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#include <linux/module.h>
45#include <linux/nsproxy.h>
46
47#include <linux/nospec.h>
48
49#include <rdma/rdma_user_cm.h>
50#include <rdma/ib_marshall.h>
51#include <rdma/rdma_cm.h>
52#include <rdma/rdma_cm_ib.h>
53#include <rdma/ib_addr.h>
54#include <rdma/ib.h>
55#include <rdma/ib_cm.h>
56#include <rdma/rdma_netlink.h>
57#include "core_priv.h"
58
59MODULE_AUTHOR("Sean Hefty");
60MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
61MODULE_LICENSE("Dual BSD/GPL");
62
63static unsigned int max_backlog = 1024;
64
65static struct ctl_table_header *ucma_ctl_table_hdr;
66static struct ctl_table ucma_ctl_table[] = {
67 {
68 .procname = "max_backlog",
69 .data = &max_backlog,
70 .maxlen = sizeof max_backlog,
71 .mode = 0644,
72 .proc_handler = proc_dointvec,
73 },
74};
75
76struct ucma_file {
77 struct mutex mut;
78 struct file *filp;
79 struct list_head ctx_list;
80 struct list_head event_list;
81 wait_queue_head_t poll_wait;
82};
83
84struct ucma_context {
85 u32 id;
86 struct completion comp;
87 refcount_t ref;
88 int events_reported;
89 atomic_t backlog;
90
91 struct ucma_file *file;
92 struct rdma_cm_id *cm_id;
93 struct mutex mutex;
94 u64 uid;
95
96 struct list_head list;
97 struct list_head mc_list;
98 struct work_struct close_work;
99};
100
101struct ucma_multicast {
102 struct ucma_context *ctx;
103 u32 id;
104 int events_reported;
105
106 u64 uid;
107 u8 join_state;
108 struct list_head list;
109 struct sockaddr_storage addr;
110};
111
112struct ucma_event {
113 struct ucma_context *ctx;
114 struct ucma_context *conn_req_ctx;
115 struct ucma_multicast *mc;
116 struct list_head list;
117 struct rdma_ucm_event_resp resp;
118};
119
120static DEFINE_XARRAY_ALLOC(ctx_table);
121static DEFINE_XARRAY_ALLOC(multicast_table);
122
123static const struct file_operations ucma_fops;
124static int ucma_destroy_private_ctx(struct ucma_context *ctx);
125
126static inline struct ucma_context *_ucma_find_context(int id,
127 struct ucma_file *file)
128{
129 struct ucma_context *ctx;
130
131 ctx = xa_load(&ctx_table, id);
132 if (!ctx)
133 ctx = ERR_PTR(-ENOENT);
134 else if (ctx->file != file)
135 ctx = ERR_PTR(-EINVAL);
136 return ctx;
137}
138
139static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
140{
141 struct ucma_context *ctx;
142
143 xa_lock(&ctx_table);
144 ctx = _ucma_find_context(id, file);
145 if (!IS_ERR(ctx))
146 if (!refcount_inc_not_zero(&ctx->ref))
147 ctx = ERR_PTR(-ENXIO);
148 xa_unlock(&ctx_table);
149 return ctx;
150}
151
152static void ucma_put_ctx(struct ucma_context *ctx)
153{
154 if (refcount_dec_and_test(&ctx->ref))
155 complete(&ctx->comp);
156}
157
158/*
159 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
160 * CM_ID is bound.
161 */
162static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
163{
164 struct ucma_context *ctx = ucma_get_ctx(file, id);
165
166 if (IS_ERR(ctx))
167 return ctx;
168 if (!ctx->cm_id->device) {
169 ucma_put_ctx(ctx);
170 return ERR_PTR(-EINVAL);
171 }
172 return ctx;
173}
174
175static void ucma_close_id(struct work_struct *work)
176{
177 struct ucma_context *ctx = container_of(work, struct ucma_context, close_work);
178
179 /* once all inflight tasks are finished, we close all underlying
180 * resources. The context is still alive till its explicit destryoing
181 * by its creator. This puts back the xarray's reference.
