1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
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
61#define P9_RDMA_MAXSIZE (1024*1024)
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86struct p9_trans_rdma {
87 enum {
88 P9_RDMA_INIT,
89 P9_RDMA_ADDR_RESOLVED,
90 P9_RDMA_ROUTE_RESOLVED,
91 P9_RDMA_CONNECTED,
92 P9_RDMA_FLUSHING,
93 P9_RDMA_CLOSING,
94 P9_RDMA_CLOSED,
95 } state;
96 struct rdma_cm_id *cm_id;
97 struct ib_pd *pd;
98 struct ib_qp *qp;
99 struct ib_cq *cq;
100 long timeout;
101 bool privport;
102 u16 port;
103 int sq_depth;
104 struct semaphore sq_sem;
105 int rq_depth;
106 struct semaphore rq_sem;
107 atomic_t excess_rc;
108 struct sockaddr_in addr;
109 spinlock_t req_lock;
110
111 struct completion cm_done;
112};
113
114
115
116
117
118
119
120
121struct p9_rdma_req;
122struct p9_rdma_context {
123 struct ib_cqe cqe;
124 dma_addr_t busa;
125 union {
126 struct p9_req_t *req;
127 struct p9_fcall *rc;
128 };
129};
130
131
132
133
134
135
136
137
138
139struct p9_rdma_opts {
140 short port;
141 bool privport;
142 int sq_depth;
143 int rq_depth;
144 long timeout;
145};
146
147
148
149
150enum {
151
152 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
153
154 Opt_privport,
155 Opt_err,
156};
157
158static match_table_t tokens = {
159 {Opt_port, "port=%u"},
160 {Opt_sq_depth, "sq=%u"},
161 {Opt_rq_depth, "rq=%u"},
162 {Opt_timeout, "timeout=%u"},
163 {Opt_privport, "privport"},
164 {Opt_err, NULL},
165};
166
167static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
168{
169 struct p9_trans_rdma *rdma = clnt->trans;
170
171 if (rdma->port != P9_PORT)
172 seq_printf(m, ",port=%u", rdma->port);
173 if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
174 seq_printf(m, ",sq=%u", rdma->sq_depth);
175 if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
176 seq_printf(m, ",rq=%u", rdma->rq_depth);
177 if (rdma->timeout != P9_RDMA_TIMEOUT)
178 seq_printf(m, ",timeout=%lu", rdma->timeout);
179 if (rdma->privport)
180 seq_puts(m, ",privport");
181 return 0;
182}
183
184
185
186
187
188
189
190
191static int parse_opts(char *params, struct p9_rdma_opts *opts)
192{
193 char *p;
194 substring_t args[MAX_OPT_ARGS];
195 int option;
196 char *options, *tmp_options;
197
198 opts->port = P9_PORT;
199 opts->sq_depth = P9_RDMA_SQ_DEPTH;
200 opts->rq_depth = P9_RDMA_RQ_DEPTH;
201 opts->timeout = P9_RDMA_TIMEOUT;
202 opts->privport = false;
203
204 if (!params)
205 return 0;
206
207 tmp_options = kstrdup(params, GFP_KERNEL);
208 if (!tmp_options) {
209 p9_debug(P9_DEBUG_ERROR,
210 "failed to allocate copy of option string\n");
211 return -ENOMEM;
212 }
213 options = tmp_options;
214
215 while ((p = strsep(&options, ",")) != NULL) {
216 int token;
217 int r;
218 if (!*p)
219 continue;
220 token = match_token(p, tokens, args);
221 if ((token != Opt_err) && (token != Opt_privport)) {
222 r = match_int(&args[0], &option);
223 if (r < 0) {
224 p9_debug(P9_DEBUG_ERROR,
225 "integer field, but no integer?\n");
226 continue;
227 }
228 }
229 switch (token) {
230 case Opt_port:
231 opts->port = option;
232 break;
233 case Opt_sq_depth:
234 opts->sq_depth = option;
235 break;
236 case Opt_rq_depth:
237 opts->rq_depth = option;
238 break;
239 case Opt_timeout:
240 opts->timeout = option;
241 break;
242 case Opt_privport:
243 opts->privport = true;
244 break;
245 default:
246 continue;
247 }
248 }
249
250 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
251 kfree(tmp_options);
252 return 0;
253}
254
255static int
256p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
257{
258 struct p9_client *c = id->context;
259 struct p9_trans_rdma *rdma = c->trans;
260 switch (event->event) {
261 case RDMA_CM_EVENT_ADDR_RESOLVED:
262 BUG_ON(rdma->state != P9_RDMA_INIT);
263 rdma->state = P9_RDMA_ADDR_RESOLVED;
264 break;
265
266 case RDMA_CM_EVENT_ROUTE_RESOLVED:
267 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
