Loading...
1/*
2 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
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
45#include <rdma/rdma_user_cm.h>
46#include <rdma/ib_marshall.h>
47#include <rdma/rdma_cm.h>
48#include <rdma/rdma_cm_ib.h>
49
50MODULE_AUTHOR("Sean Hefty");
51MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
52MODULE_LICENSE("Dual BSD/GPL");
53
54static unsigned int max_backlog = 1024;
55
56static struct ctl_table_header *ucma_ctl_table_hdr;
57static ctl_table ucma_ctl_table[] = {
58 {
59 .procname = "max_backlog",
60 .data = &max_backlog,
61 .maxlen = sizeof max_backlog,
62 .mode = 0644,
63 .proc_handler = proc_dointvec,
64 },
65 { }
66};
67
68static struct ctl_path ucma_ctl_path[] = {
69 { .procname = "net" },
70 { .procname = "rdma_ucm" },
71 { }
72};
73
74struct ucma_file {
75 struct mutex mut;
76 struct file *filp;
77 struct list_head ctx_list;
78 struct list_head event_list;
79 wait_queue_head_t poll_wait;
80};
81
82struct ucma_context {
83 int id;
84 struct completion comp;
85 atomic_t ref;
86 int events_reported;
87 int backlog;
88
89 struct ucma_file *file;
90 struct rdma_cm_id *cm_id;
91 u64 uid;
92
93 struct list_head list;
94 struct list_head mc_list;
95};
96
97struct ucma_multicast {
98 struct ucma_context *ctx;
99 int id;
100 int events_reported;
101
102 u64 uid;
103 struct list_head list;
104 struct sockaddr_storage addr;
105};
106
107struct ucma_event {
108 struct ucma_context *ctx;
109 struct ucma_multicast *mc;
110 struct list_head list;
111 struct rdma_cm_id *cm_id;
112 struct rdma_ucm_event_resp resp;
113};
114
115static DEFINE_MUTEX(mut);
116static DEFINE_IDR(ctx_idr);
117static DEFINE_IDR(multicast_idr);
118
119static inline struct ucma_context *_ucma_find_context(int id,
120 struct ucma_file *file)
121{
122 struct ucma_context *ctx;
123
124 ctx = idr_find(&ctx_idr, id);
125 if (!ctx)
126 ctx = ERR_PTR(-ENOENT);
127 else if (ctx->file != file)
128 ctx = ERR_PTR(-EINVAL);
129 return ctx;
130}
131
132static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
133{
134 struct ucma_context *ctx;
135
136 mutex_lock(&mut);
137 ctx = _ucma_find_context(id, file);
138 if (!IS_ERR(ctx))
139 atomic_inc(&ctx->ref);
140 mutex_unlock(&mut);
141 return ctx;
142}
143
144static void ucma_put_ctx(struct ucma_context *ctx)
145{
146 if (atomic_dec_and_test(&ctx->ref))
147 complete(&ctx->comp);
148}
149
150static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
151{
152 struct ucma_context *ctx;
153 int ret;
154
155 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
156 if (!ctx)
157 return NULL;
158
159 atomic_set(&ctx->ref, 1);
160 init_completion(&ctx->comp);
161 INIT_LIST_HEAD(&ctx->mc_list);
162 ctx->file = file;
163
164 do {
165 ret = idr_pre_get(&ctx_idr, GFP_KERNEL);
166 if (!ret)
167 goto error;
168
169 mutex_lock(&mut);
170 ret = idr_get_new(&ctx_idr, ctx, &ctx->id);
171 mutex_unlock(&mut);
172 } while (ret == -EAGAIN);
173
174 if (ret)
175 goto error;
176
177 list_add_tail(&ctx->list, &file->ctx_list);
178 return ctx;
179
180error:
181 kfree(ctx);
182 return NULL;
183}
184
185static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
186{
187 struct ucma_multicast *mc;
188 int ret;
189
190 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
191 if (!mc)
192 return NULL;
193
194 do {
195 ret = idr_pre_get(&multicast_idr, GFP_KERNEL);
196 if (!ret)
197 goto error;
198
199 mutex_lock(&mut);
200 ret = idr_get_new(&multicast_idr, mc, &mc->id);
201 mutex_unlock(&mut);
202 } while (ret == -EAGAIN);
203
204 if (ret)
205 goto error;
206
207 mc->ctx = ctx;
208 list_add_tail(&mc->list, &ctx->mc_list);
209 return mc;
210
211error:
212 kfree(mc);
213 return NULL;
214}
215
216static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
217 struct rdma_conn_param *src)
218{
219 if (src->private_data_len)
220 memcpy(dst->private_data, src->private_data,
221 src->private_data_len);
222 dst->private_data_len = src->private_data_len;
223 dst->responder_resources =src->responder_resources;
224 dst->initiator_depth = src->initiator_depth;
225 dst->flow_control = src->flow_control;
226 dst->retry_count = src->retry_count;
227 dst->rnr_retry_count = src->rnr_retry_count;
228 dst->srq = src->srq;
229 dst->qp_num = src->qp_num;
230}
231
232static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
233 struct rdma_ud_param *src)
234{
235 if (src->private_data_len)
236 memcpy(dst->private_data, src->private_data,
237 src->private_data_len);
238 dst->private_data_len = src->private_data_len;
239 ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
240 dst->qp_num = src->qp_num;
241 dst->qkey = src->qkey;
242}
243
244static void ucma_set_event_context(struct ucma_context *ctx,
245 struct rdma_cm_event *event,
246 struct ucma_event *uevent)
247{
248 uevent->ctx = ctx;
249 switch (event->event) {
250 case RDMA_CM_EVENT_MULTICAST_JOIN:
251 case RDMA_CM_EVENT_MULTICAST_ERROR:
252 uevent->mc = (struct ucma_multicast *)
253 event->param.ud.private_data;
254 uevent->resp.uid = uevent->mc->uid;
255 uevent->resp.id = uevent->mc->id;
256 break;
257 default:
258 uevent->resp.uid = ctx->uid;
259 uevent->resp.id = ctx->id;
260 break;
261 }
262}
263
264static int ucma_event_handler(struct rdma_cm_id *cm_id,
265 struct rdma_cm_event *event)
266{
267 struct ucma_event *uevent;
268 struct ucma_context *ctx = cm_id->context;
269 int ret = 0;
270
271 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
272 if (!uevent)
273 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
274
275 uevent->cm_id = cm_id;
276 ucma_set_event_context(ctx, event, uevent);
277 uevent->resp.event = event->event;
278 uevent->resp.status = event->status;
279 if (cm_id->ps == RDMA_PS_UDP || cm_id->ps == RDMA_PS_IPOIB)
280 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
281 else
282 ucma_copy_conn_event(&uevent->resp.param.conn,
283 &event->param.conn);
284
285 mutex_lock(&ctx->file->mut);
286 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
287 if (!ctx->backlog) {
288 ret = -ENOMEM;
289 kfree(uevent);
290 goto out;
291 }
292 ctx->backlog--;
293 } else if (!ctx->uid) {
294 /*
295 * We ignore events for new connections until userspace has set
296 * their context. This can only happen if an error occurs on a
297 * new connection before the user accepts it. This is okay,
298 * since the accept will just fail later.
299 */
300 kfree(uevent);
301 goto out;
302 }
303
304 list_add_tail(&uevent->list, &ctx->file->event_list);
305 wake_up_interruptible(&ctx->file->poll_wait);
306out:
307 mutex_unlock(&ctx->file->mut);
308 return ret;
309}
310
311static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
312 int in_len, int out_len)
313{
314 struct ucma_context *ctx;
315 struct rdma_ucm_get_event cmd;
316 struct ucma_event *uevent;
317 int ret = 0;
318 DEFINE_WAIT(wait);
319
320 if (out_len < sizeof uevent->resp)
321 return -ENOSPC;
322
323 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
324 return -EFAULT;
325
326 mutex_lock(&file->mut);
327 while (list_empty(&file->event_list)) {
328 mutex_unlock(&file->mut);
329
330 if (file->filp->f_flags & O_NONBLOCK)
331 return -EAGAIN;
332
333 if (wait_event_interruptible(file->poll_wait,
334 !list_empty(&file->event_list)))
335 return -ERESTARTSYS;
336
337 mutex_lock(&file->mut);
338 }
339
340 uevent = list_entry(file->event_list.next, struct ucma_event, list);
341
342 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
343 ctx = ucma_alloc_ctx(file);
344 if (!ctx) {
345 ret = -ENOMEM;
346 goto done;
347 }
348 uevent->ctx->backlog++;
349 ctx->cm_id = uevent->cm_id;
350 ctx->cm_id->context = ctx;
351 uevent->resp.id = ctx->id;
352 }
353
354 if (copy_to_user((void __user *)(unsigned long)cmd.response,
355 &uevent->resp, sizeof uevent->resp)) {
356 ret = -EFAULT;
357 goto done;
358 }
359
360 list_del(&uevent->list);
361 uevent->ctx->events_reported++;
362 if (uevent->mc)
363 uevent->mc->events_reported++;
364 kfree(uevent);
365done:
366 mutex_unlock(&file->mut);
367 return ret;
368}
369
370static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
371{
372 switch (cmd->ps) {
373 case RDMA_PS_TCP:
374 *qp_type = IB_QPT_RC;
375 return 0;
376 case RDMA_PS_UDP:
377 case RDMA_PS_IPOIB:
378 *qp_type = IB_QPT_UD;
379 return 0;
380 default:
381 return -EINVAL;
382 }
383}
384
385static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
386 int in_len, int out_len)
387{
388 struct rdma_ucm_create_id cmd;
389 struct rdma_ucm_create_id_resp resp;
390 struct ucma_context *ctx;
391 enum ib_qp_type qp_type;
392 int ret;
393
394 if (out_len < sizeof(resp))
395 return -ENOSPC;
396
397 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
398 return -EFAULT;
399
400 ret = ucma_get_qp_type(&cmd, &qp_type);
401 if (ret)
402 return ret;
403
404 mutex_lock(&file->mut);
405 ctx = ucma_alloc_ctx(file);
406 mutex_unlock(&file->mut);
407 if (!ctx)
408 return -ENOMEM;
409
410 ctx->uid = cmd.uid;
411 ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps, qp_type);
412 if (IS_ERR(ctx->cm_id)) {
413 ret = PTR_ERR(ctx->cm_id);
414 goto err1;
415 }
416
417 resp.id = ctx->id;
418 if (copy_to_user((void __user *)(unsigned long)cmd.response,
419 &resp, sizeof(resp))) {
420 ret = -EFAULT;
421 goto err2;
422 }
423 return 0;
424
425err2:
426 rdma_destroy_id(ctx->cm_id);
427err1:
428 mutex_lock(&mut);
429 idr_remove(&ctx_idr, ctx->id);
430 mutex_unlock(&mut);
431 kfree(ctx);
432 return ret;
433}
434
435static void ucma_cleanup_multicast(struct ucma_context *ctx)
436{
437 struct ucma_multicast *mc, *tmp;
438
439 mutex_lock(&mut);
440 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
441 list_del(&mc->list);
442 idr_remove(&multicast_idr, mc->id);
443 kfree(mc);
444 }
445 mutex_unlock(&mut);
446}
447
448static void ucma_cleanup_events(struct ucma_context *ctx)
449{
450 struct ucma_event *uevent, *tmp;
451
452 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
