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
30
31
32
33#include <linux/completion.h>
34#include <linux/file.h>
35#include <linux/mutex.h>
36#include <linux/poll.h>
37#include <linux/sched.h>
38#include <linux/idr.h>
39#include <linux/in.h>
40#include <linux/in6.h>
41#include <linux/miscdevice.h>
42#include <linux/slab.h>
43#include <linux/sysctl.h>
44#include <linux/module.h>
45
46#include <rdma/rdma_user_cm.h>
47#include <rdma/ib_marshall.h>
48#include <rdma/rdma_cm.h>
49#include <rdma/rdma_cm_ib.h>
50
51MODULE_AUTHOR("Sean Hefty");
52MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
53MODULE_LICENSE("Dual BSD/GPL");
54
55static unsigned int max_backlog = 1024;
56
57static struct ctl_table_header *ucma_ctl_table_hdr;
58static ctl_table ucma_ctl_table[] = {
59 {
60 .procname = "max_backlog",
61 .data = &max_backlog,
62 .maxlen = sizeof max_backlog,
63 .mode = 0644,
64 .proc_handler = proc_dointvec,
65 },
66 { }
67};
68
69struct ucma_file {
70 struct mutex mut;
71 struct file *filp;
72 struct list_head ctx_list;
73 struct list_head event_list;
74 wait_queue_head_t poll_wait;
75};
76
77struct ucma_context {
78 int id;
79 struct completion comp;
80 atomic_t ref;
81 int events_reported;
82 int backlog;
83
84 struct ucma_file *file;
85 struct rdma_cm_id *cm_id;
86 u64 uid;
87
88 struct list_head list;
89 struct list_head mc_list;
90};
91
92struct ucma_multicast {
93 struct ucma_context *ctx;
94 int id;
95 int events_reported;
96
97 u64 uid;
98 struct list_head list;
99 struct sockaddr_storage addr;
100};
101
102struct ucma_event {
103 struct ucma_context *ctx;
104 struct ucma_multicast *mc;
105 struct list_head list;
106 struct rdma_cm_id *cm_id;
107 struct rdma_ucm_event_resp resp;
108};
109
110static DEFINE_MUTEX(mut);
111static DEFINE_IDR(ctx_idr);
112static DEFINE_IDR(multicast_idr);
113
114static inline struct ucma_context *_ucma_find_context(int id,
115 struct ucma_file *file)
116{
117 struct ucma_context *ctx;
118
119 ctx = idr_find(&ctx_idr, id);
120 if (!ctx)
121 ctx = ERR_PTR(-ENOENT);
122 else if (ctx->file != file)
123 ctx = ERR_PTR(-EINVAL);
124 return ctx;
125}
126
127static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
128{
129 struct ucma_context *ctx;
130
131 mutex_lock(&mut);
132 ctx = _ucma_find_context(id, file);
133 if (!IS_ERR(ctx))
134 atomic_inc(&ctx->ref);
135 mutex_unlock(&mut);
136 return ctx;
137}
138
139static void ucma_put_ctx(struct ucma_context *ctx)
140{
141 if (atomic_dec_and_test(&ctx->ref))
142 complete(&ctx->comp);
143}
144
145static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
146{
147 struct ucma_context *ctx;
148 int ret;
149
150 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
151 if (!ctx)
152 return NULL;
153
154 atomic_set(&ctx->ref, 1);
155 init_completion(&ctx->comp);
156 INIT_LIST_HEAD(&ctx->mc_list);
157 ctx->file = file;
158
159 do {
160 ret = idr_pre_get(&ctx_idr, GFP_KERNEL);
161 if (!ret)
162 goto error;
163
164 mutex_lock(&mut);
165 ret = idr_get_new(&ctx_idr, ctx, &ctx->id);
166 mutex_unlock(&mut);
167 } while (ret == -EAGAIN);
168
169 if (ret)
170 goto error;
171
172 list_add_tail(&ctx->list, &file->ctx_list);
173 return ctx;
174
175error:
176 kfree(ctx);
177 return NULL;
178}
179
180static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
181{
182 struct ucma_multicast *mc;
183 int ret;
184
185 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
186 if (!mc)
187 return NULL;
188
189 do {
190 ret = idr_pre_get(&multicast_idr, GFP_KERNEL);
191 if (!ret)
192 goto error;
193
194 mutex_lock(&mut);
195 ret = idr_get_new(&multicast_idr, mc, &mc->id);
196 mutex_unlock(&mut);
197 } while (ret == -EAGAIN);
198
199 if (ret)
200 goto error;
201
202 mc->ctx = ctx;
203 list_add_tail(&mc->list, &ctx->mc_list);
204 return mc;
205
206error:
207 kfree(mc);
208 return NULL;
209}
210
211static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
212 struct rdma_conn_param *src)
213{
214 if (src->private_data_len)
215 memcpy(dst->private_data, src->private_data,
216 src->private_data_len);
217 dst->private_data_len = src->private_data_len;
218 dst->responder_resources =src->responder_resources;
219 dst->initiator_depth = src->initiator_depth;
220 dst->flow_control = src->flow_control;
