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