453 if (uevent->ctx != ctx)
454 continue;
455
456 list_del(&uevent->list);
457
458 /* clear incoming connections. */
459 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
460 rdma_destroy_id(uevent->cm_id);
461
462 kfree(uevent);
463 }
464}
465
466static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
467{
468 struct ucma_event *uevent, *tmp;
469
470 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
471 if (uevent->mc != mc)
472 continue;
473
474 list_del(&uevent->list);
475 kfree(uevent);
476 }
477}
478
479static int ucma_free_ctx(struct ucma_context *ctx)
480{
481 int events_reported;
482
483 /* No new events will be generated after destroying the id. */
484 rdma_destroy_id(ctx->cm_id);
485
486 ucma_cleanup_multicast(ctx);
487
488 /* Cleanup events not yet reported to the user. */
489 mutex_lock(&ctx->file->mut);
490 ucma_cleanup_events(ctx);
491 list_del(&ctx->list);
492 mutex_unlock(&ctx->file->mut);
493
494 events_reported = ctx->events_reported;
495 kfree(ctx);
496 return events_reported;
497}
498
499static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
500 int in_len, int out_len)
501{
502 struct rdma_ucm_destroy_id cmd;
503 struct rdma_ucm_destroy_id_resp resp;
504 struct ucma_context *ctx;
505 int ret = 0;
506
507 if (out_len < sizeof(resp))
508 return -ENOSPC;
509
510 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
511 return -EFAULT;
512
513 mutex_lock(&mut);
514 ctx = _ucma_find_context(cmd.id, file);
515 if (!IS_ERR(ctx))
516 idr_remove(&ctx_idr, ctx->id);
517 mutex_unlock(&mut);
518
519 if (IS_ERR(ctx))
520 return PTR_ERR(ctx);
521
522 ucma_put_ctx(ctx);
523 wait_for_completion(&ctx->comp);
524 resp.events_reported = ucma_free_ctx(ctx);
525
526 if (copy_to_user((void __user *)(unsigned long)cmd.response,
527 &resp, sizeof(resp)))
528 ret = -EFAULT;
529
530 return ret;
531}
532
533static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf,
534 int in_len, int out_len)
535{
536 struct rdma_ucm_bind_addr cmd;
537 struct ucma_context *ctx;
538 int ret;
539
540 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
541 return -EFAULT;
542
543 ctx = ucma_get_ctx(file, cmd.id);
544 if (IS_ERR(ctx))
545 return PTR_ERR(ctx);
546
547 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
548 ucma_put_ctx(ctx);
549 return ret;
550}
551
552static ssize_t ucma_resolve_addr(struct ucma_file *file,
553 const char __user *inbuf,
554 int in_len, int out_len)
555{
556 struct rdma_ucm_resolve_addr cmd;
557 struct ucma_context *ctx;
558 int ret;
559
560 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
561 return -EFAULT;
562
563 ctx = ucma_get_ctx(file, cmd.id);
564 if (IS_ERR(ctx))
565 return PTR_ERR(ctx);
566
567 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
568 (struct sockaddr *) &cmd.dst_addr,
569 cmd.timeout_ms);
570 ucma_put_ctx(ctx);
571 return ret;
572}
573
574static ssize_t ucma_resolve_route(struct ucma_file *file,
575 const char __user *inbuf,
576 int in_len, int out_len)
577{
578 struct rdma_ucm_resolve_route cmd;
579 struct ucma_context *ctx;
580 int ret;
581
582 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
583 return -EFAULT;
584
585 ctx = ucma_get_ctx(file, cmd.id);
586 if (IS_ERR(ctx))
587 return PTR_ERR(ctx);
588
589 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
590 ucma_put_ctx(ctx);
591 return ret;
592}
593
594static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
595 struct rdma_route *route)
596{
597 struct rdma_dev_addr *dev_addr;
598
599 resp->num_paths = route->num_paths;
600 switch (route->num_paths) {
601 case 0:
602 dev_addr = &route->addr.dev_addr;
603 rdma_addr_get_dgid(dev_addr,
604 (union ib_gid *) &resp->ib_route[0].dgid);
605 rdma_addr_get_sgid(dev_addr,
606 (union ib_gid *) &resp->ib_route[0].sgid);
607 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
608 break;
609 case 2:
610 ib_copy_path_rec_to_user(&resp->ib_route[1],
611 &route->path_rec[1]);
612 /* fall through */
613 case 1:
614 ib_copy_path_rec_to_user(&resp->ib_route[0],
615 &route->path_rec[0]);
616 break;
617 default:
618 break;
619 }
620}
621
622static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
623 struct rdma_route *route)
624{
625 struct rdma_dev_addr *dev_addr;
626 struct net_device *dev;
627 u16 vid = 0;
628
629 resp->num_paths = route->num_paths;
630 switch (route->num_paths) {
631 case 0:
632 dev_addr = &route->addr.dev_addr;
633 dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
634 if (dev) {
635 vid = rdma_vlan_dev_vlan_id(dev);
636 dev_put(dev);
637 }
638
639 iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid,
640 dev_addr->dst_dev_addr, vid);
641 iboe_addr_get_sgid(dev_addr,
642 (union ib_gid *) &resp->ib_route[0].sgid);
643 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
644 break;
645 case 2:
646 ib_copy_path_rec_to_user(&resp->ib_route[1],
647 &route->path_rec[1]);
648 /* fall through */
649 case 1:
650 ib_copy_path_rec_to_user(&resp->ib_route[0],
651 &route->path_rec[0]);
652 break;
653 default:
654 break;
655 }
656}
657
658static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
659 struct rdma_route *route)
660{
661 struct rdma_dev_addr *dev_addr;
662
663 dev_addr = &route->addr.dev_addr;
664 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
665 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
666}
667
668static ssize_t ucma_query_route(struct ucma_file *file,
669 const char __user *inbuf,
670 int in_len, int out_len)
671{
672 struct rdma_ucm_query_route cmd;
673 struct rdma_ucm_query_route_resp resp;
674 struct ucma_context *ctx;
675 struct sockaddr *addr;
676 int ret = 0;
677
678 if (out_len < sizeof(resp))
679 return -ENOSPC;
680
681 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
682 return -EFAULT;
683
684 ctx = ucma_get_ctx(file, cmd.id);
685 if (IS_ERR(ctx))
686 return PTR_ERR(ctx);
687
688 memset(&resp, 0, sizeof resp);
689 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
690 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
691 sizeof(struct sockaddr_in) :
692 sizeof(struct sockaddr_in6));
693 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
694 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
695 sizeof(struct sockaddr_in) :
696 sizeof(struct sockaddr_in6));
697 if (!ctx->cm_id->device)
698 goto out;
699
700 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
701 resp.port_num = ctx->cm_id->port_num;
702 switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
703 case RDMA_TRANSPORT_IB:
704 switch (rdma_port_get_link_layer(ctx->cm_id->device,
705 ctx->cm_id->port_num)) {
706 case IB_LINK_LAYER_INFINIBAND:
707 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
708 break;
709 case IB_LINK_LAYER_ETHERNET:
710 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
711 break;
712 default:
713 break;
714 }
715 break;
716 case RDMA_TRANSPORT_IWARP:
717 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
718 break;
719 default:
720 break;
721 }
722
723out:
724 if (copy_to_user((void __user *)(unsigned long)cmd.response,
725 &resp, sizeof(resp)))
726 ret = -EFAULT;
727
728 ucma_put_ctx(ctx);
729 return ret;
730}
731
732static void ucma_copy_conn_param(struct rdma_conn_param *dst,
733 struct rdma_ucm_conn_param *src)
734{
735 dst->private_data = src->private_data;
736 dst->private_data_len = src->private_data_len;
737 dst->responder_resources =src->responder_resources;
738 dst->initiator_depth = src->initiator_depth;
739 dst->flow_control = src->flow_control;
740 dst->retry_count = src->retry_count;
741 dst->rnr_retry_count = src->rnr_retry_count;
742 dst->srq = src->srq;
743 dst->qp_num = src->qp_num;
744}
745
746static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
747 int in_len, int out_len)
748{
749 struct rdma_ucm_connect cmd;
750 struct rdma_conn_param conn_param;
751 struct ucma_context *ctx;
752 int ret;
753
754 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
755 return -EFAULT;
756
757 if (!cmd.conn_param.valid)
758 return -EINVAL;
759
760 ctx = ucma_get_ctx(file, cmd.id);
761 if (IS_ERR(ctx))
762 return PTR_ERR(ctx);
763
764 ucma_copy_conn_param(&conn_param, &cmd.conn_param);
765 ret = rdma_connect(ctx->cm_id, &conn_param);
766 ucma_put_ctx(ctx);
767 return ret;
768}
769
770static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
771 int in_len, int out_len)
772{
773 struct rdma_ucm_listen cmd;
774 struct ucma_context *ctx;
775 int ret;
776
777 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
778 return -EFAULT;
779
780 ctx = ucma_get_ctx(file, cmd.id);
781 if (IS_ERR(ctx))
782 return PTR_ERR(ctx);
783
784 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
785 cmd.backlog : max_backlog;
786 ret = rdma_listen(ctx->cm_id, ctx->backlog);
787 ucma_put_ctx(ctx);
788 return ret;
789}
790
791static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
792 int in_len, int out_len)
793{
794 struct rdma_ucm_accept cmd;
795 struct rdma_conn_param conn_param;
796 struct ucma_context *ctx;
797 int ret;
798
799 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
800 return -EFAULT;
801
802 ctx = ucma_get_ctx(file, cmd.id);
803 if (IS_ERR(ctx))
804 return PTR_ERR(ctx);
805
806 if (cmd.conn_param.valid) {
807 ctx->uid = cmd.uid;
808 ucma_copy_conn_param(&conn_param, &cmd.conn_param);
809 ret = rdma_accept(ctx->cm_id, &conn_param);
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 default:
913 ret = -ENOSYS;
914 }
915
916 return ret;
917}
918
919static int ucma_set_ib_path(struct ucma_context *ctx,
920 struct ib_path_rec_data *path_data, size_t optlen)
921{
922 struct ib_sa_path_rec sa_path;
923 struct rdma_cm_event event;
924 int ret;
925
926 if (optlen % sizeof(*path_data))
927 return -EINVAL;
928
929 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
930 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
931 IB_PATH_BIDIRECTIONAL))
932 break;