221 dst->retry_count = src->retry_count;
222 dst->rnr_retry_count = src->rnr_retry_count;
223 dst->srq = src->srq;
224 dst->qp_num = src->qp_num;
225}
226
227static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
228 struct rdma_ud_param *src)
229{
230 if (src->private_data_len)
231 memcpy(dst->private_data, src->private_data,
232 src->private_data_len);
233 dst->private_data_len = src->private_data_len;
234 ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
235 dst->qp_num = src->qp_num;
236 dst->qkey = src->qkey;
237}
238
239static void ucma_set_event_context(struct ucma_context *ctx,
240 struct rdma_cm_event *event,
241 struct ucma_event *uevent)
242{
243 uevent->ctx = ctx;
244 switch (event->event) {
245 case RDMA_CM_EVENT_MULTICAST_JOIN:
246 case RDMA_CM_EVENT_MULTICAST_ERROR:
247 uevent->mc = (struct ucma_multicast *)
248 event->param.ud.private_data;
249 uevent->resp.uid = uevent->mc->uid;
250 uevent->resp.id = uevent->mc->id;
251 break;
252 default:
253 uevent->resp.uid = ctx->uid;
254 uevent->resp.id = ctx->id;
255 break;
256 }
257}
258
259static int ucma_event_handler(struct rdma_cm_id *cm_id,
260 struct rdma_cm_event *event)
261{
262 struct ucma_event *uevent;
263 struct ucma_context *ctx = cm_id->context;
264 int ret = 0;
265
266 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
267 if (!uevent)
268 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
269
270 mutex_lock(&ctx->file->mut);
271 uevent->cm_id = cm_id;
272 ucma_set_event_context(ctx, event, uevent);
273 uevent->resp.event = event->event;
274 uevent->resp.status = event->status;
275 if (cm_id->qp_type == IB_QPT_UD)
276 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
277 else
278 ucma_copy_conn_event(&uevent->resp.param.conn,
279 &event->param.conn);
280
281 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
282 if (!ctx->backlog) {
283 ret = -ENOMEM;
284 kfree(uevent);
285 goto out;
286 }
287 ctx->backlog--;
288 } else if (!ctx->uid) {
289
290
291
292
293
294
295 kfree(uevent);
296 goto out;
297 }
298
299 list_add_tail(&uevent->list, &ctx->file->event_list);
300 wake_up_interruptible(&ctx->file->poll_wait);
301out:
302 mutex_unlock(&ctx->file->mut);
303 return ret;
304}
305
306static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
307 int in_len, int out_len)
308{
309 struct ucma_context *ctx;
310 struct rdma_ucm_get_event cmd;
311 struct ucma_event *uevent;
312 int ret = 0;
313 DEFINE_WAIT(wait);
314
315 if (out_len < sizeof uevent->resp)
316 return -ENOSPC;
317
318 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
319 return -EFAULT;
320
321 mutex_lock(&file->mut);
322 while (list_empty(&file->event_list)) {
323 mutex_unlock(&file->mut);
324
325 if (file->filp->f_flags & O_NONBLOCK)
326 return -EAGAIN;
327
328 if (wait_event_interruptible(file->poll_wait,
329 !list_empty(&file->event_list)))
330 return -ERESTARTSYS;
331
332 mutex_lock(&file->mut);
333 }
334
335 uevent = list_entry(file->event_list.next, struct ucma_event, list);
336
337 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
338 ctx = ucma_alloc_ctx(file);
339 if (!ctx) {
340 ret = -ENOMEM;
341 goto done;
342 }
343 uevent->ctx->backlog++;
344 ctx->cm_id = uevent->cm_id;
345 ctx->cm_id->context = ctx;
346 uevent->resp.id = ctx->id;
347 }
348
349 if (copy_to_user((void __user *)(unsigned long)cmd.response,
350 &uevent->resp, sizeof uevent->resp)) {
351 ret = -EFAULT;
352 goto done;
353 }
354
355 list_del(&uevent->list);
356 uevent->ctx->events_reported++;
357 if (uevent->mc)
358 uevent->mc->events_reported++;
359 kfree(uevent);
360done:
361 mutex_unlock(&file->mut);
362 return ret;
363}
364
365static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
366{
367 switch (cmd->ps) {
368 case RDMA_PS_TCP:
369 *qp_type = IB_QPT_RC;
370 return 0;
371 case RDMA_PS_UDP:
372 case RDMA_PS_IPOIB:
373 *qp_type = IB_QPT_UD;
374 return 0;
375 case RDMA_PS_IB:
376 *qp_type = cmd->qp_type;
377 return 0;
378 default:
379 return -EINVAL;
380 }
381}
382
383static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
384 int in_len, int out_len)
385{
386 struct rdma_ucm_create_id cmd;
387 struct rdma_ucm_create_id_resp resp;
388 struct ucma_context *ctx;
389 enum ib_qp_type qp_type;
390 int ret;
391
