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#include <linux/in.h>
30#include <linux/module.h>
31#include <linux/net.h>
32#include <linux/ipv6.h>
33#include <linux/kthread.h>
34#include <linux/errno.h>
35#include <linux/kernel.h>
36#include <linux/un.h>
37#include <linux/uaccess.h>
38#include <linux/inet.h>
39#include <linux/idr.h>
40#include <linux/file.h>
41#include <linux/parser.h>
42#include <linux/semaphore.h>
43#include <linux/slab.h>
44#include <net/9p/9p.h>
45#include <net/9p/client.h>
46#include <net/9p/transport.h>
47#include <rdma/ib_verbs.h>
48#include <rdma/rdma_cm.h>
49
50#define P9_PORT 5640
51#define P9_RDMA_SQ_DEPTH 32
52#define P9_RDMA_RQ_DEPTH 32
53#define P9_RDMA_SEND_SGE 4
54#define P9_RDMA_RECV_SGE 4
55#define P9_RDMA_IRD 0
56#define P9_RDMA_ORD 0
57#define P9_RDMA_TIMEOUT 30000
58#define P9_RDMA_MAXSIZE (4*4096)
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
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
110
111
112
113
114
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
128
129
130
131
132
133
134struct p9_rdma_opts {
135 short port;
136 int sq_depth;
137 int rq_depth;
138 long timeout;
139};
140
141
142
143
144enum {
145
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
159
160
161
162
163
164static int parse_opts(char *params, struct p9_rdma_opts *opts)
165{
166 char *p;
167 substring_t args[MAX_OPT_ARGS];
168 int option;
169 char *options, *tmp_options;
170
171 opts->port = P9_PORT;
172 opts->sq_depth = P9_RDMA_SQ_DEPTH;
173 opts->rq_depth = P9_RDMA_RQ_DEPTH;
174 opts->timeout = P9_RDMA_TIMEOUT;
175
176 if (!params)
177 return 0;
178
179 tmp_options = kstrdup(params, GFP_KERNEL);
180 if (!tmp_options) {
181 P9_DPRINTK(P9_DEBUG_ERROR,
182 "failed to allocate copy of option string\n");
183 return -ENOMEM;
184 }
185 options = tmp_options;
186
187 while ((p = strsep(&options, ",")) != NULL) {
188 int token;
189 int r;
190 if (!*p)
191 continue;
192 token = match_token(p, tokens, args);
193 r = match_int(&args[0], &option);
194 if (r < 0) {
195 P9_DPRINTK(P9_DEBUG_ERROR,
196 "integer field, but no integer?\n");
197 continue;
198 }
199 switch (token) {
200 case Opt_port:
201 opts->port = option;
202 break;
203 case Opt_sq_depth:
204 opts->sq_depth = option;
205 break;
206 case Opt_rq_depth:
207 opts->rq_depth = option;
208 break;
209 case Opt_timeout:
210 opts->timeout = option;
211 break;
212 default:
213 continue;
214 }
215 }
216
217 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
218 kfree(tmp_options);
219 return 0;
220}
221
222static int
223p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
224{
225 struct p9_client *c = id->context;
226 struct p9_trans_rdma *rdma = c->trans;
227 switch (event->event) {
228 case RDMA_CM_EVENT_ADDR_RESOLVED:
229 BUG_ON(rdma->state != P9_RDMA_INIT);
230 rdma->state = P9_RDMA_ADDR_RESOLVED;
231 break;
232
233 case RDMA_CM_EVENT_ROUTE_RESOLVED:
234 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
235 rdma->state = P9_RDMA_ROUTE_RESOLVED;
236 break;
237
238 case RDMA_CM_EVENT_ESTABLISHED:
239 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
240 rdma->state = P9_RDMA_CONNECTED;
241 break;
242
243 case RDMA_CM_EVENT_DISCONNECTED:
244 if (rdma)
245 rdma->state = P9_RDMA_CLOSED;
246 if (c)
247 c->status = Disconnected;
248 break;
249
250 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
251 break;
252
253 case RDMA_CM_EVENT_ADDR_CHANGE:
254 case RDMA_CM_EVENT_ROUTE_ERROR:
255 case RDMA_CM_EVENT_DEVICE_REMOVAL:
256 case RDMA_CM_EVENT_MULTICAST_JOIN:
257 case RDMA_CM_EVENT_MULTICAST_ERROR:
258 case RDMA_CM_EVENT_REJECTED:
259 case RDMA_CM_EVENT_CONNECT_REQUEST:
260 case RDMA_CM_EVENT_CONNECT_RESPONSE:
261 case RDMA_CM_EVENT_CONNECT_ERROR:
262 case RDMA_CM_EVENT_ADDR_ERROR:
263 case RDMA_CM_EVENT_UNREACHABLE:
264 c->status = Disconnected;
265 rdma_disconnect(rdma->cm_id);
266 break;
267 default:
268 BUG();
269 }
270 complete(&rdma->cm_done);
271 return 0;
272}
273
274static void
275handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
276 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
277{
278 struct p9_req_t *req;
279 int err = 0;
280 int16_t tag;
281
282 req = NULL;
283 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
284 DMA_FROM_DEVICE);
285
286 if (status != IB_WC_SUCCESS)
287 goto err_out;
288
289 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
290 if (err)
291 goto err_out;
292
293 req = p9_tag_lookup(client, tag);
294 if (!req)
295 goto err_out;
296
297 req->rc = c->rc;
298 req->status = REQ_STATUS_RCVD;
299 p9_client_cb(client, req);
300
301 return;
302
303 err_out:
304 P9_DPRINTK(P9_DEBUG_ERROR, "req %p err %d status %d\n",
305 req, err, status);
306 rdma->state = P9_RDMA_FLUSHING;
307 client->status = Disconnected;
308}
309
310static void
311handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
312 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
313{
314 ib_dma_unmap_single(rdma->cm_id->device,
315 c->busa, c->req->tc->size,
316 DMA_TO_DEVICE);
317}
318
319static void qp_event_handler(struct ib_event *event, void *context)
320{
321 P9_DPRINTK(P9_DEBUG_ERROR, "QP event %d context %p\n", event->event,
322 context);
323}
324
325static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
326{
327 struct p9_client *client = cq_context;
328 struct p9_trans_rdma *rdma = client->trans;
329 int ret;
330 struct ib_wc wc;
331
332 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
333 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
334 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
335
336 switch (c->wc_op) {
337 case IB_WC_RECV:
338 atomic_dec(&rdma->rq_count);
339 handle_recv(client, rdma, c, wc.status, wc.byte_len);
340 break;
341
342 case IB_WC_SEND:
343 handle_send(client, rdma, c, wc.status, wc.byte_len);
344 up(&rdma->sq_sem);
345 break;
346
347 default:
348 printk(KERN_ERR "9prdma: unexpected completion type, "
349 "c->wc_op=%d, wc.opcode=%d, status=%d\n",
350 c->wc_op, wc.opcode, wc.status);
351 break;
352 }
353 kfree(c);
354 }
355}
356
357static void cq_event_handler(struct ib_event *e, void *v)
358{
359 P9_DPRINTK(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
360}
361
362static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
363{
364 if (!rdma)
365 return;
366
367 if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
368 ib_dereg_mr(rdma->dma_mr);
369
370 if (rdma->qp && !IS_ERR(rdma->qp))
371 ib_destroy_qp(rdma->qp);
372
373 if (rdma->pd && !IS_ERR(rdma->pd))
374 ib_dealloc_pd(rdma->pd);
375
376 if (rdma->cq && !IS_ERR(rdma->cq))
377 ib_destroy_cq(rdma->cq);
378
379 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
380 rdma_destroy_id(rdma->cm_id);
381
382 kfree(rdma);
383}
384
385static int
386post_recv(struct p9_client *client, struct p9_rdma_context *c)
387{
388 struct p9_trans_rdma *rdma = client->trans;
389 struct ib_recv_wr wr, *bad_wr;
390 struct ib_sge sge;
391
392 c->busa = ib_dma_map_single(rdma->cm_id->device,
393 c->rc->sdata, client->msize,
394 DMA_FROM_DEVICE);
395 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
396 goto error;
397
398 sge.addr = c->busa;
399 sge.length = client->msize;
400 sge.lkey = rdma->lkey;
401
402 wr.next = NULL;
403 c->wc_op = IB_WC_RECV;
404 wr.wr_id = (unsigned long) c;
405 wr.sg_list = &sge;
406 wr.num_sge = 1;
407 return ib_post_recv(rdma->qp, &wr, &bad_wr);
408
409 error:
410 P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
411 return -EIO;
412}
413
414static int rdma_request(struct p9_client *client, struct p9_req_t *req)
415{
416 struct p9_trans_rdma *rdma = client->trans;
417 struct ib_send_wr wr, *bad_wr;
418 struct ib_sge sge;
419 int err = 0;
420 unsigned long flags;
421 struct p9_rdma_context *c = NULL;
422 struct p9_rdma_context *rpl_context = NULL;
423
424
425 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
426 if (!rpl_context) {
427 err = -ENOMEM;
428 goto err_close;
429 }
430
431
432
433
434
435
436 if (!req->rc) {
437 req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
