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#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 */
84struct 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 */
119struct p9_rdma_req;
120struct 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 */
137struct 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 */
148enum {
149 /* Options that take integer arguments */
150 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
151 /* Options that take no argument */
152 Opt_privport,
153 Opt_err,
154};
155
156static match_table_t tokens = {
157 {Opt_port, "port=%u"},
158 {Opt_sq_depth, "sq=%u"},
159 {Opt_rq_depth, "rq=%u"},
160 {Opt_timeout, "timeout=%u"},
161 {Opt_privport, "privport"},
162 {Opt_err, NULL},
163};
164
165static 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 */
189static 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
253static int
254p9_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
305static void
306recv_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 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
324 if (err)
325 goto err_out;
326
327 req = p9_tag_lookup(client, tag);
328 if (!req)
329 goto err_out;
330
331 /* Check that we have not yet received a reply for this request.
332 */
333 if (unlikely(req->rc)) {
334 pr_err("Duplicate reply for request %d", tag);
335 goto err_out;
336 }
337
338 req->rc = c->rc;
339 p9_client_cb(client, req, REQ_STATUS_RCVD);
340
341 out:
342 up(&rdma->rq_sem);
343 kfree(c);
344 return;
345
346 err_out:
347 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
348 req, err, wc->status);
349 rdma->state = P9_RDMA_FLUSHING;
350 client->status = Disconnected;
351 goto out;
352}
353
354static void
355send_done(struct ib_cq *cq, struct ib_wc *wc)
356{
357 struct p9_client *client = cq->cq_context;
358 struct p9_trans_rdma *rdma = client->trans;
359 struct p9_rdma_context *c =
360 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
361
362 ib_dma_unmap_single(rdma->cm_id->device,
363 c->busa, c->req->tc->size,
364 DMA_TO_DEVICE);
365 up(&rdma->sq_sem);
366 kfree(c);
367}
368
369static void qp_event_handler(struct ib_event *event, void *context)
370{
371 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
372 event->event, context);
373}
374
375static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
376{
377 if (!rdma)
378 return;
379
380 if (rdma->qp && !IS_ERR(rdma->qp))
381 ib_destroy_qp(rdma->qp);
382
383 if (rdma->pd && !IS_ERR(rdma->pd))
384 ib_dealloc_pd(rdma->pd);
385
386 if (rdma->cq && !IS_ERR(rdma->cq))
387 ib_free_cq(rdma->cq);
388
389 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
390 rdma_destroy_id(rdma->cm_id);
391
392 kfree(rdma);
393}
394
395static int
396post_recv(struct p9_client *client, struct p9_rdma_context *c)
397{
398 struct p9_trans_rdma *rdma = client->trans;
399 struct ib_recv_wr wr, *bad_wr;
400 struct ib_sge sge;
401
402 c->busa = ib_dma_map_single(rdma->cm_id->device,
403 c->rc->sdata, client->msize,
404 DMA_FROM_DEVICE);
405 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
406 goto error;
407
408 c->cqe.done = recv_done;
409
410 sge.addr = c->busa;
411 sge.length = client->msize;
412 sge.lkey = rdma->pd->local_dma_lkey;
413
414 wr.next = NULL;
415 wr.wr_cqe = &c->cqe;
416 wr.sg_list = &sge;
417 wr.num_sge = 1;
418 return ib_post_recv(rdma->qp, &wr, &bad_wr);
419
420 error:
421 p9_debug(P9_DEBUG_ERROR, "EIO\n");
422 return -EIO;
423}
424
425static int rdma_request(struct p9_client *client, struct p9_req_t *req)
426{
427 struct p9_trans_rdma *rdma = client->trans;
428 struct ib_send_wr wr, *bad_wr;
429 struct ib_sge sge;
430 int err = 0;
431 unsigned long flags;
432 struct p9_rdma_context *c = NULL;
433 struct p9_rdma_context *rpl_context = NULL;
434
435 /* When an error occurs between posting the recv and the send,
436 * there will be a receive context posted without a pending request.
437 * Since there is no way to "un-post" it, we remember it and skip
438 * post_recv() for the next request.
439 * So here,
440 * see if we are this `next request' and need to absorb an excess rc.
441 * If yes, then drop and free our own, and do not recv_post().
442 **/
443 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
444 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
445 /* Got one ! */
446 kfree(req->rc);
447 req->rc = NULL;
448 goto dont_need_post_recv;
449 } else {
450 /* We raced and lost. */
451 atomic_inc(&rdma->excess_rc);
452 }
453 }
454
455 /* Allocate an fcall for the reply */
456 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
457 if (!rpl_context) {
458 err = -ENOMEM;
459 goto recv_error;
460 }
461 rpl_context->rc = req->rc;
462
463 /*
464 * Post a receive buffer for this request. We need to ensure
465 * there is a reply buffer available for every outstanding
466 * request. A flushed request can result in no reply for an
467 * outstanding request, so we must keep a count to avoid
468 * overflowing the RQ.