933 }
934
935 if (!optlen)
936 return -EINVAL;
937
938 ib_sa_unpack_path(path_data->path_rec, &sa_path);
939 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
940 if (ret)
941 return ret;
942
943 memset(&event, 0, sizeof event);
944 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
945 return ucma_event_handler(ctx->cm_id, &event);
946}
947
948static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
949 void *optval, size_t optlen)
950{
951 int ret;
952
953 switch (optname) {
954 case RDMA_OPTION_IB_PATH:
955 ret = ucma_set_ib_path(ctx, optval, optlen);
956 break;
957 default:
958 ret = -ENOSYS;
959 }
960
961 return ret;
962}
963
964static int ucma_set_option_level(struct ucma_context *ctx, int level,
965 int optname, void *optval, size_t optlen)
966{
967 int ret;
968
969 switch (level) {
970 case RDMA_OPTION_ID:
971 ret = ucma_set_option_id(ctx, optname, optval, optlen);
972 break;
973 case RDMA_OPTION_IB:
974 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
975 break;
976 default:
977 ret = -ENOSYS;
978 }
979
980 return ret;
981}
982
983static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
984 int in_len, int out_len)
985{
986 struct rdma_ucm_set_option cmd;
987 struct ucma_context *ctx;
988 void *optval;
989 int ret;
990
991 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
992 return -EFAULT;
993
994 ctx = ucma_get_ctx(file, cmd.id);
995 if (IS_ERR(ctx))
996 return PTR_ERR(ctx);
997
998 optval = kmalloc(cmd.optlen, GFP_KERNEL);
999 if (!optval) {
1000 ret = -ENOMEM;
1001 goto out1;
1002 }
1003
1004 if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval,
1005 cmd.optlen)) {
1006 ret = -EFAULT;
1007 goto out2;
1008 }
1009
1010 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1011 cmd.optlen);
1012out2:
1013 kfree(optval);
1014out1:
1015 ucma_put_ctx(ctx);
1016 return ret;
1017}
1018
1019static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1020 int in_len, int out_len)
1021{
1022 struct rdma_ucm_notify cmd;
1023 struct ucma_context *ctx;
1024 int ret;
1025
1026 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1027 return -EFAULT;
1028
1029 ctx = ucma_get_ctx(file, cmd.id);
1030 if (IS_ERR(ctx))
1031 return PTR_ERR(ctx);
1032
1033 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1034 ucma_put_ctx(ctx);
1035 return ret;
1036}
1037
1038static ssize_t ucma_join_multicast(struct ucma_file *file,
1039 const char __user *inbuf,
1040 int in_len, int out_len)
1041{
1042 struct rdma_ucm_join_mcast cmd;
1043 struct rdma_ucm_create_id_resp resp;
1044 struct ucma_context *ctx;
1045 struct ucma_multicast *mc;
1046 int ret;
1047
1048 if (out_len < sizeof(resp))
1049 return -ENOSPC;
1050
1051 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1052 return -EFAULT;
1053
1054 ctx = ucma_get_ctx(file, cmd.id);
1055 if (IS_ERR(ctx))
1056 return PTR_ERR(ctx);
1057
1058 mutex_lock(&file->mut);
1059 mc = ucma_alloc_multicast(ctx);
1060 if (!mc) {
1061 ret = -ENOMEM;
1062 goto err1;
1063 }
1064
1065 mc->uid = cmd.uid;
1066 memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr);
1067 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1068 if (ret)
1069 goto err2;
1070
1071 resp.id = mc->id;
1072 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1073 &resp, sizeof(resp))) {
1074 ret = -EFAULT;
1075 goto err3;
1076 }
1077
1078 mutex_unlock(&file->mut);
1079 ucma_put_ctx(ctx);
1080 return 0;
1081
1082err3:
1083 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1084 ucma_cleanup_mc_events(mc);
1085err2:
1086 mutex_lock(&mut);
1087 idr_remove(&multicast_idr, mc->id);
1088 mutex_unlock(&mut);
1089 list_del(&mc->list);
1090 kfree(mc);
1091err1:
1092 mutex_unlock(&file->mut);
1093 ucma_put_ctx(ctx);
1094 return ret;
1095}
1096
1097static ssize_t ucma_leave_multicast(struct ucma_file *file,
1098 const char __user *inbuf,
1099 int in_len, int out_len)
1100{
1101 struct rdma_ucm_destroy_id cmd;
1102 struct rdma_ucm_destroy_id_resp resp;
1103 struct ucma_multicast *mc;
1104 int ret = 0;
1105
1106 if (out_len < sizeof(resp))
1107 return -ENOSPC;
1108
1109 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1110 return -EFAULT;
1111
1112 mutex_lock(&mut);
1113 mc = idr_find(&multicast_idr, cmd.id);
1114 if (!mc)
1115 mc = ERR_PTR(-ENOENT);
1116 else if (mc->ctx->file != file)
1117 mc = ERR_PTR(-EINVAL);
1118 else {
1119 idr_remove(&multicast_idr, mc->id);
1120 atomic_inc(&mc->ctx->ref);
1121 }
1122 mutex_unlock(&mut);
1123
1124 if (IS_ERR(mc)) {
1125 ret = PTR_ERR(mc);
1126 goto out;
1127 }
1128
1129 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1130 mutex_lock(&mc->ctx->file->mut);
1131 ucma_cleanup_mc_events(mc);
1132 list_del(&mc->list);
1133 mutex_unlock(&mc->ctx->file->mut);
1134
1135 ucma_put_ctx(mc->ctx);
1136 resp.events_reported = mc->events_reported;
1137 kfree(mc);
1138
1139 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1140 &resp, sizeof(resp)))
1141 ret = -EFAULT;
1142out:
1143 return ret;
1144}
1145
1146static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1147{
1148 /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1149 if (file1 < file2) {
1150 mutex_lock(&file1->mut);
1151 mutex_lock(&file2->mut);
1152 } else {
1153 mutex_lock(&file2->mut);
1154 mutex_lock(&file1->mut);
1155 }
1156}
1157
1158static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1159{
1160 if (file1 < file2) {
1161 mutex_unlock(&file2->mut);
1162 mutex_unlock(&file1->mut);
1163 } else {
1164 mutex_unlock(&file1->mut);
1165 mutex_unlock(&file2->mut);
1166 }
1167}
1168
1169static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1170{
1171 struct ucma_event *uevent, *tmp;
1172
1173 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1174 if (uevent->ctx == ctx)
1175 list_move_tail(&uevent->list, &file->event_list);
1176}
1177
1178static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1179 const char __user *inbuf,
1180 int in_len, int out_len)
1181{
1182 struct rdma_ucm_migrate_id cmd;
1183 struct rdma_ucm_migrate_resp resp;
1184 struct ucma_context *ctx;
1185 struct file *filp;
1186 struct ucma_file *cur_file;
1187 int ret = 0;
1188
1189 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1190 return -EFAULT;
1191
1192 /* Get current fd to protect against it being closed */
1193 filp = fget(cmd.fd);
1194 if (!filp)
1195 return -ENOENT;
1196
1197 /* Validate current fd and prevent destruction of id. */
1198 ctx = ucma_get_ctx(filp->private_data, cmd.id);
1199 if (IS_ERR(ctx)) {
1200 ret = PTR_ERR(ctx);
1201 goto file_put;
1202 }
1203
1204 cur_file = ctx->file;
1205 if (cur_file == new_file) {
1206 resp.events_reported = ctx->events_reported;
1207 goto response;
1208 }
1209
1210 /*
1211 * Migrate events between fd's, maintaining order, and avoiding new
1212 * events being added before existing events.
1213 */
1214 ucma_lock_files(cur_file, new_file);
1215 mutex_lock(&mut);
1216
1217 list_move_tail(&ctx->list, &new_file->ctx_list);
1218 ucma_move_events(ctx, new_file);
1219 ctx->file = new_file;
1220 resp.events_reported = ctx->events_reported;
1221
1222 mutex_unlock(&mut);
1223 ucma_unlock_files(cur_file, new_file);
1224
1225response:
1226 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1227 &resp, sizeof(resp)))
1228 ret = -EFAULT;
1229
1230 ucma_put_ctx(ctx);
1231file_put:
1232 fput(filp);
1233 return ret;
1234}
1235
1236static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1237 const char __user *inbuf,
1238 int in_len, int out_len) = {
1239 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1240 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1241 [RDMA_USER_CM_CMD_BIND_ADDR] = ucma_bind_addr,
1242 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1243 [RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route,
1244 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1245 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1246 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1247 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1248 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1249 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1250 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1251 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1252 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1253 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1254 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1255 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast,
1256 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1257 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id
1258};
1259
1260static ssize_t ucma_write(struct file *filp, const char __user *buf,
1261 size_t len, loff_t *pos)
1262{
1263 struct ucma_file *file = filp->private_data;
1264 struct rdma_ucm_cmd_hdr hdr;
1265 ssize_t ret;
1266
1267 if (len < sizeof(hdr))
1268 return -EINVAL;
1269
1270 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1271 return -EFAULT;
1272
1273 if (hdr.cmd < 0 || hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1274 return -EINVAL;
1275
1276 if (hdr.in + sizeof(hdr) > len)
1277 return -EINVAL;
1278
1279 if (!ucma_cmd_table[hdr.cmd])
1280 return -ENOSYS;
1281
1282 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1283 if (!ret)
1284 ret = len;
1285
1286 return ret;
1287}
1288
1289static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1290{
1291 struct ucma_file *file = filp->private_data;
1292 unsigned int mask = 0;
1293
1294 poll_wait(filp, &file->poll_wait, wait);
1295
1296 if (!list_empty(&file->event_list))
1297 mask = POLLIN | POLLRDNORM;
1298
1299 return mask;
1300}
1301
1302/*
1303 * ucma_open() does not need the BKL:
1304 *
1305 * - no global state is referred to;
1306 * - there is no ioctl method to race against;
1307 * - no further module initialization is required for open to work
1308 * after the device is registered.