182 */
183 ucma_put_ctx(ctx);
184 wait_for_completion(&ctx->comp);
185 /* No new events will be generated after destroying the id. */
186 rdma_destroy_id(ctx->cm_id);
187
188 /* Reading the cm_id without holding a positive ref is not allowed */
189 ctx->cm_id = NULL;
190}
191
192static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
193{
194 struct ucma_context *ctx;
195
196 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
197 if (!ctx)
198 return NULL;
199
200 INIT_WORK(&ctx->close_work, ucma_close_id);
201 init_completion(&ctx->comp);
202 INIT_LIST_HEAD(&ctx->mc_list);
203 /* So list_del() will work if we don't do ucma_finish_ctx() */
204 INIT_LIST_HEAD(&ctx->list);
205 ctx->file = file;
206 mutex_init(&ctx->mutex);
207
208 if (xa_alloc(&ctx_table, &ctx->id, NULL, xa_limit_32b, GFP_KERNEL)) {
209 kfree(ctx);
210 return NULL;
211 }
212 return ctx;
213}
214
215static void ucma_set_ctx_cm_id(struct ucma_context *ctx,
216 struct rdma_cm_id *cm_id)
217{
218 refcount_set(&ctx->ref, 1);
219 ctx->cm_id = cm_id;
220}
221
222static void ucma_finish_ctx(struct ucma_context *ctx)
223{
224 lockdep_assert_held(&ctx->file->mut);
225 list_add_tail(&ctx->list, &ctx->file->ctx_list);
226 xa_store(&ctx_table, ctx->id, ctx, GFP_KERNEL);
227}
228
229static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
230 struct rdma_conn_param *src)
231{
232 if (src->private_data_len)
233 memcpy(dst->private_data, src->private_data,
234 src->private_data_len);
235 dst->private_data_len = src->private_data_len;
236 dst->responder_resources = src->responder_resources;
237 dst->initiator_depth = src->initiator_depth;
238 dst->flow_control = src->flow_control;
239 dst->retry_count = src->retry_count;
240 dst->rnr_retry_count = src->rnr_retry_count;
241 dst->srq = src->srq;
242 dst->qp_num = src->qp_num;
243}
244
245static void ucma_copy_ud_event(struct ib_device *device,
246 struct rdma_ucm_ud_param *dst,
247 struct rdma_ud_param *src)
248{
249 if (src->private_data_len)
250 memcpy(dst->private_data, src->private_data,
251 src->private_data_len);
252 dst->private_data_len = src->private_data_len;
253 ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
254 dst->qp_num = src->qp_num;
255 dst->qkey = src->qkey;
256}
257
258static struct ucma_event *ucma_create_uevent(struct ucma_context *ctx,
259 struct rdma_cm_event *event)
260{
261 struct ucma_event *uevent;
262
263 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
264 if (!uevent)
265 return NULL;
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 uevent->resp.event = event->event;
282 uevent->resp.status = event->status;
283 if (ctx->cm_id->qp_type == IB_QPT_UD)
284 ucma_copy_ud_event(ctx->cm_id->device, &uevent->resp.param.ud,
285 &event->param.ud);
286 else
287 ucma_copy_conn_event(&uevent->resp.param.conn,
288 &event->param.conn);
289
290 uevent->resp.ece.vendor_id = event->ece.vendor_id;
291 uevent->resp.ece.attr_mod = event->ece.attr_mod;
292 return uevent;
293}
294
295static int ucma_connect_event_handler(struct rdma_cm_id *cm_id,
296 struct rdma_cm_event *event)
297{
298 struct ucma_context *listen_ctx = cm_id->context;
299 struct ucma_context *ctx;
300 struct ucma_event *uevent;
301
302 if (!atomic_add_unless(&listen_ctx->backlog, -1, 0))
303 return -ENOMEM;
304 ctx = ucma_alloc_ctx(listen_ctx->file);
305 if (!ctx)
306 goto err_backlog;
307 ucma_set_ctx_cm_id(ctx, cm_id);
308
309 uevent = ucma_create_uevent(listen_ctx, event);
310 if (!uevent)
311 goto err_alloc;
312 uevent->conn_req_ctx = ctx;
313 uevent->resp.id = ctx->id;
314
315 ctx->cm_id->context = ctx;
316
317 mutex_lock(&ctx->file->mut);
318 ucma_finish_ctx(ctx);
319 list_add_tail(&uevent->list, &ctx->file->event_list);
320 mutex_unlock(&ctx->file->mut);
321 wake_up_interruptible(&ctx->file->poll_wait);
322 return 0;
323
324err_alloc:
325 ucma_destroy_private_ctx(ctx);
326err_backlog:
327 atomic_inc(&listen_ctx->backlog);
328 /* Returning error causes the new ID to be destroyed */
329 return -ENOMEM;
330}
331
332static int ucma_event_handler(struct rdma_cm_id *cm_id,
333 struct rdma_cm_event *event)
334{
335 struct ucma_event *uevent;
336 struct ucma_context *ctx = cm_id->context;
337
338 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST)
339 return ucma_connect_event_handler(cm_id, event);
340
341 /*
342 * We ignore events for new connections until userspace has set their
343 * context. This can only happen if an error occurs on a new connection
344 * before the user accepts it. This is okay, since the accept will just
345 * fail later. However, we do need to release the underlying HW
346 * resources in case of a device removal event.