268 rdma->state = P9_RDMA_ROUTE_RESOLVED;
269 break;
270
271 case RDMA_CM_EVENT_ESTABLISHED:
272 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
273 rdma->state = P9_RDMA_CONNECTED;
274 break;
275
276 case RDMA_CM_EVENT_DISCONNECTED:
277 if (rdma)
278 rdma->state = P9_RDMA_CLOSED;
279 if (c)
280 c->status = Disconnected;
281 break;
282
283 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
284 break;
285
286 case RDMA_CM_EVENT_ADDR_CHANGE:
287 case RDMA_CM_EVENT_ROUTE_ERROR:
288 case RDMA_CM_EVENT_DEVICE_REMOVAL:
289 case RDMA_CM_EVENT_MULTICAST_JOIN:
290 case RDMA_CM_EVENT_MULTICAST_ERROR:
291 case RDMA_CM_EVENT_REJECTED:
292 case RDMA_CM_EVENT_CONNECT_REQUEST:
293 case RDMA_CM_EVENT_CONNECT_RESPONSE:
294 case RDMA_CM_EVENT_CONNECT_ERROR:
295 case RDMA_CM_EVENT_ADDR_ERROR:
296 case RDMA_CM_EVENT_UNREACHABLE:
297 c->status = Disconnected;
298 rdma_disconnect(rdma->cm_id);
299 break;
300 default:
301 BUG();
302 }
303 complete(&rdma->cm_done);
304 return 0;
305}
306
307static void
308recv_done(struct ib_cq *cq, struct ib_wc *wc)
309{
310 struct p9_client *client = cq->cq_context;
311 struct p9_trans_rdma *rdma = client->trans;
312 struct p9_rdma_context *c =
313 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
314 struct p9_req_t *req;
315 int err = 0;
316 int16_t tag;
317
318 req = NULL;
319 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
320 DMA_FROM_DEVICE);
321
322 if (wc->status != IB_WC_SUCCESS)
323 goto err_out;
324
325 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
326 if (err)
327 goto err_out;
328
329 req = p9_tag_lookup(client, tag);
330 if (!req)
331 goto err_out;
332
333
334
335 if (unlikely(req->rc)) {
336 pr_err("Duplicate reply for request %d", tag);
337 goto err_out;
338 }
339
340 req->rc = c->rc;
341 p9_client_cb(client, req, REQ_STATUS_RCVD);
342
343 out:
344 up(&rdma->rq_sem);
345 kfree(c);
346 return;
347
348 err_out:
349 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
350 req, err, wc->status);
351 rdma->state = P9_RDMA_FLUSHING;
352 client->status = Disconnected;
353 goto out;
354}
355
356static void
357send_done(struct ib_cq *cq, struct ib_wc *wc)
358{
359 struct p9_client *client = cq->cq_context;
360 struct p9_trans_rdma *rdma = client->trans;
361 struct p9_rdma_context *c =
362 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
363
364 ib_dma_unmap_single(rdma->cm_id->device,
365 c->busa, c->req->tc->size,
366 DMA_TO_DEVICE);
367 up(&rdma->sq_sem);
368 kfree(c);
369}
370
371static void qp_event_handler(struct ib_event *event, void *context)
372{
373 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
374 event->event, context);
375}
376
377static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
378{
379 if (!rdma)
380 return;
381
382 if (rdma->qp && !IS_ERR(rdma->qp))
383 ib_destroy_qp(rdma->qp);
384
385 if (rdma->pd && !IS_ERR(rdma->pd))
386 ib_dealloc_pd(rdma->pd);
387
388 if (rdma->cq && !IS_ERR(rdma->cq))
389 ib_free_cq(rdma->cq);
390
391 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
392 rdma_destroy_id(rdma->cm_id);
393
394 kfree(rdma);
395}
396
397static int
398post_recv(struct p9_client *client, struct p9_rdma_context *c)
399{
400 struct p9_trans_rdma *rdma = client->trans;
401 struct ib_recv_wr wr, *bad_wr;
402 struct ib_sge sge;
403
404 c->busa = ib_dma_map_single(rdma->cm_id->device,
405 c->rc->sdata, client->msize,
406 DMA_FROM_DEVICE);
407 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
408 goto error;
409
410 c->cqe.done = recv_done;
411
412 sge.addr = c->busa;
413 sge.length = client->msize;
414 sge.lkey = rdma->pd->local_dma_lkey;
415
416 wr.next = NULL;
417 wr.wr_cqe = &c->cqe;
418 wr.sg_list = &sge;
419 wr.num_sge = 1;
420 return ib_post_recv(rdma->qp, &wr, &bad_wr);
421
422 error:
423 p9_debug(P9_DEBUG_ERROR, "EIO\n");
424 return -EIO;
425}
426
427static int rdma_request(struct p9_client *client, struct p9_req_t *req)