392 if (out_len < sizeof(resp))
393 return -ENOSPC;
394
395 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
396 return -EFAULT;
397
398 ret = ucma_get_qp_type(&cmd, &qp_type);
399 if (ret)
400 return ret;
401
402 mutex_lock(&file->mut);
403 ctx = ucma_alloc_ctx(file);
404 mutex_unlock(&file->mut);
405 if (!ctx)
406 return -ENOMEM;
407
408 ctx->uid = cmd.uid;
409 ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps, qp_type);
410 if (IS_ERR(ctx->cm_id)) {
411 ret = PTR_ERR(ctx->cm_id);
412 goto err1;
413 }
414
415 resp.id = ctx->id;
416 if (copy_to_user((void __user *)(unsigned long)cmd.response,
417 &resp, sizeof(resp))) {
418 ret = -EFAULT;
419 goto err2;
420 }
421 return 0;
422
423err2:
424 rdma_destroy_id(ctx->cm_id);
425err1:
426 mutex_lock(&mut);
427 idr_remove(&ctx_idr, ctx->id);
428 mutex_unlock(&mut);
429 kfree(ctx);
430 return ret;
431}
432
433static void ucma_cleanup_multicast(struct ucma_context *ctx)
434{
435 struct ucma_multicast *mc, *tmp;
436
437 mutex_lock(&mut);
438 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
439 list_del(&mc->list);
440 idr_remove(&multicast_idr, mc->id);
441 kfree(mc);
442 }
443 mutex_unlock(&mut);
444}
445
446static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
447{
448 struct ucma_event *uevent, *tmp;
449
450 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
451 if (uevent->mc != mc)
452 continue;
453
454 list_del(&uevent->list);
455 kfree(uevent);
456 }
457}
458
459
460
461
462
463
464static int ucma_free_ctx(struct ucma_context *ctx)
465{
466 int events_reported;
467 struct ucma_event *uevent, *tmp;
468 LIST_HEAD(list);
469
470
471 rdma_destroy_id(ctx->cm_id);
472
473 ucma_cleanup_multicast(ctx);
474
475
476 mutex_lock(&ctx->file->mut);
477 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
478 if (uevent->ctx == ctx)
479 list_move_tail(&uevent->list, &list);
480 }
481 list_del(&ctx->list);
482 mutex_unlock(&ctx->file->mut);
483
484 list_for_each_entry_safe(uevent, tmp, &list, list) {
485 list_del(&uevent->list);
486 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
487 rdma_destroy_id(uevent->cm_id);
488 kfree(uevent);
489 }
490
491 events_reported = ctx->events_reported;
492 kfree(ctx);
493 return events_reported;
494}
495
496static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
497 int in_len, int out_len)
498{
499 struct rdma_ucm_destroy_id cmd;
500 struct rdma_ucm_destroy_id_resp resp;
501 struct ucma_context *ctx;
502 int ret = 0;
503
504 if (out_len < sizeof(resp))
505 return -ENOSPC;
506
507 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
508 return -EFAULT;
509
510 mutex_lock(&mut);
511 ctx = _ucma_find_context(cmd.id, file);
512 if (!IS_ERR(ctx))
513 idr_remove(&ctx_idr, ctx->id);
514 mutex_unlock(&mut);
515
516 if (IS_ERR(ctx))
517 return PTR_ERR(ctx);
518
519 ucma_put_ctx(ctx);
520 wait_for_completion(&ctx->comp);
521 resp.events_reported = ucma_free_ctx(ctx);
522
523 if (copy_to_user((void __user *)(unsigned long)cmd.response,
524 &resp, sizeof(resp)))
525 ret = -EFAULT;
526
527 return ret;
528}
529
530static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf,
531 int in_len, int out_len)
532{
533 struct rdma_ucm_bind_addr cmd;
534 struct ucma_context *ctx;
535 int ret;
536
537 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
538 return -EFAULT;
539
540 ctx = ucma_get_ctx(file, cmd.id);
541 if (IS_ERR(ctx))
542 return PTR_ERR(ctx);
543
544 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
545 ucma_put_ctx(ctx);
546 return ret;
547}
548
549static ssize_t ucma_resolve_addr(struct ucma_file *file,
550 const char __user *inbuf,
551 int in_len, int out_len)
552{
553 struct rdma_ucm_resolve_addr cmd;
554 struct ucma_context *ctx;
555 int ret;
556
557 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
558 return -EFAULT;
559
560 ctx = ucma_get_ctx(file, cmd.id);
561 if (IS_ERR(ctx))
562 return PTR_ERR(ctx);
563
564 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
565 (struct sockaddr *) &cmd.dst_addr,
566 cmd.timeout_ms);
567 ucma_put_ctx(ctx);
568 return ret;
569}
570
571static ssize_t ucma_resolve_route(struct ucma_file *file,
572 const char __user *inbuf,
573 int in_len, int out_len)
574{
575 struct rdma_ucm_resolve_route cmd;