469 */
470 if (down_interruptible(&rdma->rq_sem)) {
471 err = -EINTR;
472 goto recv_error;
473 }
474
475 err = post_recv(client, rpl_context);
476 if (err) {
477 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
478 goto recv_error;
479 }
480 /* remove posted receive buffer from request structure */
481 req->rc = NULL;
482
483dont_need_post_recv:
484 /* Post the request */
485 c = kmalloc(sizeof *c, GFP_NOFS);
486 if (!c) {
487 err = -ENOMEM;
488 goto send_error;
489 }
490 c->req = req;
491
492 c->busa = ib_dma_map_single(rdma->cm_id->device,
493 c->req->tc->sdata, c->req->tc->size,
494 DMA_TO_DEVICE);
495 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
496 err = -EIO;
497 goto send_error;
498 }
499
500 c->cqe.done = send_done;
501
502 sge.addr = c->busa;
503 sge.length = c->req->tc->size;
504 sge.lkey = rdma->pd->local_dma_lkey;
505
506 wr.next = NULL;
507 wr.wr_cqe = &c->cqe;
508 wr.opcode = IB_WR_SEND;
509 wr.send_flags = IB_SEND_SIGNALED;
510 wr.sg_list = &sge;
511 wr.num_sge = 1;
512
513 if (down_interruptible(&rdma->sq_sem)) {
514 err = -EINTR;
515 goto send_error;
516 }
517
518 /* Mark request as `sent' *before* we actually send it,
519 * because doing if after could erase the REQ_STATUS_RCVD
520 * status in case of a very fast reply.
521 */
522 req->status = REQ_STATUS_SENT;
523 err = ib_post_send(rdma->qp, &wr, &bad_wr);
524 if (err)
525 goto send_error;
526
527 /* Success */
528 return 0;
529
530 /* Handle errors that happened during or while preparing the send: */
531 send_error:
532 req->status = REQ_STATUS_ERROR;
533 kfree(c);
534 p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
535
536 /* Ach.
537 * We did recv_post(), but not send. We have one recv_post in excess.
538 */
539 atomic_inc(&rdma->excess_rc);
540 return err;
541
542 /* Handle errors that happened during or while preparing post_recv(): */
543 recv_error:
544 kfree(rpl_context);
545 spin_lock_irqsave(&rdma->req_lock, flags);
546 if (rdma->state < P9_RDMA_CLOSING) {
547 rdma->state = P9_RDMA_CLOSING;
548 spin_unlock_irqrestore(&rdma->req_lock, flags);
549 rdma_disconnect(rdma->cm_id);
550 } else
551 spin_unlock_irqrestore(&rdma->req_lock, flags);
552 return err;
553}
554
555static void rdma_close(struct p9_client *client)
556{
557 struct p9_trans_rdma *rdma;
558
559 if (!client)
560 return;
561
562 rdma = client->trans;
563 if (!rdma)
564 return;
565
566 client->status = Disconnected;
567 rdma_disconnect(rdma->cm_id);
568 rdma_destroy_trans(rdma);
569}
570
571/**
572 * alloc_rdma - Allocate and initialize the rdma transport structure
573 * @opts: Mount options structure
574 */
575static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
576{
577 struct p9_trans_rdma *rdma;
578
579 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
580 if (!rdma)
581 return NULL;
582
583 rdma->port = opts->port;
584 rdma->privport = opts->privport;
585 rdma->sq_depth = opts->sq_depth;
586 rdma->rq_depth = opts->rq_depth;
587 rdma->timeout = opts->timeout;
588 spin_lock_init(&rdma->req_lock);
589 init_completion(&rdma->cm_done);
590 sema_init(&rdma->sq_sem, rdma->sq_depth);
591 sema_init(&rdma->rq_sem, rdma->rq_depth);
592 atomic_set(&rdma->excess_rc, 0);
593
594 return rdma;
595}
596
597static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
598{
599 /* Nothing to do here.
600 * We will take care of it (if we have to) in rdma_cancelled()
601 */
602 return 1;
603}
604
605/* A request has been fully flushed without a reply.
606 * That means we have posted one buffer in excess.