428{
429 struct p9_trans_rdma *rdma = client->trans;
430 struct ib_send_wr wr, *bad_wr;
431 struct ib_sge sge;
432 int err = 0;
433 unsigned long flags;
434 struct p9_rdma_context *c = NULL;
435 struct p9_rdma_context *rpl_context = NULL;
436
437
438
439
440
441
442
443
444
445 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
446 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
447
448 kfree(req->rc);
449 req->rc = NULL;
450 goto dont_need_post_recv;
451 } else {
452
453 atomic_inc(&rdma->excess_rc);
454 }
455 }
456
457
458 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
459 if (!rpl_context) {
460 err = -ENOMEM;
461 goto recv_error;
462 }
463 rpl_context->rc = req->rc;
464
465
466
467
468
469
470
471
472 if (down_interruptible(&rdma->rq_sem)) {
473 err = -EINTR;
474 goto recv_error;
475 }
476
477 err = post_recv(client, rpl_context);
478 if (err) {
479 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
480 goto recv_error;
481 }
482
483 req->rc = NULL;
484
485dont_need_post_recv:
486
487 c = kmalloc(sizeof *c, GFP_NOFS);
488 if (!c) {
489 err = -ENOMEM;
490 goto send_error;
491 }
492 c->req = req;
493
494 c->busa = ib_dma_map_single(rdma->cm_id->device,
495 c->req->tc->sdata, c->req->tc->size,
496 DMA_TO_DEVICE);
497 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
498 err = -EIO;
499 goto send_error;
500 }
501
502 c->cqe.done = send_done;
503
504 sge.addr = c->busa;
505 sge.length = c->req->tc->size;
506 sge.lkey = rdma->pd->local_dma_lkey;
507
508 wr.next = NULL;
509 wr.wr_cqe = &c->cqe;
510 wr.opcode = IB_WR_SEND;
511 wr.send_flags = IB_SEND_SIGNALED;
512 wr.sg_list = &sge;
513 wr.num_sge = 1;
514
515 if (down_interruptible(&rdma->sq_sem)) {
516 err = -EINTR;
517 goto send_error;
518 }
519
520
521
522
523
524 req->status = REQ_STATUS_SENT;
525 err = ib_post_send(rdma->qp, &wr, &bad_wr);
526 if (err)
527 goto send_error;
528
529
530 return 0;
531
532
533 send_error:
534 req->status = REQ_STATUS_ERROR;
535 kfree(c);
536 p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
537
538
539
540
541 atomic_inc(&rdma->excess_rc);
542 return err;
543
544
545 recv_error:
546 kfree(rpl_context);
547 spin_lock_irqsave(&rdma->req_lock, flags);
548 if (rdma->state < P9_RDMA_CLOSING) {
549 rdma->state = P9_RDMA_CLOSING;
550 spin_unlock_irqrestore(&rdma->req_lock, flags);
551 rdma_disconnect(rdma->cm_id);
552 } else
553 spin_unlock_irqrestore(&rdma->req_lock, flags);
554 return err;
555}
556
557static void rdma_close(struct p9_client *client)
558{
559 struct p9_trans_rdma *rdma;
560
561 if (!client)
562 return;
563
564 rdma = client->trans;
565 if (!rdma)
566 return;
567
568 client->status = Disconnected;
569 rdma_disconnect(rdma->cm_id);
570 rdma_destroy_trans(rdma);
571}
572
573
574
575
576
577static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
578{
579 struct p9_trans_rdma *rdma;
580
581 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
582 if (!rdma)
583 return NULL;
584
585 rdma->port = opts->port;
586 rdma->privport = opts->privport;
587 rdma->sq_depth = opts->sq_depth;
588 rdma->rq_depth = opts->rq_depth;
589 rdma->timeout = opts->timeout;
590 spin_lock_init(&rdma->req_lock);
591 init_completion(&rdma->cm_done);
592 sema_init(&rdma->sq_sem, rdma->sq_depth);
593 sema_init(&rdma->rq_sem, rdma->rq_depth);
594 atomic_set(&rdma->excess_rc, 0);
595
596 return rdma;
597}
598
599static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
600{
601
602
603
604 return 1;
605}
606
607
608
609
610static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
611{
612 struct p9_trans_rdma *rdma = client->trans;
613 atomic_inc(&rdma->excess_rc);
614 return 0;
615}
616
617static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
618{
619 struct sockaddr_in cl = {
620 .sin_family = AF_INET,
621 .sin_addr.s_addr = htonl(INADDR_ANY),
622 };
623 int port, err = -EINVAL;
624