576 struct ucma_context *ctx;
577 int ret;
578
579 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
580 return -EFAULT;
581
582 ctx = ucma_get_ctx(file, cmd.id);
583 if (IS_ERR(ctx))
584 return PTR_ERR(ctx);
585
586 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
587 ucma_put_ctx(ctx);
588 return ret;
589}
590
591static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
592 struct rdma_route *route)
593{
594 struct rdma_dev_addr *dev_addr;
595
596 resp->num_paths = route->num_paths;
597 switch (route->num_paths) {
598 case 0:
599 dev_addr = &route->addr.dev_addr;
600 rdma_addr_get_dgid(dev_addr,
601 (union ib_gid *) &resp->ib_route[0].dgid);
602 rdma_addr_get_sgid(dev_addr,
603 (union ib_gid *) &resp->ib_route[0].sgid);
604 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
605 break;
606 case 2:
607 ib_copy_path_rec_to_user(&resp->ib_route[1],
608 &route->path_rec[1]);
609
610 case 1:
611 ib_copy_path_rec_to_user(&resp->ib_route[0],
612 &route->path_rec[0]);
613 break;
614 default:
615 break;
616 }
617}
618
619static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
620 struct rdma_route *route)
621{
622 struct rdma_dev_addr *dev_addr;
623 struct net_device *dev;
624 u16 vid = 0;
625
626 resp->num_paths = route->num_paths;
627 switch (route->num_paths) {
628 case 0:
629 dev_addr = &route->addr.dev_addr;
630 dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
631 if (dev) {
632 vid = rdma_vlan_dev_vlan_id(dev);
633 dev_put(dev);
634 }
635
636 iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid,
637 dev_addr->dst_dev_addr, vid);
638 iboe_addr_get_sgid(dev_addr,
639 (union ib_gid *) &resp->ib_route[0].sgid);
640 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
641 break;
642 case 2:
643 ib_copy_path_rec_to_user(&resp->ib_route[1],
644 &route->path_rec[1]);
645
646 case 1:
647 ib_copy_path_rec_to_user(&resp->ib_route[0],
648 &route->path_rec[0]);
649 break;
650 default:
651 break;
652 }
653}
654
655static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
656 struct rdma_route *route)
657{
658 struct rdma_dev_addr *dev_addr;
659
660 dev_addr = &route->addr.dev_addr;
661 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
662 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
663}
664
665static ssize_t ucma_query_route(struct ucma_file *file,
666 const char __user *inbuf,
667 int in_len, int out_len)
668{
669 struct rdma_ucm_query_route cmd;
670 struct rdma_ucm_query_route_resp resp;
671 struct ucma_context *ctx;
672 struct sockaddr *addr;
673 int ret = 0;
674
675 if (out_len < sizeof(resp))
676 return -ENOSPC;
677
678 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
679 return -EFAULT;
680
681 ctx = ucma_get_ctx(file, cmd.id);
682 if (IS_ERR(ctx))
683 return PTR_ERR(ctx);
684
685 memset(&resp, 0, sizeof resp);
686 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
687 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
688 sizeof(struct sockaddr_in) :
689 sizeof(struct sockaddr_in6));
690 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
691 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
692 sizeof(struct sockaddr_in) :
693 sizeof(struct sockaddr_in6));
694 if (!ctx->cm_id->device)
695 goto out;
696
697 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
698 resp.port_num = ctx->cm_id->port_num;
699 switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
700 case RDMA_TRANSPORT_IB:
701 switch (rdma_port_get_link_layer(ctx->cm_id->device,
702 ctx->cm_id->port_num)) {
703 case IB_LINK_LAYER_INFINIBAND:
704 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
705 break;
706 case IB_LINK_LAYER_ETHERNET:
707 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
708 break;
709 default:
710 break;
711 }
712 break;
713 case RDMA_TRANSPORT_IWARP:
714 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
715 break;
716 default:
717 break;
718 }
719
720out:
721 if (copy_to_user((void __user *)(unsigned long)cmd.response,
722 &resp, sizeof(resp)))
723 ret = -EFAULT;
724
725 ucma_put_ctx(ctx);
726 return ret;
727}
728
729static void ucma_copy_conn_param(struct rdma_conn_param *dst,
730 struct rdma_ucm_conn_param *src)
731{
732 dst->private_data = src->private_data;
733 dst->private_data_len = src->private_data_len;