1309 */
1310static int ucma_open(struct inode *inode, struct file *filp)
1311{
1312 struct ucma_file *file;
1313
1314 file = kmalloc(sizeof *file, GFP_KERNEL);
1315 if (!file)
1316 return -ENOMEM;
1317
1318 INIT_LIST_HEAD(&file->event_list);
1319 INIT_LIST_HEAD(&file->ctx_list);
1320 init_waitqueue_head(&file->poll_wait);
1321 mutex_init(&file->mut);
1322
1323 filp->private_data = file;
1324 file->filp = filp;
1325
1326 return nonseekable_open(inode, filp);
1327}
1328
1329static int ucma_close(struct inode *inode, struct file *filp)
1330{
1331 struct ucma_file *file = filp->private_data;
1332 struct ucma_context *ctx, *tmp;
1333
1334 mutex_lock(&file->mut);
1335 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1336 mutex_unlock(&file->mut);
1337
1338 mutex_lock(&mut);
1339 idr_remove(&ctx_idr, ctx->id);
1340 mutex_unlock(&mut);
1341
1342 ucma_free_ctx(ctx);
1343 mutex_lock(&file->mut);
1344 }
1345 mutex_unlock(&file->mut);
1346 kfree(file);
1347 return 0;
1348}
1349
1350static const struct file_operations ucma_fops = {
1351 .owner = THIS_MODULE,
1352 .open = ucma_open,
1353 .release = ucma_close,
1354 .write = ucma_write,
1355 .poll = ucma_poll,
1356 .llseek = no_llseek,
1357};
1358
1359static struct miscdevice ucma_misc = {
1360 .minor = MISC_DYNAMIC_MINOR,
1361 .name = "rdma_cm",
1362 .nodename = "infiniband/rdma_cm",
1363 .mode = 0666,
1364 .fops = &ucma_fops,
1365};
1366
1367static ssize_t show_abi_version(struct device *dev,
1368 struct device_attribute *attr,
1369 char *buf)
1370{
1371 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1372}
1373static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1374
1375static int __init ucma_init(void)
1376{
1377 int ret;
1378
1379 ret = misc_register(&ucma_misc);
1380 if (ret)
1381 return ret;
1382
1383 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1384 if (ret) {
1385 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1386 goto err1;
1387 }
1388
1389 ucma_ctl_table_hdr = register_sysctl_paths(ucma_ctl_path, ucma_ctl_table);
1390 if (!ucma_ctl_table_hdr) {
1391 printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n");
1392 ret = -ENOMEM;
1393 goto err2;
1394 }
1395 return 0;
1396err2:
1397 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1398err1:
1399 misc_deregister(&ucma_misc);
1400 return ret;
1401}
1402
1403static void __exit ucma_cleanup(void)
1404{
1405 unregister_sysctl_table(ucma_ctl_table_hdr);
1406 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1407 misc_deregister(&ucma_misc);
1408 idr_destroy(&ctx_idr);
1409}
1410
1411module_init(ucma_init);
1412module_exit(ucma_cleanup);
1/*
2 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
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#include <linux/nsproxy.h>
46
47#include <linux/nospec.h>
48
49#include <rdma/rdma_user_cm.h>
50#include <rdma/ib_marshall.h>
51#include <rdma/rdma_cm.h>
52#include <rdma/rdma_cm_ib.h>
53#include <rdma/ib_addr.h>
54#include <rdma/ib.h>
55#include <rdma/rdma_netlink.h>
56#include "core_priv.h"
57
58MODULE_AUTHOR("Sean Hefty");
59MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
60MODULE_LICENSE("Dual BSD/GPL");
61
62static unsigned int max_backlog = 1024;
63
64static struct ctl_table_header *ucma_ctl_table_hdr;
65static struct ctl_table ucma_ctl_table[] = {
66 {
67 .procname = "max_backlog",
68 .data = &max_backlog,
69 .maxlen = sizeof max_backlog,
70 .mode = 0644,
71 .proc_handler = proc_dointvec,
72 },
73 { }
74};
75
76struct ucma_file {
77 struct mutex mut;
78 struct file *filp;
79 struct list_head ctx_list;
80 struct list_head event_list;
81 wait_queue_head_t poll_wait;
82 struct workqueue_struct *close_wq;
83};
84
85struct ucma_context {
86 u32 id;
87 struct completion comp;
88 atomic_t ref;
89 int events_reported;
90 int backlog;
91
92 struct ucma_file *file;
93 struct rdma_cm_id *cm_id;
94 u64 uid;
95
96 struct list_head list;
97 struct list_head mc_list;
98 /* mark that device is in process of destroying the internal HW
99 * resources, protected by the ctx_table lock
100 */
101 int closing;
102 /* sync between removal event and id destroy, protected by file mut */
103 int destroying;
104 struct work_struct close_work;
105};
106
107struct ucma_multicast {
108 struct ucma_context *ctx;
109 u32 id;
110 int events_reported;
111
112 u64 uid;
113 u8 join_state;
114 struct list_head list;
115 struct sockaddr_storage addr;
116};
117
118struct ucma_event {
119 struct ucma_context *ctx;
120 struct ucma_multicast *mc;
121 struct list_head list;
122 struct rdma_cm_id *cm_id;
123 struct rdma_ucm_event_resp resp;
124 struct work_struct close_work;
125};
126
127static DEFINE_XARRAY_ALLOC(ctx_table);
128static DEFINE_XARRAY_ALLOC(multicast_table);
129
130static const struct file_operations ucma_fops;
131
132static inline struct ucma_context *_ucma_find_context(int id,
133 struct ucma_file *file)
134{
135 struct ucma_context *ctx;
136
137 ctx = xa_load(&ctx_table, id);
138 if (!ctx)
139 ctx = ERR_PTR(-ENOENT);
140 else if (ctx->file != file || !ctx->cm_id)
141 ctx = ERR_PTR(-EINVAL);
142 return ctx;
143}
144
145static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
146{
147 struct ucma_context *ctx;
148
149 xa_lock(&ctx_table);
150 ctx = _ucma_find_context(id, file);
151 if (!IS_ERR(ctx)) {
152 if (ctx->closing)
153 ctx = ERR_PTR(-EIO);
154 else
155 atomic_inc(&ctx->ref);
156 }
157 xa_unlock(&ctx_table);
158 return ctx;
159}
160
161static void ucma_put_ctx(struct ucma_context *ctx)
162{
163 if (atomic_dec_and_test(&ctx->ref))
164 complete(&ctx->comp);
165}
166
167/*
168 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
169 * CM_ID is bound.
170 */
171static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
172{
173 struct ucma_context *ctx = ucma_get_ctx(file, id);
174
175 if (IS_ERR(ctx))
176 return ctx;
177 if (!ctx->cm_id->device) {
178 ucma_put_ctx(ctx);
179 return ERR_PTR(-EINVAL);
180 }
181 return ctx;
182}
183
184static void ucma_close_event_id(struct work_struct *work)
185{
186 struct ucma_event *uevent_close = container_of(work, struct ucma_event, close_work);
187
188 rdma_destroy_id(uevent_close->cm_id);
189 kfree(uevent_close);
190}
191
192static void ucma_close_id(struct work_struct *work)
193{
194 struct ucma_context *ctx = container_of(work, struct ucma_context, close_work);
195
196 /* once all inflight tasks are finished, we close all underlying
197 * resources. The context is still alive till its explicit destryoing
198 * by its creator.
199 */
200 ucma_put_ctx(ctx);
201 wait_for_completion(&ctx->comp);
202 /* No new events will be generated after destroying the id. */
203 rdma_destroy_id(ctx->cm_id);
204}
205
206static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
207{
208 struct ucma_context *ctx;
209
210 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
211 if (!ctx)
212 return NULL;
213
214 INIT_WORK(&ctx->close_work, ucma_close_id);
215 atomic_set(&ctx->ref, 1);
216 init_completion(&ctx->comp);
217 INIT_LIST_HEAD(&ctx->mc_list);
218 ctx->file = file;
219
220 if (xa_alloc(&ctx_table, &ctx->id, ctx, xa_limit_32b, GFP_KERNEL))
221 goto error;
222
223 list_add_tail(&ctx->list, &file->ctx_list);
224 return ctx;
225
226error:
227 kfree(ctx);
228 return NULL;
229}
230
231static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
232{
233 struct ucma_multicast *mc;
234
235 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
236 if (!mc)
237 return NULL;
238
239 mc->ctx = ctx;
240 if (xa_alloc(&multicast_table, &mc->id, NULL, xa_limit_32b, GFP_KERNEL))
241 goto error;
242
243 list_add_tail(&mc->list, &ctx->mc_list);
244 return mc;
245
246error:
247 kfree(mc);
248 return NULL;
249}
250
251static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
252 struct rdma_conn_param *src)
253{
254 if (src->private_data_len)
255 memcpy(dst->private_data, src->private_data,
256 src->private_data_len);
257 dst->private_data_len = src->private_data_len;
258 dst->responder_resources =src->responder_resources;
259 dst->initiator_depth = src->initiator_depth;
260 dst->flow_control = src->flow_control;
261 dst->retry_count = src->retry_count;
262 dst->rnr_retry_count = src->rnr_retry_count;
263 dst->srq = src->srq;
264 dst->qp_num = src->qp_num;
265}
266
267static void ucma_copy_ud_event(struct ib_device *device,
268 struct rdma_ucm_ud_param *dst,
269 struct rdma_ud_param *src)
270{
271 if (src->private_data_len)
272 memcpy(dst->private_data, src->private_data,
273 src->private_data_len);
274 dst->private_data_len = src->private_data_len;
275 ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
276 dst->qp_num = src->qp_num;
277 dst->qkey = src->qkey;
278}
279
280static void ucma_set_event_context(struct ucma_context *ctx,
281 struct rdma_cm_event *event,
282 struct ucma_event *uevent)
283{
284 uevent->ctx = ctx;
285 switch (event->event) {
286 case RDMA_CM_EVENT_MULTICAST_JOIN:
287 case RDMA_CM_EVENT_MULTICAST_ERROR:
288 uevent->mc = (struct ucma_multicast *)
289 event->param.ud.private_data;
290 uevent->resp.uid = uevent->mc->uid;
291 uevent->resp.id = uevent->mc->id;
292 break;
293 default:
294 uevent->resp.uid = ctx->uid;
295 uevent->resp.id = ctx->id;
296 break;
297 }
298}
299
300/* Called with file->mut locked for the relevant context. */
301static void ucma_removal_event_handler(struct rdma_cm_id *cm_id)
302{
303 struct ucma_context *ctx = cm_id->context;
304 struct ucma_event *con_req_eve;
305 int event_found = 0;
306
307 if (ctx->destroying)
308 return;
309
310 /* only if context is pointing to cm_id that it owns it and can be
311 * queued to be closed, otherwise that cm_id is an inflight one that
312 * is part of that context event list pending to be detached and
313 * reattached to its new context as part of ucma_get_event,
314 * handled separately below.