625 for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
626 cl.sin_port = htons((ushort)port);
627 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
628 if (err != -EADDRINUSE)
629 break;
630 }
631 return err;
632}
633
634
635
636
637
638
639
640static int
641rdma_create_trans(struct p9_client *client, const char *addr, char *args)
642{
643 int err;
644 struct p9_rdma_opts opts;
645 struct p9_trans_rdma *rdma;
646 struct rdma_conn_param conn_param;
647 struct ib_qp_init_attr qp_attr;
648
649
650 err = parse_opts(args, &opts);
651 if (err < 0)
652 return err;
653
654
655 rdma = alloc_rdma(&opts);
656 if (!rdma)
657 return -ENOMEM;
658
659
660 rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
661 RDMA_PS_TCP, IB_QPT_RC);
662 if (IS_ERR(rdma->cm_id))
663 goto error;
664
665
666 client->trans = rdma;
667
668
669 if (opts.privport) {
670 err = p9_rdma_bind_privport(rdma);
671 if (err < 0) {
672 pr_err("%s (%d): problem binding to privport: %d\n",
673 __func__, task_pid_nr(current), -err);
674 goto error;
675 }
676 }
677
678
679 rdma->addr.sin_family = AF_INET;
680 rdma->addr.sin_addr.s_addr = in_aton(addr);
681 rdma->addr.sin_port = htons(opts.port);
682 err = rdma_resolve_addr(rdma->cm_id, NULL,
683 (struct sockaddr *)&rdma->addr,
684 rdma->timeout);
685 if (err)
686 goto error;
687 err = wait_for_completion_interruptible(&rdma->cm_done);
688 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
689 goto error;
690
691
692 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
693 if (err)
694 goto error;
695 err = wait_for_completion_interruptible(&rdma->cm_done);
696 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
697 goto error;
698
699
700 rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
701 opts.sq_depth + opts.rq_depth + 1,
702 0, IB_POLL_SOFTIRQ);
703 if (IS_ERR(rdma->cq))
704 goto error;
705
706
707 rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
708 if (IS_ERR(rdma->pd))
709 goto error;
710
711
712 memset(&qp_attr, 0, sizeof qp_attr);
713 qp_attr.event_handler = qp_event_handler;
714 qp_attr.qp_context = client;
715 qp_attr.cap.max_send_wr = opts.sq_depth;
716 qp_attr.cap.max_recv_wr = opts.rq_depth;
717 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
718 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
719 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
720 qp_attr.qp_type = IB_QPT_RC;
721 qp_attr.send_cq = rdma->cq;
722 qp_attr.recv_cq = rdma->cq;
723 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
724 if (err)
725 goto error;
726 rdma->qp = rdma->cm_id->qp;
727
728
729 memset(&conn_param, 0, sizeof(conn_param));
730 conn_param.private_data = NULL;
731 conn_param.private_data_len = 0;
732 conn_param.responder_resources = P9_RDMA_IRD;
733 conn_param.initiator_depth = P9_RDMA_ORD;
734 err = rdma_connect(rdma->cm_id, &conn_param);
735 if (err)
736 goto error;
737 err = wait_for_completion_interruptible(&rdma->cm_done);
738 if (err || (rdma->state != P9_RDMA_CONNECTED))
739 goto error;
740
741 client->status = Connected;
742
743 return 0;
744
745error:
746 rdma_destroy_trans(rdma);
747 return -ENOTCONN;
748}
749
750static struct p9_trans_module p9_rdma_trans = {
751 .name = "rdma",
752 .maxsize = P9_RDMA_MAXSIZE,
753 .def = 0,
754 .owner = THIS_MODULE,
755 .create = rdma_create_trans,
756 .close = rdma_close,
757 .request = rdma_request,
758 .cancel = rdma_cancel,
759 .cancelled = rdma_cancelled,
760 .show_options = p9_rdma_show_options,
761};
762
763
764
765
766static int __init p9_trans_rdma_init(void)
767{
768 v9fs_register_trans(&p9_rdma_trans);
769 return 0;
770}
771
772static void __exit p9_trans_rdma_exit(void)
773{
774 v9fs_unregister_trans(&p9_rdma_trans);
775}
776
777module_init(p9_trans_rdma_init);
778module_exit(p9_trans_rdma_exit);
779
780MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
781MODULE_DESCRIPTION("RDMA Transport for 9P");
782MODULE_LICENSE("Dual BSD/GPL");
783