347 */
348 if (ctx->uid) {
349 uevent = ucma_create_uevent(ctx, event);
350 if (!uevent)
351 return 0;
352
353 mutex_lock(&ctx->file->mut);
354 list_add_tail(&uevent->list, &ctx->file->event_list);
355 mutex_unlock(&ctx->file->mut);
356 wake_up_interruptible(&ctx->file->poll_wait);
357 }
358
359 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL) {
360 xa_lock(&ctx_table);
361 if (xa_load(&ctx_table, ctx->id) == ctx)
362 queue_work(system_unbound_wq, &ctx->close_work);
363 xa_unlock(&ctx_table);
364 }
365 return 0;
366}
367
368static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
369 int in_len, int out_len)
370{
371 struct rdma_ucm_get_event cmd;
372 struct ucma_event *uevent;
373
374 /*
375 * Old 32 bit user space does not send the 4 byte padding in the
376 * reserved field. We don't care, allow it to keep working.
377 */
378 if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved) -
379 sizeof(uevent->resp.ece))
380 return -ENOSPC;
381
382 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
383 return -EFAULT;
384
385 mutex_lock(&file->mut);
386 while (list_empty(&file->event_list)) {
387 mutex_unlock(&file->mut);
388
389 if (file->filp->f_flags & O_NONBLOCK)
390 return -EAGAIN;
391
392 if (wait_event_interruptible(file->poll_wait,
393 !list_empty(&file->event_list)))
394 return -ERESTARTSYS;
395
396 mutex_lock(&file->mut);
397 }
398
399 uevent = list_first_entry(&file->event_list, struct ucma_event, list);
400
401 if (copy_to_user(u64_to_user_ptr(cmd.response),
402 &uevent->resp,
403 min_t(size_t, out_len, sizeof(uevent->resp)))) {
404 mutex_unlock(&file->mut);
405 return -EFAULT;
406 }
407
408 list_del(&uevent->list);
409 uevent->ctx->events_reported++;
410 if (uevent->mc)
411 uevent->mc->events_reported++;
412 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
413 atomic_inc(&uevent->ctx->backlog);
414 mutex_unlock(&file->mut);
415
416 kfree(uevent);
417 return 0;
418}
419
420static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
421{
422 switch (cmd->ps) {
423 case RDMA_PS_TCP:
424 *qp_type = IB_QPT_RC;
425 return 0;
426 case RDMA_PS_UDP:
427 case RDMA_PS_IPOIB:
428 *qp_type = IB_QPT_UD;
429 return 0;
430 case RDMA_PS_IB:
431 *qp_type = cmd->qp_type;
432 return 0;
433 default:
434 return -EINVAL;
435 }
436}
437
438static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
439 int in_len, int out_len)
440{
441 struct rdma_ucm_create_id cmd;
442 struct rdma_ucm_create_id_resp resp;
443 struct ucma_context *ctx;
444 struct rdma_cm_id *cm_id;
445 enum ib_qp_type qp_type;
446 int ret;
447
448 if (out_len < sizeof(resp))
449 return -ENOSPC;
450
451 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
452 return -EFAULT;
453
454 ret = ucma_get_qp_type(&cmd, &qp_type);
455 if (ret)
456 return ret;
457
458 ctx = ucma_alloc_ctx(file);
459 if (!ctx)
460 return -ENOMEM;
461
462 ctx->uid = cmd.uid;
463 cm_id = rdma_create_user_id(ucma_event_handler, ctx, cmd.ps, qp_type);
464 if (IS_ERR(cm_id)) {
465 ret = PTR_ERR(cm_id);
466 goto err1;
467 }
468 ucma_set_ctx_cm_id(ctx, cm_id);
469
470 resp.id = ctx->id;
471 if (copy_to_user(u64_to_user_ptr(cmd.response),
472 &resp, sizeof(resp))) {
473 ret = -EFAULT;
474 goto err1;
475 }
476
477 mutex_lock(&file->mut);
478 ucma_finish_ctx(ctx);
479 mutex_unlock(&file->mut);
480 return 0;
481
482err1:
483 ucma_destroy_private_ctx(ctx);
484 return ret;
485}
486
487static void ucma_cleanup_multicast(struct ucma_context *ctx)
488{
489 struct ucma_multicast *mc, *tmp;
490
491 xa_lock(&multicast_table);
492 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
493 list_del(&mc->list);
494 /*
495 * At this point mc->ctx->ref is 0 so the mc cannot leave the