315 */
316 if (ctx->cm_id == cm_id) {
317 xa_lock(&ctx_table);
318 ctx->closing = 1;
319 xa_unlock(&ctx_table);
320 queue_work(ctx->file->close_wq, &ctx->close_work);
321 return;
322 }
323
324 list_for_each_entry(con_req_eve, &ctx->file->event_list, list) {
325 if (con_req_eve->cm_id == cm_id &&
326 con_req_eve->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
327 list_del(&con_req_eve->list);
328 INIT_WORK(&con_req_eve->close_work, ucma_close_event_id);
329 queue_work(ctx->file->close_wq, &con_req_eve->close_work);
330 event_found = 1;
331 break;
332 }
333 }
334 if (!event_found)
335 pr_err("ucma_removal_event_handler: warning: connect request event wasn't found\n");
336}
337
338static int ucma_event_handler(struct rdma_cm_id *cm_id,
339 struct rdma_cm_event *event)
340{
341 struct ucma_event *uevent;
342 struct ucma_context *ctx = cm_id->context;
343 int ret = 0;
344
345 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
346 if (!uevent)
347 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
348
349 mutex_lock(&ctx->file->mut);
350 uevent->cm_id = cm_id;
351 ucma_set_event_context(ctx, event, uevent);
352 uevent->resp.event = event->event;
353 uevent->resp.status = event->status;
354 if (cm_id->qp_type == IB_QPT_UD)
355 ucma_copy_ud_event(cm_id->device, &uevent->resp.param.ud,
356 &event->param.ud);
357 else
358 ucma_copy_conn_event(&uevent->resp.param.conn,
359 &event->param.conn);
360
361 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
362 if (!ctx->backlog) {
363 ret = -ENOMEM;
364 kfree(uevent);
365 goto out;
366 }
367 ctx->backlog--;
368 } else if (!ctx->uid || ctx->cm_id != cm_id) {
369 /*
370 * We ignore events for new connections until userspace has set
371 * their context. This can only happen if an error occurs on a
372 * new connection before the user accepts it. This is okay,
373 * since the accept will just fail later. However, we do need
374 * to release the underlying HW resources in case of a device
375 * removal event.
376 */
377 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
378 ucma_removal_event_handler(cm_id);
379
380 kfree(uevent);
381 goto out;
382 }
383
384 list_add_tail(&uevent->list, &ctx->file->event_list);
385 wake_up_interruptible(&ctx->file->poll_wait);
386 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
387 ucma_removal_event_handler(cm_id);
388out:
389 mutex_unlock(&ctx->file->mut);
390 return ret;
391}
392
393static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
394 int in_len, int out_len)
395{
396 struct ucma_context *ctx;
397 struct rdma_ucm_get_event cmd;
398 struct ucma_event *uevent;
399 int ret = 0;
400
401 /*
402 * Old 32 bit user space does not send the 4 byte padding in the
403 * reserved field. We don't care, allow it to keep working.
404 */
405 if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved))
406 return -ENOSPC;
407
408 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
409 return -EFAULT;
410
411 mutex_lock(&file->mut);
412 while (list_empty(&file->event_list)) {
413 mutex_unlock(&file->mut);
414
415 if (file->filp->f_flags & O_NONBLOCK)
416 return -EAGAIN;
417
418 if (wait_event_interruptible(file->poll_wait,
419 !list_empty(&file->event_list)))
420 return -ERESTARTSYS;
421
422 mutex_lock(&file->mut);
423 }
424
425 uevent = list_entry(file->event_list.next, struct ucma_event, list);
426
427 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
428 ctx = ucma_alloc_ctx(file);
429 if (!ctx) {
430 ret = -ENOMEM;
431 goto done;
432 }
433 uevent->ctx->backlog++;
434 ctx->cm_id = uevent->cm_id;
435 ctx->cm_id->context = ctx;
436 uevent->resp.id = ctx->id;
437 }
438
439 if (copy_to_user(u64_to_user_ptr(cmd.response),
440 &uevent->resp,
441 min_t(size_t, out_len, sizeof(uevent->resp)))) {
442 ret = -EFAULT;
443 goto done;
444 }
445
446 list_del(&uevent->list);
447 uevent->ctx->events_reported++;
448 if (uevent->mc)
449 uevent->mc->events_reported++;
450 kfree(uevent);
451done:
452 mutex_unlock(&file->mut);
453 return ret;
454}
455
456static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
457{
458 switch (cmd->ps) {
459 case RDMA_PS_TCP:
460 *qp_type = IB_QPT_RC;
461 return 0;
462 case RDMA_PS_UDP:
463 case RDMA_PS_IPOIB:
464 *qp_type = IB_QPT_UD;
465 return 0;
466 case RDMA_PS_IB:
467 *qp_type = cmd->qp_type;
468 return 0;
469 default:
470 return -EINVAL;
471 }
472}
473
474static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
475 int in_len, int out_len)
476{
477 struct rdma_ucm_create_id cmd;
478 struct rdma_ucm_create_id_resp resp;
479 struct ucma_context *ctx;
480 struct rdma_cm_id *cm_id;
481 enum ib_qp_type qp_type;
482 int ret;
483
484 if (out_len < sizeof(resp))
485 return -ENOSPC;
486
487 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
488 return -EFAULT;
489
490 ret = ucma_get_qp_type(&cmd, &qp_type);
491 if (ret)
492 return ret;
493
494 mutex_lock(&file->mut);
495 ctx = ucma_alloc_ctx(file);
496 mutex_unlock(&file->mut);
497 if (!ctx)
498 return -ENOMEM;
499
500 ctx->uid = cmd.uid;
501 cm_id = __rdma_create_id(current->nsproxy->net_ns,
502 ucma_event_handler, ctx, cmd.ps, qp_type, NULL);
503 if (IS_ERR(cm_id)) {
504 ret = PTR_ERR(cm_id);
505 goto err1;
506 }
507
508 resp.id = ctx->id;
509 if (copy_to_user(u64_to_user_ptr(cmd.response),
510 &resp, sizeof(resp))) {
511 ret = -EFAULT;
512 goto err2;
513 }
514
515 ctx->cm_id = cm_id;
516 return 0;
517
518err2:
519 rdma_destroy_id(cm_id);
520err1:
521 xa_erase(&ctx_table, ctx->id);
522 mutex_lock(&file->mut);
523 list_del(&ctx->list);
524 mutex_unlock(&file->mut);
525 kfree(ctx);
526 return ret;
527}
528
529static void ucma_cleanup_multicast(struct ucma_context *ctx)
530{
531 struct ucma_multicast *mc, *tmp;
532
533 mutex_lock(&ctx->file->mut);
534 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
535 list_del(&mc->list);
536 xa_erase(&multicast_table, mc->id);
537 kfree(mc);
538 }
539 mutex_unlock(&ctx->file->mut);
540}
541
542static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
543{
544 struct ucma_event *uevent, *tmp;
545
546 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
547 if (uevent->mc != mc)
548 continue;
549
550 list_del(&uevent->list);
551 kfree(uevent);
552 }
553}
554
555/*
556 * ucma_free_ctx is called after the underlying rdma CM-ID is destroyed. At
557 * this point, no new events will be reported from the hardware. However, we
558 * still need to cleanup the UCMA context for this ID. Specifically, there
559 * might be events that have not yet been consumed by the user space software.
560 * These might include pending connect requests which we have not completed
561 * processing. We cannot call rdma_destroy_id while holding the lock of the
562 * context (file->mut), as it might cause a deadlock. We therefore extract all
563 * relevant events from the context pending events list while holding the
564 * mutex. After that we release them as needed.
565 */
566static int ucma_free_ctx(struct ucma_context *ctx)
567{
568 int events_reported;
569 struct ucma_event *uevent, *tmp;
570 LIST_HEAD(list);
571
572
573 ucma_cleanup_multicast(ctx);
574
575 /* Cleanup events not yet reported to the user. */
576 mutex_lock(&ctx->file->mut);
577 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
578 if (uevent->ctx == ctx)
579 list_move_tail(&uevent->list, &list);
580 }
581 list_del(&ctx->list);
582 mutex_unlock(&ctx->file->mut);
583
584 list_for_each_entry_safe(uevent, tmp, &list, list) {
585 list_del(&uevent->list);
586 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
587 rdma_destroy_id(uevent->cm_id);
588 kfree(uevent);
589 }
590
591 events_reported = ctx->events_reported;
592 kfree(ctx);
593 return events_reported;
594}
595
596static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
597 int in_len, int out_len)
598{
599 struct rdma_ucm_destroy_id cmd;
600 struct rdma_ucm_destroy_id_resp resp;
601 struct ucma_context *ctx;
602 int ret = 0;
603
604 if (out_len < sizeof(resp))
605 return -ENOSPC;
606
607 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
608 return -EFAULT;
609
610 xa_lock(&ctx_table);
611 ctx = _ucma_find_context(cmd.id, file);
612 if (!IS_ERR(ctx))
613 __xa_erase(&ctx_table, ctx->id);
614 xa_unlock(&ctx_table);
615
616 if (IS_ERR(ctx))
617 return PTR_ERR(ctx);
618
619 mutex_lock(&ctx->file->mut);
620 ctx->destroying = 1;
621 mutex_unlock(&ctx->file->mut);
622
623 flush_workqueue(ctx->file->close_wq);
624 /* At this point it's guaranteed that there is no inflight
625 * closing task */
626 xa_lock(&ctx_table);
627 if (!ctx->closing) {
628 xa_unlock(&ctx_table);
629 ucma_put_ctx(ctx);
630 wait_for_completion(&ctx->comp);
631 rdma_destroy_id(ctx->cm_id);
632 } else {
633 xa_unlock(&ctx_table);
634 }
635
636 resp.events_reported = ucma_free_ctx(ctx);
637 if (copy_to_user(u64_to_user_ptr(cmd.response),
638 &resp, sizeof(resp)))
639 ret = -EFAULT;
640
641 return ret;
642}
643
644static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
645 int in_len, int out_len)
646{
647 struct rdma_ucm_bind_ip cmd;
648 struct ucma_context *ctx;
649 int ret;
650
651 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
652 return -EFAULT;
653
654 if (!rdma_addr_size_in6(&cmd.addr))
655 return -EINVAL;
656
657 ctx = ucma_get_ctx(file, cmd.id);
658 if (IS_ERR(ctx))
659 return PTR_ERR(ctx);
660
661 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
662 ucma_put_ctx(ctx);
663 return ret;
664}
665
666static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
667 int in_len, int out_len)
668{
669 struct rdma_ucm_bind cmd;
670 struct ucma_context *ctx;
671 int ret;
672
673 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
674 return -EFAULT;
675
676 if (cmd.reserved || !cmd.addr_size ||
677 cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
678 return -EINVAL;
679
680 ctx = ucma_get_ctx(file, cmd.id);
681 if (IS_ERR(ctx))
682 return PTR_ERR(ctx);
683
684 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
685 ucma_put_ctx(ctx);
686 return ret;
687}
688
689static ssize_t ucma_resolve_ip(struct ucma_file *file,
690 const char __user *inbuf,
691 int in_len, int out_len)
692{
693 struct rdma_ucm_resolve_ip cmd;
694 struct ucma_context *ctx;
695 int ret;
696
697 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
698 return -EFAULT;
699
700 if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
701 !rdma_addr_size_in6(&cmd.dst_addr))
702 return -EINVAL;
703
704 ctx = ucma_get_ctx(file, cmd.id);
705 if (IS_ERR(ctx))