734 dst->responder_resources =src->responder_resources;
735 dst->initiator_depth = src->initiator_depth;
736 dst->flow_control = src->flow_control;
737 dst->retry_count = src->retry_count;
738 dst->rnr_retry_count = src->rnr_retry_count;
739 dst->srq = src->srq;
740 dst->qp_num = src->qp_num;
741}
742
743static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
744 int in_len, int out_len)
745{
746 struct rdma_ucm_connect cmd;
747 struct rdma_conn_param conn_param;
748 struct ucma_context *ctx;
749 int ret;
750
751 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
752 return -EFAULT;
753
754 if (!cmd.conn_param.valid)
755 return -EINVAL;
756
757 ctx = ucma_get_ctx(file, cmd.id);
758 if (IS_ERR(ctx))
759 return PTR_ERR(ctx);
760
761 ucma_copy_conn_param(&conn_param, &cmd.conn_param);
762 ret = rdma_connect(ctx->cm_id, &conn_param);
763 ucma_put_ctx(ctx);
764 return ret;
765}
766
767static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
768 int in_len, int out_len)
769{
770 struct rdma_ucm_listen cmd;
771 struct ucma_context *ctx;
772 int ret;
773
774 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
775 return -EFAULT;
776
777 ctx = ucma_get_ctx(file, cmd.id);
778 if (IS_ERR(ctx))
779 return PTR_ERR(ctx);
780
781 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
782 cmd.backlog : max_backlog;
783 ret = rdma_listen(ctx->cm_id, ctx->backlog);
784 ucma_put_ctx(ctx);
785 return ret;
786}
787
788static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
789 int in_len, int out_len)
790{
791 struct rdma_ucm_accept cmd;
792 struct rdma_conn_param conn_param;
793 struct ucma_context *ctx;
794 int ret;
795
796 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
797 return -EFAULT;
798
799 ctx = ucma_get_ctx(file, cmd.id);
800 if (IS_ERR(ctx))
801 return PTR_ERR(ctx);
802
803 if (cmd.conn_param.valid) {
804 ucma_copy_conn_param(&conn_param, &cmd.conn_param);
805 mutex_lock(&file->mut);
806 ret = rdma_accept(ctx->cm_id, &conn_param);
807 if (!ret)
808 ctx->uid = cmd.uid;
809 mutex_unlock(&file->mut);
810 } else
811 ret = rdma_accept(ctx->cm_id, NULL);
812
813 ucma_put_ctx(ctx);
814 return ret;
815}
816
817static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
818 int in_len, int out_len)
819{
820 struct rdma_ucm_reject cmd;
821 struct ucma_context *ctx;
822 int ret;
823
824 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
825 return -EFAULT;
826
827 ctx = ucma_get_ctx(file, cmd.id);
828 if (IS_ERR(ctx))
829 return PTR_ERR(ctx);
830
831 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
832 ucma_put_ctx(ctx);
833 return ret;
834}
835
836static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
837 int in_len, int out_len)
838{
839 struct rdma_ucm_disconnect cmd;
840 struct ucma_context *ctx;
841 int ret;
842
843 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
844 return -EFAULT;
845
846 ctx = ucma_get_ctx(file, cmd.id);
847 if (IS_ERR(ctx))
848 return PTR_ERR(ctx);
849
850 ret = rdma_disconnect(ctx->cm_id);
851 ucma_put_ctx(ctx);
852 return ret;
853}
854
855static ssize_t ucma_init_qp_attr(struct ucma_file *file,
856 const char __user *inbuf,
857 int in_len, int out_len)
858{
859 struct rdma_ucm_init_qp_attr cmd;
860 struct ib_uverbs_qp_attr resp;
861 struct ucma_context *ctx;
862 struct ib_qp_attr qp_attr;
863 int ret;
864
865 if (out_len < sizeof(resp))
866 return -ENOSPC;
867
868 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
869 return -EFAULT;
870
871 ctx = ucma_get_ctx(file, cmd.id);
872 if (IS_ERR(ctx))
873 return PTR_ERR(ctx);
874
875 resp.qp_attr_mask = 0;
876 memset(&qp_attr, 0, sizeof qp_attr);
877 qp_attr.qp_state = cmd.qp_state;
878 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
879 if (ret)
880 goto out;
881
882 ib_copy_qp_attr_to_user(&resp, &qp_attr);
883 if (copy_to_user((void __user *)(unsigned long)cmd.response,
884 &resp, sizeof(resp)))
885 ret = -EFAULT;
886
887out:
888 ucma_put_ctx(ctx);
889 return ret;
890}
891
892static int ucma_set_option_id(struct ucma_context *ctx, int optname,
893 void *optval, size_t optlen)
894{
895 int ret = 0;
896
897 switch (optname) {
898 case RDMA_OPTION_ID_TOS:
899 if (optlen != sizeof(u8)) {
900 ret = -EINVAL;
901 break;
902 }
903 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
