma->rq_count);
466
467 /* remove posted receive buffer from request structure */
468 req->rc = NULL;
469
470 /* Post the request */
471 c = kmalloc(sizeof *c, GFP_KERNEL);
472 if (!c)
473 goto err_close;
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 P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
501 return -EIO;
502
503 err_close:
504 spin_lock_irqsave(&rdma->req_lock, flags);
505 if (rdma->state < P9_RDMA_CLOSING) {
506 rdma->state = P9_RDMA_CLOSING;
507 spin_unlock_irqrestore(&rdma->req_lock, flags);
508 rdma_disconnect(rdma->cm_id);
509 } else
510 spin_unlock_irqrestore(&rdma->req_lock, flags);
511 return err;
512}
513
514static void rdma_close(struct p9_client *client)
515{
516 struct p9_trans_rdma *rdma;
517
518 if (!client)
519 return;
520
521 rdma = client->trans;
522 if (!rdma)
523 return;
524
525 client->status = Disconnected;
526 rdma_disconnect(rdma->cm_id);
527 rdma_destroy_trans(rdma);
528}
529
530/**
531 * alloc_rdma - Allocate and initialize the rdma transport structure
532 * @msize: MTU
533 * @dotu: Extension attribute
534 * @opts: Mount options structure
535 */
536static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
537{
538 struct p9_trans_rdma *rdma;
539
540 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
541 if (!rdma)
542 return NULL;
543
544 rdma->sq_depth = opts->sq_depth;
545 rdma->rq_depth = opts->rq_depth;
546 rdma->timeout = opts->timeout;
547 spin_lock_init(&rdma->req_lock);
548 init_completion(&rdma->cm_done);
549 sema_init(&rdma->sq_sem, rdma->sq_depth);
550 atomic_set(&rdma->rq_count, 0);
551
552 return rdma;
553}
554
555/* its not clear to me we can do anything after send has been posted */
556static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
557{
558 return 1;
559}
560
561/**
562 * trans_create_rdma - Transport method for creating atransport instance
563 * @client: client instance
564 * @addr: IP address string
565 * @args: Mount options string
566 */
567static int
568rdma_create_trans(struct p9_client *client, const char *addr, char *args)
569{
570 int err;
571 struct p9_rdma_opts opts;
572 struct p9_trans_rdma *rdma;
573 struct rdma_conn_param conn_param;
574 struct ib_qp_init_attr qp_attr;
575 struct ib_device_attr devattr;
576
577 /* Parse the transport specific mount options */
578 err = parse_opts(args, &opts);
579 if (err < 0)
580 return err;
581
582 /* Create and initialize the RDMA transport structure */
583 rdma = alloc_rdma(&opts);
584 if (!rdma)
585 return -ENOMEM;
586
587 /* Create the RDMA CM ID */
588 rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP);
589 if (IS_ERR(rdma->cm_id))
590 goto error;
591
592 /* Resolve the server's address */
593 rdma->addr.sin_family = AF_INET;
594 rdma->addr.sin_addr.s_addr = in_aton(addr);
595 rdma->addr.sin_port = htons(opts.port);
596 err = rdma_resolve_addr(rdma->cm_id, NULL,
597 (struct sockaddr *)&rdma->addr,
598 rdma->timeout);
599 if (err)
600 goto error;
601 err = wait_for_completion_interruptible(&rdma->cm_done);
602 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
603 goto error;
604
605 /* Resolve the route to the server */
606 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
607 if (err)
608 goto error;
609 err = wait_for_completion_interruptible(&rdma->cm_done);
610 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
611 goto error;
612
613 /* Query the device attributes */
614 err = ib_query_device(rdma->cm_id->device, &devattr);
615 if (err)
616 goto error;
617
618 /* Create the Completion Queue */
619 rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
620 cq_event_handler, client,
621 opts.sq_depth + opts.rq_depth + 1, 0);
622 if (IS_ERR(rdma->cq))
623 goto error;
624 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
625
626 /* Create the Protection Domain */
627 rdma->pd = ib_alloc_pd(rdma->cm_id->device);
628 if (IS_ERR(rdma->pd))
629 goto error;
630
631 /* Cache the DMA lkey in the transport */
632 rdma->dma_mr = NULL;
633 if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
634 rdma->lkey = rdma->cm_id->device->local_dma_lkey;
635 else {
636 rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
637 if (IS_ERR(rdma->dma_mr))
638 goto error;
639 rdma->lkey = rdma->dma_mr->lkey;
640 }
641
642 /* Create the Queue Pair */
643 memset(&qp_attr, 0, sizeof qp_attr);
644 qp_attr.event_handler = qp_event_handler;
645 qp_attr.qp_context = client;
646 qp_attr.cap.max_send_wr = opts.sq_depth;
647 qp_attr.cap.max_recv_wr = opts.rq_depth;
648 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
649 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
650 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
651 qp_attr.qp_type = IB_QPT_RC;
652 qp_attr.send_cq = rdma->cq;
653 qp_attr.recv_cq = rdma->cq;
654 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
655 if (err)
656 goto error;
657 rdma->qp = rdma->cm_id->qp;
658
659 /* Request a connection */
660 memset(&conn_param, 0, sizeof(conn_param));
661 conn_param.private_data = NULL;
662 conn_param.private_data_len = 0;
663 conn_param.responder_resources = P9_RDMA_IRD;
664 conn_param.initiator_depth = P9_RDMA_ORD;
665 err = rdma_connect(rdma->cm_id, &conn_param);
666 if (err)
667 goto error;
668 err = wait_for_completion_interruptible(&rdma->cm_done);
669 if (err || (rdma->state != P9_RDMA_CONNECTED))
670 goto error;
671
672 client->trans = rdma;
673 client->status = Connected;
674
675 return 0;
676
677error:
678 rdma_destroy_trans(rdma);
679 return -ENOTCONN;
680}
681
682static struct p9_trans_module p9_rdma_trans = {
683 .name = "rdma",
684 .maxsize = P9_RDMA_MAXSIZE,
685 .def = 0,
686 .owner = THIS_MODULE,
687 .create = rdma_create_trans,
688 .close = rdma_close,
689 .request = rdma_request,
690 .cancel = rdma_cancel,
691};
692
693/**
694 * p9_trans_rdma_init - Register the 9P RDMA transport driver
695 */
696static int __init p9_trans_rdma_init(void)
697{
698 v9fs_register_trans(&p9_rdma_trans);
699 return 0;
700}
701
702static void __exit p9_trans_rdma_exit(void)
703{
704 v9fs_unregister_trans(&p9_rdma_trans);
705}
706
707module_init(p9_trans_rdma_init);
708module_exit(p9_trans_rdma_exit);
709
710MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
711MODULE_DESCRIPTION("RDMA Transport for 9P");
712MODULE_LICENSE("Dual BSD/GPL");