706 return PTR_ERR(ctx);
707
708 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
709 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
710 ucma_put_ctx(ctx);
711 return ret;
712}
713
714static ssize_t ucma_resolve_addr(struct ucma_file *file,
715 const char __user *inbuf,
716 int in_len, int out_len)
717{
718 struct rdma_ucm_resolve_addr cmd;
719 struct ucma_context *ctx;
720 int ret;
721
722 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
723 return -EFAULT;
724
725 if (cmd.reserved ||
726 (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
727 !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
728 return -EINVAL;
729
730 ctx = ucma_get_ctx(file, cmd.id);
731 if (IS_ERR(ctx))
732 return PTR_ERR(ctx);
733
734 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
735 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
736 ucma_put_ctx(ctx);
737 return ret;
738}
739
740static ssize_t ucma_resolve_route(struct ucma_file *file,
741 const char __user *inbuf,
742 int in_len, int out_len)
743{
744 struct rdma_ucm_resolve_route cmd;
745 struct ucma_context *ctx;
746 int ret;
747
748 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
749 return -EFAULT;
750
751 ctx = ucma_get_ctx_dev(file, cmd.id);
752 if (IS_ERR(ctx))
753 return PTR_ERR(ctx);
754
755 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
756 ucma_put_ctx(ctx);
757 return ret;
758}
759
760static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
761 struct rdma_route *route)
762{
763 struct rdma_dev_addr *dev_addr;
764
765 resp->num_paths = route->num_paths;
766 switch (route->num_paths) {
767 case 0:
768 dev_addr = &route->addr.dev_addr;
769 rdma_addr_get_dgid(dev_addr,
770 (union ib_gid *) &resp->ib_route[0].dgid);
771 rdma_addr_get_sgid(dev_addr,
772 (union ib_gid *) &resp->ib_route[0].sgid);
773 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
774 break;
775 case 2:
776 ib_copy_path_rec_to_user(&resp->ib_route[1],
777 &route->path_rec[1]);
778 /* fall through */
779 case 1:
780 ib_copy_path_rec_to_user(&resp->ib_route[0],
781 &route->path_rec[0]);
782 break;
783 default:
784 break;
785 }
786}
787
788static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
789 struct rdma_route *route)
790{
791
792 resp->num_paths = route->num_paths;
793 switch (route->num_paths) {
794 case 0:
795 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
796 (union ib_gid *)&resp->ib_route[0].dgid);
797 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
798 (union ib_gid *)&resp->ib_route[0].sgid);
799 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
800 break;
801 case 2:
802 ib_copy_path_rec_to_user(&resp->ib_route[1],
803 &route->path_rec[1]);
804 /* fall through */
805 case 1:
806 ib_copy_path_rec_to_user(&resp->ib_route[0],
807 &route->path_rec[0]);
808 break;
809 default:
810 break;
811 }
812}
813
814static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
815 struct rdma_route *route)
816{
817 struct rdma_dev_addr *dev_addr;
818
819 dev_addr = &route->addr.dev_addr;
820 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
821 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
822}
823
824static ssize_t ucma_query_route(struct ucma_file *file,
825 const char __user *inbuf,
826 int in_len, int out_len)
827{
828 struct rdma_ucm_query cmd;
829 struct rdma_ucm_query_route_resp resp;
830 struct ucma_context *ctx;
831 struct sockaddr *addr;
832 int ret = 0;
833
834 if (out_len < sizeof(resp))
835 return -ENOSPC;
836
837 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
838 return -EFAULT;
839
840 ctx = ucma_get_ctx(file, cmd.id);
841 if (IS_ERR(ctx))
842 return PTR_ERR(ctx);
843
844 memset(&resp, 0, sizeof resp);
845 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
846 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
847 sizeof(struct sockaddr_in) :
848 sizeof(struct sockaddr_in6));
849 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
850 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
851 sizeof(struct sockaddr_in) :
852 sizeof(struct sockaddr_in6));
853 if (!ctx->cm_id->device)
854 goto out;
855
856 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
857 resp.port_num = ctx->cm_id->port_num;
858
859 if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
860 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
861 else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
862 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
863 else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
864 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
865
866out:
867 if (copy_to_user(u64_to_user_ptr(cmd.response),
868 &resp, sizeof(resp)))
869 ret = -EFAULT;
870
871 ucma_put_ctx(ctx);
872 return ret;
873}
874
875static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
876 struct rdma_ucm_query_addr_resp *resp)
877{
878 if (!cm_id->device)
879 return;
880
881 resp->node_guid = (__force __u64) cm_id->device->node_guid;
882 resp->port_num = cm_id->port_num;
883 resp->pkey = (__force __u16) cpu_to_be16(
884 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
885}
886
887static ssize_t ucma_query_addr(struct ucma_context *ctx,
888 void __user *response, int out_len)
889{
890 struct rdma_ucm_query_addr_resp resp;
891 struct sockaddr *addr;
892 int ret = 0;
893
894 if (out_len < sizeof(resp))
895 return -ENOSPC;
896
897 memset(&resp, 0, sizeof resp);
898
899 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
900 resp.src_size = rdma_addr_size(addr);
901 memcpy(&resp.src_addr, addr, resp.src_size);
902
903 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
904 resp.dst_size = rdma_addr_size(addr);
905 memcpy(&resp.dst_addr, addr, resp.dst_size);
906
907 ucma_query_device_addr(ctx->cm_id, &resp);
908
909 if (copy_to_user(response, &resp, sizeof(resp)))
910 ret = -EFAULT;
911
912 return ret;
913}
914
915static ssize_t ucma_query_path(struct ucma_context *ctx,
916 void __user *response, int out_len)
917{
918 struct rdma_ucm_query_path_resp *resp;
919 int i, ret = 0;
920
921 if (out_len < sizeof(*resp))
922 return -ENOSPC;
923
924 resp = kzalloc(out_len, GFP_KERNEL);
925 if (!resp)
926 return -ENOMEM;
927
928 resp->num_paths = ctx->cm_id->route.num_paths;
929 for (i = 0, out_len -= sizeof(*resp);
930 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
931 i++, out_len -= sizeof(struct ib_path_rec_data)) {
932 struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
933
934 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
935 IB_PATH_BIDIRECTIONAL;
936 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
937 struct sa_path_rec ib;
938
939 sa_convert_path_opa_to_ib(&ib, rec);
940 ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
941
942 } else {
943 ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
944 }
945 }
946
947 if (copy_to_user(response, resp, struct_size(resp, path_data, i)))
948 ret = -EFAULT;
949
950 kfree(resp);
951 return ret;
952}
953
954static ssize_t ucma_query_gid(struct ucma_context *ctx,
955 void __user *response, int out_len)
956{
957 struct rdma_ucm_query_addr_resp resp;
958 struct sockaddr_ib *addr;
959 int ret = 0;
960
961 if (out_len < sizeof(resp))
962 return -ENOSPC;
963
964 memset(&resp, 0, sizeof resp);
965
966 ucma_query_device_addr(ctx->cm_id, &resp);
967
968 addr = (struct sockaddr_ib *) &resp.src_addr;
969 resp.src_size = sizeof(*addr);
970 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
971 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
972 } else {
973 addr->sib_family = AF_IB;
974 addr->sib_pkey = (__force __be16) resp.pkey;
975 rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
976 NULL);
977 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
978 &ctx->cm_id->route.addr.src_addr);
979 }
980
981 addr = (struct sockaddr_ib *) &resp.dst_addr;
982 resp.dst_size = sizeof(*addr);
983 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
984 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
985 } else {
986 addr->sib_family = AF_IB;
987 addr->sib_pkey = (__force __be16) resp.pkey;
988 rdma_read_gids(ctx->cm_id, NULL,
989 (union ib_gid *)&addr->sib_addr);
990 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
991 &ctx->cm_id->route.addr.dst_addr);
992 }
993
994 if (copy_to_user(response, &resp, sizeof(resp)))
995 ret = -EFAULT;
996
997 return ret;
998}
999
1000static ssize_t ucma_query(struct ucma_file *file,
1001 const char __user *inbuf,
1002 int in_len, int out_len)
1003{
1004 struct rdma_ucm_query cmd;
1005 struct ucma_context *ctx;
1006 void __user *response;
1007 int ret;
1008
1009 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1010 return -EFAULT;
1011
1012 response = u64_to_user_ptr(cmd.response);
1013 ctx = ucma_get_ctx(file, cmd.id);
1014 if (IS_ERR(ctx))
1015 return PTR_ERR(ctx);
1016
1017 switch (cmd.option) {
1018 case RDMA_USER_CM_QUERY_ADDR:
1019 ret = ucma_query_addr(ctx, response, out_len);
1020 break;
1021 case RDMA_USER_CM_QUERY_PATH:
1022 ret = ucma_query_path(ctx, response, out_len);
1023 break;
1024 case RDMA_USER_CM_QUERY_GID:
1025 ret = ucma_query_gid(ctx, response, out_len);
1026 break;
1027 default:
1028 ret = -ENOSYS;
1029 break;
1030 }
1031
1032 ucma_put_ctx(ctx);
1033 return ret;
1034}
1035
1036static void ucma_copy_conn_param(struct rdma_cm_id *id,
1037 struct rdma_conn_param *dst,
1038 struct rdma_ucm_conn_param *src)
1039{
1040 dst->private_data = src->private_data;
1041 dst->private_data_len = src->private_data_len;
1042 dst->responder_resources =src->responder_resources;
1043 dst->initiator_depth = src->initiator_depth;
1044 dst->flow_control = src->flow_control;
1045 dst->retry_count = src->retry_count;
1046 dst->rnr_retry_count = src->rnr_retry_count;
1047 dst->srq = src->srq;
1048 dst->qp_num = src->qp_num;
1049 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1050}
1051
1052static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1053 int in_len, int out_len)
1054{
1055 struct rdma_ucm_connect cmd;
1056 struct rdma_conn_param conn_param;
1057 struct ucma_context *ctx;
1058 int ret;
1059
1060 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1061 return -EFAULT;
1062
1063 if (!cmd.conn_param.valid)
1064 return -EINVAL;
1065
1066 ctx = ucma_get_ctx_dev(file, cmd.id);
1067 if (IS_ERR(ctx))
1068 return PTR_ERR(ctx);
1069
1070 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1071 ret = rdma_connect(ctx->cm_id, &conn_param);
1072 ucma_put_ctx(ctx);
1073 return ret;
1074}
1075