904 break;
905 case RDMA_OPTION_ID_REUSEADDR:
906 if (optlen != sizeof(int)) {
907 ret = -EINVAL;
908 break;
909 }
910 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
911 break;
912 case RDMA_OPTION_ID_AFONLY:
913 if (optlen != sizeof(int)) {
914 ret = -EINVAL;
915 break;
916 }
917 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
918 break;
919 default:
920 ret = -ENOSYS;
921 }
922
923 return ret;
924}
925
926static int ucma_set_ib_path(struct ucma_context *ctx,
927 struct ib_path_rec_data *path_data, size_t optlen)
928{
929 struct ib_sa_path_rec sa_path;
930 struct rdma_cm_event event;
931 int ret;
932
933 if (optlen % sizeof(*path_data))
934 return -EINVAL;
935
936 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
937 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
938 IB_PATH_BIDIRECTIONAL))
939 break;
940 }
941
942 if (!optlen)
943 return -EINVAL;
944
945 ib_sa_unpack_path(path_data->path_rec, &sa_path);
946 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
947 if (ret)
948 return ret;
949
950 memset(&event, 0, sizeof event);
951 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
952 return ucma_event_handler(ctx->cm_id, &event);
953}
954
955static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
956 void *optval, size_t optlen)
957{
958 int ret;
959
960 switch (optname) {
961 case RDMA_OPTION_IB_PATH:
962 ret = ucma_set_ib_path(ctx, optval, optlen);
963 break;
964 default:
965 ret = -ENOSYS;
966 }
967
968 return ret;
969}
970
971static int ucma_set_option_level(struct ucma_context *ctx, int level,
972 int optname, void *optval, size_t optlen)
973{
974 int ret;
975
976 switch (level) {
977 case RDMA_OPTION_ID:
978 ret = ucma_set_option_id(ctx, optname, optval, optlen);
979 break;
980 case RDMA_OPTION_IB:
981 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
982 break;
983 default:
984 ret = -ENOSYS;
985 }
986
987 return ret;
988}
989
990static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
991 int in_len, int out_len)
992{
993 struct rdma_ucm_set_option cmd;
994 struct ucma_context *ctx;
995 void *optval;
996 int ret;
997
998 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
999 return -EFAULT;
1000
1001 ctx = ucma_get_ctx(file, cmd.id);
1002 if (IS_ERR(ctx))
1003 return PTR_ERR(ctx);
1004
1005 optval = memdup_user((void __user *) (unsigned long) cmd.optval,
1006 cmd.optlen);
1007 if (IS_ERR(optval)) {
1008 ret = PTR_ERR(optval);
1009 goto out;
1010 }
1011
1012 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1013 cmd.optlen);
1014 kfree(optval);
1015
1016out:
1017 ucma_put_ctx(ctx);
1018 return ret;
1019}
1020
1021static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1022 int in_len, int out_len)
1023{
1024 struct rdma_ucm_notify cmd;
1025 struct ucma_context *ctx;
1026 int ret;
1027
1028 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1029 return -EFAULT;
1030
1031 ctx = ucma_get_ctx(file, cmd.id);
1032 if (IS_ERR(ctx))
1033 return PTR_ERR(ctx);
1034
1035 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1036 ucma_put_ctx(ctx);
1037 return ret;
1038}
1039
1040static ssize_t ucma_join_multicast(struct ucma_file *file,
1041 const char __user *inbuf,
1042 int in_len, int out_len)
1043{
1044 struct rdma_ucm_join_mcast cmd;
1045 struct rdma_ucm_create_id_resp resp;
1046 struct ucma_context *ctx;
1047 struct ucma_multicast *mc;
1048 int ret;
1049
1050 if (out_len < sizeof(resp))
1051 return -ENOSPC;
1052
1053 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1054 return -EFAULT;
1055
1056 ctx = ucma_get_ctx(file, cmd.id);
1057 if (IS_ERR(ctx))
1058 return PTR_ERR(ctx);
1059
1060 mutex_lock(&file->mut);
1061 mc = ucma_alloc_multicast(ctx);
1062 if (!mc) {
1063 ret = -ENOMEM;
1064 goto err1;
1065 }
1066
1067 mc->uid = cmd.uid;
1068 memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr);
1069 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1070 if (ret)
1071 goto err2;
1072
1073 resp.id = mc->id;
1074 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1075 &resp, sizeof(resp))) {
1076 ret = -EFAULT;
1077 goto err3;
1078 }
1079
1080 mutex_unlock(&file->mut);
1081 ucma_put_ctx(ctx);
1082 return 0;
1083
1084err3:
1085 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1086 ucma_cleanup_mc_events(mc);
1087err2:
1088 mutex_lock(&mut);
1089 idr_remove(&multicast_idr, mc->id);
1090 mutex_unlock(&mut);
1091 list_del(&mc->list);
1092 kfree(mc);
1093err1:
1094 mutex_unlock(&file->mut);
1095 ucma_put_ctx(ctx);
1096 return ret;
1097}
1098
1099static ssize_t ucma_leave_multicast(struct ucma_file *file,
1100 const char __user *inbuf,
1101 int in_len, int out_len)
1102{
1103 struct rdma_ucm_destroy_id cmd;
1104 struct rdma_ucm_destroy_id_resp resp;
1105 struct ucma_multicast *mc;
1106 int ret = 0;
1107
1108 if (out_len < sizeof(resp))
1109 return -ENOSPC;
1110
1111 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1112 return -EFAULT;
1113
1114 mutex_lock(&mut);
1115 mc = idr_find(&multicast_idr, cmd.id);
1116 if (!mc)
1117 mc = ERR_PTR(-ENOENT);
1118 else if (mc->ctx->file != file)
1119 mc = ERR_PTR(-EINVAL);
1120 else {
1121 idr_remove(&multicast_idr, mc->id);
1122 atomic_inc(&mc->ctx->ref);
1123 }
1124 mutex_unlock(&mut);
1125
1126 if (IS_ERR(mc)) {
1127 ret = PTR_ERR(mc);
1128 goto out;
1129 }
1130
1131 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1132 mutex_lock(&mc->ctx->file->mut);
1133 ucma_cleanup_mc_events(mc);
1134 list_del(&mc->list);
1135 mutex_unlock(&mc->ctx->file->mut);
1136
1137 ucma_put_ctx(mc->ctx);
1138 resp.events_reported = mc->events_reported;
1139 kfree(mc);
1140
1141 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1142 &resp, sizeof(resp)))
1143 ret = -EFAULT;
1144out:
1145 return ret;
1146}
1147
1148static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1149{
1150
1151 if (file1 < file2) {
1152 mutex_lock(&file1->mut);
1153 mutex_lock(&file2->mut);
1154 } else {
1155 mutex_lock(&file2->mut);
1156 mutex_lock(&file1->mut);
1157 }
1158}
1159
1160static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1161{
1162 if (file1 < file2) {
1163 mutex_unlock(&file2->mut);
1164 mutex_unlock(&file1->mut);
1165 } else {
1166 mutex_unlock(&file1->mut);
1167 mutex_unlock(&file2->mut);
1168 }
1169}
1170
1171static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1172{
1173 struct ucma_event *uevent, *tmp;
1174
1175 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1176 if (uevent->ctx == ctx)
1177 list_move_tail(&uevent->list, &file->event_list);
1178}
1179
1180static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1181 const char __user *inbuf,
1182 int in_len, int out_len)
1183{
1184 struct rdma_ucm_migrate_id cmd;
1185 struct rdma_ucm_migrate_resp resp;
1186 struct ucma_context *ctx;
1187 struct file *filp;
1188 struct ucma_file *cur_file;
1189 int ret = 0;
1190
1191 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1192 return -EFAULT;
1193
1194
1195 filp = fget(cmd.fd);
1196 if (!filp)
1197 return -ENOENT;
1198
1199
1200 ctx = ucma_get_ctx(filp->private_data, cmd.id);
1201 if (IS_ERR(ctx)) {
1202 ret = PTR_ERR(ctx);
1203 goto file_put;
1204 }
1205
1206 cur_file = ctx->file;
1207 if (cur_file == new_file) {
1208 resp.events_reported = ctx->events_reported;
1209 goto response;
1210 }
1211
1212
1213
1214
1215
1216 ucma_lock_files(cur_file, new_file);
1217 mutex_lock(&mut);
1218
1219 list_move_tail(&ctx->list, &new_file->ctx_list);
1220 ucma_move_events(ctx, new_file);
1221 ctx->file = new_file;
1222 resp.events_reported = ctx->events_reported;
1223
1224 mutex_unlock(&mut);
1225 ucma_unlock_files(cur_file, new_file);
1226
1227response:
1228 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1229 &resp, sizeof(resp)))
1230 ret = -EFAULT;
1231
1232 ucma_put_ctx(ctx);
1233file_put:
1234 fput(filp);
1235 return ret;
1236}
1237
1238static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1239 const char __user *inbuf,
1240 int in_len, int out_len) = {
1241 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1242 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1243 [RDMA_USER_CM_CMD_BIND_ADDR] = ucma_bind_addr,
1244 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1245 [RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route,