496 * lock on the reader and this is enough serialization
497 */
498 __xa_erase(&multicast_table, mc->id);
499 kfree(mc);
500 }
501 xa_unlock(&multicast_table);
502}
503
504static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
505{
506 struct ucma_event *uevent, *tmp;
507
508 rdma_lock_handler(mc->ctx->cm_id);
509 mutex_lock(&mc->ctx->file->mut);
510 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
511 if (uevent->mc != mc)
512 continue;
513
514 list_del(&uevent->list);
515 kfree(uevent);
516 }
517 mutex_unlock(&mc->ctx->file->mut);
518 rdma_unlock_handler(mc->ctx->cm_id);
519}
520
521static int ucma_cleanup_ctx_events(struct ucma_context *ctx)
522{
523 int events_reported;
524 struct ucma_event *uevent, *tmp;
525 LIST_HEAD(list);
526
527 /* Cleanup events not yet reported to the user.*/
528 mutex_lock(&ctx->file->mut);
529 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
530 if (uevent->ctx != ctx)
531 continue;
532
533 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST &&
534 xa_cmpxchg(&ctx_table, uevent->conn_req_ctx->id,
535 uevent->conn_req_ctx, XA_ZERO_ENTRY,
536 GFP_KERNEL) == uevent->conn_req_ctx) {
537 list_move_tail(&uevent->list, &list);
538 continue;
539 }
540 list_del(&uevent->list);
541 kfree(uevent);
542 }
543 list_del(&ctx->list);
544 events_reported = ctx->events_reported;
545 mutex_unlock(&ctx->file->mut);
546
547 /*
548 * If this was a listening ID then any connections spawned from it that
549 * have not been delivered to userspace are cleaned up too. Must be done
550 * outside any locks.
551 */
552 list_for_each_entry_safe(uevent, tmp, &list, list) {
553 ucma_destroy_private_ctx(uevent->conn_req_ctx);
554 kfree(uevent);
555 }
556 return events_reported;
557}
558
559/*
560 * When this is called the xarray must have a XA_ZERO_ENTRY in the ctx->id (ie
561 * the ctx is not public to the user). This either because:
562 * - ucma_finish_ctx() hasn't been called
563 * - xa_cmpxchg() succeed to remove the entry (only one thread can succeed)
564 */
565static int ucma_destroy_private_ctx(struct ucma_context *ctx)
566{
567 int events_reported;
568
569 /*
570 * Destroy the underlying cm_id. New work queuing is prevented now by
571 * the removal from the xarray. Once the work is cancled ref will either
572 * be 0 because the work ran to completion and consumed the ref from the
573 * xarray, or it will be positive because we still have the ref from the
574 * xarray. This can also be 0 in cases where cm_id was never set
575 */
576 cancel_work_sync(&ctx->close_work);
577 if (refcount_read(&ctx->ref))
578 ucma_close_id(&ctx->close_work);
579
580 events_reported = ucma_cleanup_ctx_events(ctx);
581 ucma_cleanup_multicast(ctx);
582
583 WARN_ON(xa_cmpxchg(&ctx_table, ctx->id, XA_ZERO_ENTRY, NULL,
584 GFP_KERNEL) != NULL);
585 mutex_destroy(&ctx->mutex);
586 kfree(ctx);
587 return events_reported;
588}
589
590static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
591 int in_len, int out_len)
592{
593 struct rdma_ucm_destroy_id cmd;
594 struct rdma_ucm_destroy_id_resp resp;
595 struct ucma_context *ctx;
596 int ret = 0;
597
598 if (out_len < sizeof(resp))
599 return -ENOSPC;
600
601 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
602 return -EFAULT;
603
604 xa_lock(&ctx_table);
605 ctx = _ucma_find_context(cmd.id, file);
606 if (!IS_ERR(ctx)) {
607 if (__xa_cmpxchg(&ctx_table, ctx->id, ctx, XA_ZERO_ENTRY,
608 GFP_KERNEL) != ctx)
609 ctx = ERR_PTR(-ENOENT);
610 }
611 xa_unlock(&ctx_table);
612
613 if (IS_ERR(ctx))
614 return PTR_ERR(ctx);
615
616 resp.events_reported = ucma_destroy_private_ctx(ctx);
617 if (copy_to_user(u64_to_user_ptr(cmd.response),
618 &resp, sizeof(resp)))
619 ret = -EFAULT;