1076static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1077 int in_len, int out_len)
1078{
1079 struct rdma_ucm_listen cmd;
1080 struct ucma_context *ctx;
1081 int ret;
1082
1083 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1084 return -EFAULT;
1085
1086 ctx = ucma_get_ctx(file, cmd.id);
1087 if (IS_ERR(ctx))
1088 return PTR_ERR(ctx);
1089
1090 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
1091 cmd.backlog : max_backlog;
1092 ret = rdma_listen(ctx->cm_id, ctx->backlog);
1093 ucma_put_ctx(ctx);
1094 return ret;
1095}
1096
1097static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1098 int in_len, int out_len)
1099{
1100 struct rdma_ucm_accept cmd;
1101 struct rdma_conn_param conn_param;
1102 struct ucma_context *ctx;
1103 int ret;
1104
1105 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1106 return -EFAULT;
1107
1108 ctx = ucma_get_ctx_dev(file, cmd.id);
1109 if (IS_ERR(ctx))
1110 return PTR_ERR(ctx);
1111
1112 if (cmd.conn_param.valid) {
1113 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1114 mutex_lock(&file->mut);
1115 ret = __rdma_accept(ctx->cm_id, &conn_param, NULL);
1116 if (!ret)
1117 ctx->uid = cmd.uid;
1118 mutex_unlock(&file->mut);
1119 } else
1120 ret = __rdma_accept(ctx->cm_id, NULL, NULL);
1121
1122 ucma_put_ctx(ctx);
1123 return ret;
1124}
1125
1126static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1127 int in_len, int out_len)
1128{
1129 struct rdma_ucm_reject cmd;
1130 struct ucma_context *ctx;
1131 int ret;
1132
1133 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1134 return -EFAULT;
1135
1136 ctx = ucma_get_ctx_dev(file, cmd.id);
1137 if (IS_ERR(ctx))
1138 return PTR_ERR(ctx);
1139
1140 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
1141 ucma_put_ctx(ctx);
1142 return ret;
1143}
1144
1145static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1146 int in_len, int out_len)
1147{
1148 struct rdma_ucm_disconnect cmd;
1149 struct ucma_context *ctx;
1150 int ret;
1151
1152 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1153 return -EFAULT;
1154
1155 ctx = ucma_get_ctx_dev(file, cmd.id);
1156 if (IS_ERR(ctx))
1157 return PTR_ERR(ctx);
1158
1159 ret = rdma_disconnect(ctx->cm_id);
1160 ucma_put_ctx(ctx);
1161 return ret;
1162}
1163
1164static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1165 const char __user *inbuf,
1166 int in_len, int out_len)
1167{
1168 struct rdma_ucm_init_qp_attr cmd;
1169 struct ib_uverbs_qp_attr resp;
1170 struct ucma_context *ctx;
1171 struct ib_qp_attr qp_attr;
1172 int ret;
1173
1174 if (out_len < sizeof(resp))
1175 return -ENOSPC;
1176
1177 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1178 return -EFAULT;
1179
1180 if (cmd.qp_state > IB_QPS_ERR)
1181 return -EINVAL;
1182
1183 ctx = ucma_get_ctx_dev(file, cmd.id);
1184 if (IS_ERR(ctx))
1185 return PTR_ERR(ctx);
1186
1187 resp.qp_attr_mask = 0;
1188 memset(&qp_attr, 0, sizeof qp_attr);
1189 qp_attr.qp_state = cmd.qp_state;
1190 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1191 if (ret)
1192 goto out;
1193
1194 ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1195 if (copy_to_user(u64_to_user_ptr(cmd.response),
1196 &resp, sizeof(resp)))
1197 ret = -EFAULT;
1198
1199out:
1200 ucma_put_ctx(ctx);
1201 return ret;
1202}
1203
1204static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1205 void *optval, size_t optlen)
1206{
1207 int ret = 0;
1208
1209 switch (optname) {
1210 case RDMA_OPTION_ID_TOS:
1211 if (optlen != sizeof(u8)) {
1212 ret = -EINVAL;
1213 break;
1214 }
1215 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1216 break;
1217 case RDMA_OPTION_ID_REUSEADDR:
1218 if (optlen != sizeof(int)) {
1219 ret = -EINVAL;
1220 break;
1221 }
1222 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1223 break;
1224 case RDMA_OPTION_ID_AFONLY:
1225 if (optlen != sizeof(int)) {
1226 ret = -EINVAL;
1227 break;
1228 }
1229 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1230 break;
1231 case RDMA_OPTION_ID_ACK_TIMEOUT:
1232 if (optlen != sizeof(u8)) {
1233 ret = -EINVAL;
1234 break;
1235 }
1236 ret = rdma_set_ack_timeout(ctx->cm_id, *((u8 *)optval));
1237 break;
1238 default:
1239 ret = -ENOSYS;
1240 }
1241
1242 return ret;
1243}
1244
1245static int ucma_set_ib_path(struct ucma_context *ctx,
1246 struct ib_path_rec_data *path_data, size_t optlen)
1247{
1248 struct sa_path_rec sa_path;
1249 struct rdma_cm_event event;
1250 int ret;
1251
1252 if (optlen % sizeof(*path_data))
1253 return -EINVAL;
1254
1255 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1256 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1257 IB_PATH_BIDIRECTIONAL))
1258 break;
1259 }
1260
1261 if (!optlen)
1262 return -EINVAL;
1263
1264 if (!ctx->cm_id->device)
1265 return -EINVAL;
1266
1267 memset(&sa_path, 0, sizeof(sa_path));
1268
1269 sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1270 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1271
1272 if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1273 struct sa_path_rec opa;
1274
1275 sa_convert_path_ib_to_opa(&opa, &sa_path);
1276 ret = rdma_set_ib_path(ctx->cm_id, &opa);
1277 } else {
1278 ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
1279 }
1280 if (ret)
1281 return ret;
1282
1283 memset(&event, 0, sizeof event);
1284 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1285 return ucma_event_handler(ctx->cm_id, &event);
1286}
1287
1288static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1289 void *optval, size_t optlen)
1290{
1291 int ret;
1292
1293 switch (optname) {
1294 case RDMA_OPTION_IB_PATH:
1295 ret = ucma_set_ib_path(ctx, optval, optlen);
1296 break;
1297 default:
1298 ret = -ENOSYS;
1299 }
1300
1301 return ret;
1302}
1303
1304static int ucma_set_option_level(struct ucma_context *ctx, int level,
1305 int optname, void *optval, size_t optlen)
1306{
1307 int ret;
1308
1309 switch (level) {
1310 case RDMA_OPTION_ID:
1311 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1312 break;
1313 case RDMA_OPTION_IB:
1314 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1315 break;
1316 default:
1317 ret = -ENOSYS;
1318 }
1319
1320 return ret;
1321}
1322
1323static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1324 int in_len, int out_len)
1325{
1326 struct rdma_ucm_set_option cmd;
1327 struct ucma_context *ctx;
1328 void *optval;
1329 int ret;
1330
1331 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1332 return -EFAULT;
1333
1334 if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1335 return -EINVAL;
1336
1337 ctx = ucma_get_ctx(file, cmd.id);
1338 if (IS_ERR(ctx))
1339 return PTR_ERR(ctx);
1340
1341 optval = memdup_user(u64_to_user_ptr(cmd.optval),
1342 cmd.optlen);
1343 if (IS_ERR(optval)) {
1344 ret = PTR_ERR(optval);
1345 goto out;
1346 }
1347
1348 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1349 cmd.optlen);
1350 kfree(optval);
1351
1352out:
1353 ucma_put_ctx(ctx);
1354 return ret;
1355}
1356
1357static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1358 int in_len, int out_len)
1359{
1360 struct rdma_ucm_notify cmd;
1361 struct ucma_context *ctx;
1362 int ret = -EINVAL;
1363
1364 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1365 return -EFAULT;
1366
1367 ctx = ucma_get_ctx(file, cmd.id);
1368 if (IS_ERR(ctx))
1369 return PTR_ERR(ctx);
1370
1371 if (ctx->cm_id->device)
1372 ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
1373
1374 ucma_put_ctx(ctx);
1375 return ret;
1376}
1377
1378static ssize_t ucma_process_join(struct ucma_file *file,
1379 struct rdma_ucm_join_mcast *cmd, int out_len)
1380{
1381 struct rdma_ucm_create_id_resp resp;
1382 struct ucma_context *ctx;
1383 struct ucma_multicast *mc;
1384 struct sockaddr *addr;
1385 int ret;
1386 u8 join_state;
1387
1388 if (out_len < sizeof(resp))
1389 return -ENOSPC;
1390
1391 addr = (struct sockaddr *) &cmd->addr;
1392 if (cmd->addr_size != rdma_addr_size(addr))
1393 return -EINVAL;
1394
1395 if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1396 join_state = BIT(FULLMEMBER_JOIN);
1397 else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1398 join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1399 else
1400 return -EINVAL;
1401
1402 ctx = ucma_get_ctx_dev(file, cmd->id);
1403 if (IS_ERR(ctx))
1404 return PTR_ERR(ctx);
1405
1406 mutex_lock(&file->mut);
1407 mc = ucma_alloc_multicast(ctx);
1408 if (!mc) {
1409 ret = -ENOMEM;
1410 goto err1;
1411 }
1412 mc->join_state = join_state;
1413 mc->uid = cmd->uid;
1414 memcpy(&mc->addr, addr, cmd->addr_size);
1415 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1416 join_state, mc);
1417 if (ret)
1418 goto err2;
1419
1420 resp.id = mc->id;
1421 if (copy_to_user(u64_to_user_ptr(cmd->response),
1422 &resp, sizeof(resp))) {
1423 ret = -EFAULT;
1424 goto err3;
1425 }
1426
1427 xa_store(&multicast_table, mc->id, mc, 0);
1428
1429 mutex_unlock(&file->mut);
1430 ucma_put_ctx(ctx);
1431 return 0;
1432
1433err3:
1434 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1435 ucma_cleanup_mc_events(mc);
1436err2:
1437 xa_erase(&multicast_table, mc->id);
1438 list_del(&mc->list);
1439 kfree(mc);
1440err1:
1441 mutex_unlock(&file->mut);
1442 ucma_put_ctx(ctx);
1443 return ret;
1444}
1445
1446static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1447 const char __user *inbuf,
1448 int in_len, int out_len)
1449{
1450 struct rdma_ucm_join_ip_mcast cmd;
1451 struct rdma_ucm_join_mcast join_cmd;
1452
1453 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1454 return -EFAULT;
1455
1456 join_cmd.response = cmd.response;
1457 join_cmd.uid = cmd.uid;
1458 join_cmd.id = cmd.id;
1459 join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1460 if (!join_cmd.addr_size)
1461 return -EINVAL;
1462
1463 join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1464 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1465
1466 return ucma_process_join(file, &join_cmd, out_len);
1467}
1468
1469static ssize_t ucma_join_multicast(struct ucma_file *file,
1470 const char __user *inbuf,
1471 int in_len, int out_len)
1472{
1473 struct rdma_ucm_join_mcast cmd;
1474
1475 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1476 return -EFAULT;
1477
1478 if (!rdma_addr_size_kss(&cmd.addr))
1479 return -EINVAL;
1480
1481 return ucma_process_join(file, &cmd, out_len);
1482}
1483
1484static ssize_t ucma_leave_multicast(struct ucma_file *file,
1485 const char __user *inbuf,
1486 int in_len, int out_len)
1487{
1488 struct rdma_ucm_destroy_id cmd;
1489 struct rdma_ucm_destroy_id_resp resp;