1246 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1247 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1248 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1249 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1250 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1251 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1252 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1253 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1254 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1255 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1256 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1257 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast,
1258 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1259 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id
1260};
1261
1262static ssize_t ucma_write(struct file *filp, const char __user *buf,
1263 size_t len, loff_t *pos)
1264{
1265 struct ucma_file *file = filp->private_data;
1266 struct rdma_ucm_cmd_hdr hdr;
1267 ssize_t ret;
1268
1269 if (len < sizeof(hdr))
1270 return -EINVAL;
1271
1272 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1273 return -EFAULT;
1274
1275 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1276 return -EINVAL;
1277
1278 if (hdr.in + sizeof(hdr) > len)
1279 return -EINVAL;
1280
1281 if (!ucma_cmd_table[hdr.cmd])
1282 return -ENOSYS;
1283
1284 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1285 if (!ret)
1286 ret = len;
1287
1288 return ret;
1289}
1290
1291static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1292{
1293 struct ucma_file *file = filp->private_data;
1294 unsigned int mask = 0;
1295
1296 poll_wait(filp, &file->poll_wait, wait);
1297
1298 if (!list_empty(&file->event_list))
1299 mask = POLLIN | POLLRDNORM;
1300
1301 return mask;
1302}
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312static int ucma_open(struct inode *inode, struct file *filp)
1313{
1314 struct ucma_file *file;
1315
1316 file = kmalloc(sizeof *file, GFP_KERNEL);
1317 if (!file)
1318 return -ENOMEM;
1319
1320 INIT_LIST_HEAD(&file->event_list);
1321 INIT_LIST_HEAD(&file->ctx_list);
1322 init_waitqueue_head(&file->poll_wait);
1323 mutex_init(&file->mut);
1324
1325 filp->private_data = file;
1326 file->filp = filp;
1327
1328 return nonseekable_open(inode, filp);
1329}
1330
1331static int ucma_close(struct inode *inode, struct file *filp)
1332{
1333 struct ucma_file *file = filp->private_data;
1334 struct ucma_context *ctx, *tmp;
1335
1336 mutex_lock(&file->mut);
1337 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1338 mutex_unlock(&file->mut);
1339
1340 mutex_lock(&mut);
1341 idr_remove(&ctx_idr, ctx->id);
1342 mutex_unlock(&mut);
1343
1344 ucma_free_ctx(ctx);
1345 mutex_lock(&file->mut);
1346 }
1347 mutex_unlock(&file->mut);
1348 kfree(file);
1349 return 0;
1350}
1351
1352static const struct file_operations ucma_fops = {
1353 .owner = THIS_MODULE,
1354 .open = ucma_open,
1355 .release = ucma_close,
1356 .write = ucma_write,
1357 .poll = ucma_poll,
1358 .llseek = no_llseek,
1359};
1360
1361static struct miscdevice ucma_misc = {
1362 .minor = MISC_DYNAMIC_MINOR,
1363 .name = "rdma_cm",
1364 .nodename = "infiniband/rdma_cm",
1365 .mode = 0666,
1366 .fops = &ucma_fops,
1367};
1368
1369static ssize_t show_abi_version(struct device *dev,
1370 struct device_attribute *attr,
1371 char *buf)
1372{
1373 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1374}
1375static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1376
1377static int __init ucma_init(void)
1378{
1379 int ret;
1380
1381 ret = misc_register(&ucma_misc);
1382 if (ret)
1383 return ret;
1384
1385 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1386 if (ret) {
1387 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1388 goto err1;
1389 }
1390
1391 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1392 if (!ucma_ctl_table_hdr) {
1393 printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n");
1394 ret = -ENOMEM;
1395 goto err2;
1396 }
1397 return 0;
1398err2:
1399 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1400err1:
1401 misc_deregister(&ucma_misc);
1402 return ret;
1403}
1404
1405static void __exit ucma_cleanup(void)
1406{
1407 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1408 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1409 misc_deregister(&ucma_misc);
1410 idr_destroy(&ctx_idr);
1411}
1412
1413module_init(ucma_init);
1414module_exit(ucma_cleanup);
1415