620
621 return ret;
622}
623
624static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
625 int in_len, int out_len)
626{
627 struct rdma_ucm_bind_ip cmd;
628 struct ucma_context *ctx;
629 int ret;
630
631 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
632 return -EFAULT;
633
634 if (!rdma_addr_size_in6(&cmd.addr))
635 return -EINVAL;
636
637 ctx = ucma_get_ctx(file, cmd.id);
638 if (IS_ERR(ctx))
639 return PTR_ERR(ctx);
640
641 mutex_lock(&ctx->mutex);
642 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
643 mutex_unlock(&ctx->mutex);
644
645 ucma_put_ctx(ctx);
646 return ret;
647}
648
649static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
650 int in_len, int out_len)
651{
652 struct rdma_ucm_bind cmd;
653 struct ucma_context *ctx;
654 int ret;
655
656 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
657 return -EFAULT;
658
659 if (cmd.reserved || !cmd.addr_size ||
660 cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
661 return -EINVAL;
662
663 ctx = ucma_get_ctx(file, cmd.id);
664 if (IS_ERR(ctx))
665 return PTR_ERR(ctx);
666
667 mutex_lock(&ctx->mutex);
668 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
669 mutex_unlock(&ctx->mutex);
670 ucma_put_ctx(ctx);
671 return ret;
672}
673
674static ssize_t ucma_resolve_ip(struct ucma_file *file,
675 const char __user *inbuf,
676 int in_len, int out_len)
677{
678 struct rdma_ucm_resolve_ip cmd;
679 struct ucma_context *ctx;
680 int ret;
681
682 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
683 return -EFAULT;
684
685 if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
686 !rdma_addr_size_in6(&cmd.dst_addr))
687 return -EINVAL;
688
689 ctx = ucma_get_ctx(file, cmd.id);
690 if (IS_ERR(ctx))
691 return PTR_ERR(ctx);
692
693 mutex_lock(&ctx->mutex);
694 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
695 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
696 mutex_unlock(&ctx->mutex);
697 ucma_put_ctx(ctx);
698 return ret;
699}
700
701static ssize_t ucma_resolve_addr(struct ucma_file *file,
702 const char __user *inbuf,
703 int in_len, int out_len)
704{
705 struct rdma_ucm_resolve_addr cmd;
706 struct ucma_context *ctx;
707 int ret;
708
709 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
710 return -EFAULT;
711
712 if (cmd.reserved ||
713 (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
714 !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
715 return -EINVAL;
716
717 ctx = ucma_get_ctx(file, cmd.id);
718 if (IS_ERR(ctx))
719 return PTR_ERR(ctx);
720
721 mutex_lock(&ctx->mutex);
722 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
723 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
724 mutex_unlock(&ctx->mutex);
725 ucma_put_ctx(ctx);
726 return ret;
727}
728
729static ssize_t ucma_resolve_route(struct ucma_file *file,
730 const char __user *inbuf,
731 int in_len, int out_len)
732{
733 struct rdma_ucm_resolve_route cmd;
734 struct ucma_context *ctx;
735 int ret;
736
737 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
738 return -EFAULT;
739
740 ctx = ucma_get_ctx_dev(file, cmd.id);
741 if (IS_ERR(ctx))
742 return PTR_ERR(ctx);
743
744 mutex_lock(&ctx->mutex);
745 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
746 mutex_unlock(&ctx->mutex);
747 ucma_put_ctx(ctx);
748 return ret;
749}
750
751static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
752 struct rdma_route *route)
753{
754 struct rdma_dev_addr *dev_addr;
755
756 resp->num_paths = route->num_pri_alt_paths;
757 switch (route->num_pri_alt_paths) {
758 case 0:
759 dev_addr = &route->addr.dev_addr;
760 rdma_addr_get_dgid(dev_addr,
761 (union ib_gid *) &resp->ib_route[0].dgid);
762 rdma_addr_get_sgid(dev_addr,