438 GFP_NOFS);
439 if (req->rc) {
440 req->rc->sdata = (char *) req->rc +
441 sizeof(struct p9_fcall);
442 req->rc->capacity = client->msize;
443 }
444 }
445 rpl_context->rc = req->rc;
446 if (!rpl_context->rc) {
447 err = -ENOMEM;
448 goto err_free2;
449 }
450
451
452
453
454
455
456
457
458 if (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
459 err = post_recv(client, rpl_context);
460 if (err)
461 goto err_free1;
462 } else
463 atomic_dec(&rdma->rq_count);
464
465
466 req->rc = NULL;
467
468
469 c = kmalloc(sizeof *c, GFP_NOFS);
470 if (!c) {
471 err = -ENOMEM;
472 goto err_free1;
473 }
474 c->req = req;
475
476 c->busa = ib_dma_map_single(rdma->cm_id->device,
477 c->req->tc->sdata, c->req->tc->size,
478 DMA_TO_DEVICE);
479 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
480 goto error;
481
482 sge.addr = c->busa;
483 sge.length = c->req->tc->size;
484 sge.lkey = rdma->lkey;
485
486 wr.next = NULL;
487 c->wc_op = IB_WC_SEND;
488 wr.wr_id = (unsigned long) c;
489 wr.opcode = IB_WR_SEND;
490 wr.send_flags = IB_SEND_SIGNALED;
491 wr.sg_list = &sge;
492 wr.num_sge = 1;
493
494 if (down_interruptible(&rdma->sq_sem))
495 goto error;
496
497 return ib_post_send(rdma->qp, &wr, &bad_wr);
498
499 error:
500 kfree(c);
501 kfree(rpl_context->rc);
502 kfree(rpl_context);
503 P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
504 return -EIO;
505 err_free1:
506 kfree(rpl_context->rc);
507 err_free2:
508 kfree(rpl_context);
509 err_close:
510 spin_lock_irqsave(&rdma->req_lock, flags);
511 if (rdma->state < P9_RDMA_CLOSING) {
512 rdma->state = P9_RDMA_CLOSING;
513 spin_unlock_irqrestore(&rdma->req_lock, flags);
514 rdma_disconnect(rdma->cm_id);
515 } else
516 spin_unlock_irqrestore(&rdma->req_lock, flags);
517 return err;
518}
519
520static void rdma_close(struct p9_client *client)
521{
522 struct p9_trans_rdma *rdma;
523
524 if (!client)
525 return;
526
527 rdma = client->trans;
528 if (!rdma)
529 return;
530
531 client->status = Disconnected;
532 rdma_disconnect(rdma->cm_id);
533 rdma_destroy_trans(rdma);
534}
535
536
537
538
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
560static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
561{
562 return 1;
563}
564
565
566
567
568
569
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
582 err = parse_opts(args, &opts);
583 if (err < 0)
584 return err;
585
586
587 rdma = alloc_rdma(&opts);
588 if (!rdma)
589 return -ENOMEM;
590
591
592 rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
593 IB_QPT_RC);
594 if (IS_ERR(rdma->cm_id))
595 goto error;
596
597
598 client->trans = rdma;
599
600
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
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
622 err = ib_query_device(rdma->cm_id->device, &devattr);
623 if (err)
624 goto error;
625
626
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
635 rdma->pd = ib_alloc_pd(rdma->cm_id->device);
636 if (IS_ERR(rdma->pd))
637 goto error;
638
639
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
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
668 memset(&conn_param, 0, sizeof(conn_param));
669 conn_param.private_data = NULL;
670 conn_param.private_data_len = 0;
671 conn_param.responder_resources = P9_RDMA_IRD;
672 conn_param.initiator_depth = P9_RDMA_ORD;
673 err = rdma_connect(rdma->cm_id, &conn_param);
674 if (err)
675 goto error;
676 err = wait_for_completion_interruptible(&rdma->cm_done);
677 if (err || (rdma->state != P9_RDMA_CONNECTED))
678 goto error;
679
680 client->status = Connected;
681
682 return 0;
683
684error:
685 rdma_destroy_trans(rdma);
686 return -ENOTCONN;
687}
688
689static struct p9_trans_module p9_rdma_trans = {
690 .name = "rdma",
691 .maxsize = P9_RDMA_MAXSIZE,
692 .def = 0,
693 .owner = THIS_MODULE,
694 .create = rdma_create_trans,
695 .close = rdma_close,
696 .request = rdma_request,
697 .cancel = rdma_cancel,
698};
699
700
701
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");
720