1490 struct ucma_multicast *mc;
1491 int ret = 0;
1492
1493 if (out_len < sizeof(resp))
1494 return -ENOSPC;
1495
1496 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1497 return -EFAULT;
1498
1499 xa_lock(&multicast_table);
1500 mc = xa_load(&multicast_table, cmd.id);
1501 if (!mc)
1502 mc = ERR_PTR(-ENOENT);
1503 else if (mc->ctx->file != file)
1504 mc = ERR_PTR(-EINVAL);
1505 else if (!atomic_inc_not_zero(&mc->ctx->ref))
1506 mc = ERR_PTR(-ENXIO);
1507 else
1508 __xa_erase(&multicast_table, mc->id);
1509 xa_unlock(&multicast_table);
1510
1511 if (IS_ERR(mc)) {
1512 ret = PTR_ERR(mc);
1513 goto out;
1514 }
1515
1516 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1517 mutex_lock(&mc->ctx->file->mut);
1518 ucma_cleanup_mc_events(mc);
1519 list_del(&mc->list);
1520 mutex_unlock(&mc->ctx->file->mut);
1521
1522 ucma_put_ctx(mc->ctx);
1523 resp.events_reported = mc->events_reported;
1524 kfree(mc);
1525
1526 if (copy_to_user(u64_to_user_ptr(cmd.response),
1527 &resp, sizeof(resp)))
1528 ret = -EFAULT;
1529out:
1530 return ret;
1531}
1532
1533static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1534{
1535 /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1536 if (file1 < file2) {
1537 mutex_lock(&file1->mut);
1538 mutex_lock_nested(&file2->mut, SINGLE_DEPTH_NESTING);
1539 } else {
1540 mutex_lock(&file2->mut);
1541 mutex_lock_nested(&file1->mut, SINGLE_DEPTH_NESTING);
1542 }
1543}
1544
1545static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1546{
1547 if (file1 < file2) {
1548 mutex_unlock(&file2->mut);
1549 mutex_unlock(&file1->mut);
1550 } else {
1551 mutex_unlock(&file1->mut);
1552 mutex_unlock(&file2->mut);
1553 }
1554}
1555
1556static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1557{
1558 struct ucma_event *uevent, *tmp;
1559
1560 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1561 if (uevent->ctx == ctx)
1562 list_move_tail(&uevent->list, &file->event_list);
1563}
1564
1565static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1566 const char __user *inbuf,
1567 int in_len, int out_len)
1568{
1569 struct rdma_ucm_migrate_id cmd;
1570 struct rdma_ucm_migrate_resp resp;
1571 struct ucma_context *ctx;
1572 struct fd f;
1573 struct ucma_file *cur_file;
1574 int ret = 0;
1575
1576 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1577 return -EFAULT;
1578
1579 /* Get current fd to protect against it being closed */
1580 f = fdget(cmd.fd);
1581 if (!f.file)
1582 return -ENOENT;
1583 if (f.file->f_op != &ucma_fops) {
1584 ret = -EINVAL;
1585 goto file_put;
1586 }
1587
1588 /* Validate current fd and prevent destruction of id. */
1589 ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1590 if (IS_ERR(ctx)) {
1591 ret = PTR_ERR(ctx);
1592 goto file_put;
1593 }
1594
1595 cur_file = ctx->file;
1596 if (cur_file == new_file) {
1597 resp.events_reported = ctx->events_reported;
1598 goto response;
1599 }
1600
1601 /*
1602 * Migrate events between fd's, maintaining order, and avoiding new
1603 * events being added before existing events.
1604 */
1605 ucma_lock_files(cur_file, new_file);
1606 xa_lock(&ctx_table);
1607
1608 list_move_tail(&ctx->list, &new_file->ctx_list);
1609 ucma_move_events(ctx, new_file);
1610 ctx->file = new_file;
1611 resp.events_reported = ctx->events_reported;
1612
1613 xa_unlock(&ctx_table);
1614 ucma_unlock_files(cur_file, new_file);
1615
1616response:
1617 if (copy_to_user(u64_to_user_ptr(cmd.response),
1618 &resp, sizeof(resp)))
1619 ret = -EFAULT;
1620
1621 ucma_put_ctx(ctx);
1622file_put:
1623 fdput(f);
1624 return ret;
1625}
1626
1627static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1628 const char __user *inbuf,
1629 int in_len, int out_len) = {
1630 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1631 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1632 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1633 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1634 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1635 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1636 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1637 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1638 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1639 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1640 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1641 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1642 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1643 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1644 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1645 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1646 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1647 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1648 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1649 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1650 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1651 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1652 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1653};
1654
1655static ssize_t ucma_write(struct file *filp, const char __user *buf,
1656 size_t len, loff_t *pos)
1657{
1658 struct ucma_file *file = filp->private_data;
1659 struct rdma_ucm_cmd_hdr hdr;
1660 ssize_t ret;
1661
1662 if (!ib_safe_file_access(filp)) {
1663 pr_err_once("ucma_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1664 task_tgid_vnr(current), current->comm);
1665 return -EACCES;
1666 }
1667
1668 if (len < sizeof(hdr))
1669 return -EINVAL;
1670
1671 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1672 return -EFAULT;
1673
1674 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1675 return -EINVAL;
1676 hdr.cmd = array_index_nospec(hdr.cmd, ARRAY_SIZE(ucma_cmd_table));
1677
1678 if (hdr.in + sizeof(hdr) > len)
1679 return -EINVAL;
1680
1681 if (!ucma_cmd_table[hdr.cmd])
1682 return -ENOSYS;
1683
1684 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1685 if (!ret)
1686 ret = len;
1687
1688 return ret;
1689}
1690
1691static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1692{
1693 struct ucma_file *file = filp->private_data;
1694 __poll_t mask = 0;
1695
1696 poll_wait(filp, &file->poll_wait, wait);
1697
1698 if (!list_empty(&file->event_list))
1699 mask = EPOLLIN | EPOLLRDNORM;
1700
1701 return mask;
1702}
1703
1704/*
1705 * ucma_open() does not need the BKL:
1706 *
1707 * - no global state is referred to;
1708 * - there is no ioctl method to race against;
1709 * - no further module initialization is required for open to work
1710 * after the device is registered.
1711 */
1712static int ucma_open(struct inode *inode, struct file *filp)
1713{
1714 struct ucma_file *file;
1715
1716 file = kmalloc(sizeof *file, GFP_KERNEL);
1717 if (!file)
1718 return -ENOMEM;
1719
1720 file->close_wq = alloc_ordered_workqueue("ucma_close_id",
1721 WQ_MEM_RECLAIM);
1722 if (!file->close_wq) {
1723 kfree(file);
1724 return -ENOMEM;
1725 }
1726
1727 INIT_LIST_HEAD(&file->event_list);
1728 INIT_LIST_HEAD(&file->ctx_list);
1729 init_waitqueue_head(&file->poll_wait);
1730 mutex_init(&file->mut);
1731
1732 filp->private_data = file;
1733 file->filp = filp;
1734
1735 return stream_open(inode, filp);
1736}
1737
1738static int ucma_close(struct inode *inode, struct file *filp)
1739{
1740 struct ucma_file *file = filp->private_data;
1741 struct ucma_context *ctx, *tmp;
1742
1743 mutex_lock(&file->mut);
1744 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1745 ctx->destroying = 1;
1746 mutex_unlock(&file->mut);
1747
1748 xa_erase(&ctx_table, ctx->id);
1749 flush_workqueue(file->close_wq);
1750 /* At that step once ctx was marked as destroying and workqueue
1751 * was flushed we are safe from any inflights handlers that
1752 * might put other closing task.
1753 */
1754 xa_lock(&ctx_table);
1755 if (!ctx->closing) {
1756 xa_unlock(&ctx_table);
1757 ucma_put_ctx(ctx);
1758 wait_for_completion(&ctx->comp);
1759 /* rdma_destroy_id ensures that no event handlers are
1760 * inflight for that id before releasing it.
1761 */
1762 rdma_destroy_id(ctx->cm_id);
1763 } else {
1764 xa_unlock(&ctx_table);
1765 }
1766
1767 ucma_free_ctx(ctx);
1768 mutex_lock(&file->mut);
1769 }
1770 mutex_unlock(&file->mut);
1771 destroy_workqueue(file->close_wq);
1772 kfree(file);
1773 return 0;
1774}
1775
1776static const struct file_operations ucma_fops = {
1777 .owner = THIS_MODULE,
1778 .open = ucma_open,
1779 .release = ucma_close,
1780 .write = ucma_write,
1781 .poll = ucma_poll,
1782 .llseek = no_llseek,
1783};
1784
1785static struct miscdevice ucma_misc = {
1786 .minor = MISC_DYNAMIC_MINOR,
1787 .name = "rdma_cm",
1788 .nodename = "infiniband/rdma_cm",
1789 .mode = 0666,
1790 .fops = &ucma_fops,
1791};
1792
1793static int ucma_get_global_nl_info(struct ib_client_nl_info *res)
1794{
1795 res->abi = RDMA_USER_CM_ABI_VERSION;
1796 res->cdev = ucma_misc.this_device;
1797 return 0;
1798}
1799
1800static struct ib_client rdma_cma_client = {
1801 .name = "rdma_cm",
1802 .get_global_nl_info = ucma_get_global_nl_info,
1803};
1804MODULE_ALIAS_RDMA_CLIENT("rdma_cm");
1805
1806static ssize_t show_abi_version(struct device *dev,
1807 struct device_attribute *attr,
1808 char *buf)
1809{
1810 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1811}
1812static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1813
1814static int __init ucma_init(void)
1815{
1816 int ret;
1817
1818 ret = misc_register(&ucma_misc);
1819 if (ret)
1820 return ret;
1821
1822 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1823 if (ret) {
1824 pr_err("rdma_ucm: couldn't create abi_version attr\n");
1825 goto err1;
1826 }
1827
1828 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1829 if (!ucma_ctl_table_hdr) {
1830 pr_err("rdma_ucm: couldn't register sysctl paths\n");
1831 ret = -ENOMEM;
1832 goto err2;
1833 }
1834
1835 ret = ib_register_client(&rdma_cma_client);
1836 if (ret)
1837 goto err3;
1838
1839 return 0;
1840err3:
1841 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1842err2:
1843 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1844err1:
1845 misc_deregister(&ucma_misc);
1846 return ret;
1847}
1848
1849static void __exit ucma_cleanup(void)
1850{
1851 ib_unregister_client(&rdma_cma_client);
1852 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1853 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1854 misc_deregister(&ucma_misc);
1855}
1856
1857module_init(ucma_init);
1858module_exit(ucma_cleanup);