607 */
608static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
609{
610 struct p9_trans_rdma *rdma = client->trans;
611 atomic_inc(&rdma->excess_rc);
612 return 0;
613}
614
615static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
616{
617 struct sockaddr_in cl = {
618 .sin_family = AF_INET,
619 .sin_addr.s_addr = htonl(INADDR_ANY),
620 };
621 int port, err = -EINVAL;
622
623 for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
624 cl.sin_port = htons((ushort)port);
625 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
626 if (err != -EADDRINUSE)
627 break;
628 }
629 return err;
630}
631
632/**
633 * rdma_create_trans - Transport method for creating a transport instance
634 * @client: client instance
635 * @addr: IP address string
636 * @args: Mount options string
637 */
638static int
639rdma_create_trans(struct p9_client *client, const char *addr, char *args)
640{
641 int err;
642 struct p9_rdma_opts opts;
643 struct p9_trans_rdma *rdma;
644 struct rdma_conn_param conn_param;
645 struct ib_qp_init_attr qp_attr;
646
647 /* Parse the transport specific mount options */
648 err = parse_opts(args, &opts);
649 if (err < 0)
650 return err;
651
652 /* Create and initialize the RDMA transport structure */
653 rdma = alloc_rdma(&opts);
654 if (!rdma)
655 return -ENOMEM;
656
657 /* Create the RDMA CM ID */
658 rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
659 RDMA_PS_TCP, IB_QPT_RC);
660 if (IS_ERR(rdma->cm_id))
661 goto error;
662
663 /* Associate the client with the transport */
664 client->trans = rdma;
665
666 /* Bind to a privileged port if we need to */
667 if (opts.privport) {
668 err = p9_rdma_bind_privport(rdma);
669 if (err < 0) {
670 pr_err("%s (%d): problem binding to privport: %d\n",
671 __func__, task_pid_nr(current), -err);
672 goto error;
673 }
674 }
675
676 /* Resolve the server's address */
677 rdma->addr.sin_family = AF_INET;
678 rdma->addr.sin_addr.s_addr = in_aton(addr);
679 rdma->addr.sin_port = htons(opts.port);
680 err = rdma_resolve_addr(rdma->cm_id, NULL,
681 (struct sockaddr *)&rdma->addr,
682 rdma->timeout);
683 if (err)
684 goto error;
685 err = wait_for_completion_interruptible(&rdma->cm_done);
686 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
687 goto error;
688
689 /* Resolve the route to the server */
690 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
691 if (err)
692 goto error;
693 err = wait_for_completion_interruptible(&rdma->cm_done);
694 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
695 goto error;
696
697 /* Create the Completion Queue */
698 rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
699 opts.sq_depth + opts.rq_depth + 1,
700 0, IB_POLL_SOFTIRQ);
701 if (IS_ERR(rdma->cq))
702 goto error;
703
704 /* Create the Protection Domain */
705 rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
706 if (IS_ERR(rdma->pd))
707 goto error;
708
709 /* Create the Queue Pair */
710 memset(&qp_attr, 0, sizeof qp_attr);
711 qp_attr.event_handler = qp_event_handler;
712 qp_attr.qp_context = client;
713 qp_attr.cap.max_send_wr = opts.sq_depth;
714 qp_attr.cap.max_recv_wr = opts.rq_depth;
715 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
716 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
717 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
718 qp_attr.qp_type = IB_QPT_RC;
719 qp_attr.send_cq = rdma->cq;
720 qp_attr.recv_cq = rdma->cq;
721 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
722 if (err)
723 goto error;
724 rdma->qp = rdma->cm_id->qp;
725
726 /* Request a connection */
727 memset(&conn_param, 0, sizeof(conn_param));
728 conn_param.private_data = NULL;
729 conn_param.private_data_len = 0;
730 conn_param.responder_resources = P9_RDMA_IRD;
731 conn_param.initiator_depth = P9_RDMA_ORD;
732 err = rdma_connect(rdma->cm_id, &conn_param);
733 if (err)
734 goto error;
735 err = wait_for_completion_interruptible(&rdma->cm_done);
736 if (err || (rdma->state != P9_RDMA_CONNECTED))
737 goto error;
738
739 client->status = Connected;
740
741 return 0;
742
743error:
744 rdma_destroy_trans(rdma);
745 return -ENOTCONN;
746}
747
748static struct p9_trans_module p9_rdma_trans = {
749 .name = "rdma",
750 .maxsize = P9_RDMA_MAXSIZE,
751 .def = 0,
752 .owner = THIS_MODULE,
753 .create = rdma_create_trans,
754 .close = rdma_close,
755 .request = rdma_request,
756 .cancel = rdma_cancel,
757 .cancelled = rdma_cancelled,
758 .show_options = p9_rdma_show_options,
759};
760
761/**
762 * p9_trans_rdma_init - Register the 9P RDMA transport driver
763 */
764static int __init p9_trans_rdma_init(void)
765{
766 v9fs_register_trans(&p9_rdma_trans);
767 return 0;
768}
769
770static void __exit p9_trans_rdma_exit(void)
771{
772 v9fs_unregister_trans(&p9_rdma_trans);
773}
774
775module_init(p9_trans_rdma_init);
776module_exit(p9_trans_rdma_exit);
777
778MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
779MODULE_DESCRIPTION("RDMA Transport for 9P");
780MODULE_LICENSE("Dual BSD/GPL");