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