763 (union ib_gid *) &resp->ib_route[0].sgid);
764 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
765 break;
766 case 2:
767 ib_copy_path_rec_to_user(&resp->ib_route[1],
768 &route->path_rec[1]);
769 fallthrough;
770 case 1:
771 ib_copy_path_rec_to_user(&resp->ib_route[0],
772 &route->path_rec[0]);
773 break;
774 default:
775 break;
776 }
777}
778
779static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
780 struct rdma_route *route)
781{
782
783 resp->num_paths = route->num_pri_alt_paths;
784 switch (route->num_pri_alt_paths) {
785 case 0:
786 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
787 (union ib_gid *)&resp->ib_route[0].dgid);
788 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
789 (union ib_gid *)&resp->ib_route[0].sgid);
790 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
791 break;
792 case 2:
793 ib_copy_path_rec_to_user(&resp->ib_route[1],
794 &route->path_rec[1]);
795 fallthrough;
796 case 1:
797 ib_copy_path_rec_to_user(&resp->ib_route[0],
798 &route->path_rec[0]);
799 break;
800 default:
801 break;
802 }
803}
804
805static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
806 struct rdma_route *route)
807{
808 struct rdma_dev_addr *dev_addr;
809
810 dev_addr = &route->addr.dev_addr;
811 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
812 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
813}
814
815static ssize_t ucma_query_route(struct ucma_file *file,
816 const char __user *inbuf,
817 int in_len, int out_len)
818{
819 struct rdma_ucm_query cmd;
820 struct rdma_ucm_query_route_resp resp;
821 struct ucma_context *ctx;
822 struct sockaddr *addr;
823 int ret = 0;
824
825 if (out_len < offsetof(struct rdma_ucm_query_route_resp, ibdev_index))
826 return -ENOSPC;
827
828 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
829 return -EFAULT;
830
831 ctx = ucma_get_ctx(file, cmd.id);
832 if (IS_ERR(ctx))
833 return PTR_ERR(ctx);
834
835 mutex_lock(&ctx->mutex);
836 memset(&resp, 0, sizeof resp);
837 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
838 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
839 sizeof(struct sockaddr_in) :
840 sizeof(struct sockaddr_in6));
841 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
842 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
843 sizeof(struct sockaddr_in) :
844 sizeof(struct sockaddr_in6));
845 if (!ctx->cm_id->device)
846 goto out;
847
848 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
849 resp.ibdev_index = ctx->cm_id->device->index;
850 resp.port_num = ctx->cm_id->port_num;
851
852 if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
853 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
854 else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
855 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
856 else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
857 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
858
859out:
860 mutex_unlock(&ctx->mutex);
861 if (copy_to_user(u64_to_user_ptr(cmd.response), &resp,
862 min_t(size_t, out_len, sizeof(resp))))
863 ret = -EFAULT;
864
865 ucma_put_ctx(ctx);
866 return ret;
867}
868
869static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
870 struct rdma_ucm_query_addr_resp *resp)
871{
872 if (!cm_id->device)
873 return;
874
875 resp->node_guid = (__force __u64) cm_id->device->node_guid;
876 resp->ibdev_index = cm_id->device->index;
877 resp->port_num = cm_id->port_num;
878 resp->pkey = (__force __u16) cpu_to_be16(
879 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
880}
881
882static ssize_t ucma_query_addr(struct ucma_context *ctx,
883 void __user *response, int out_len)
884{
885 struct rdma_ucm_query_addr_resp resp;
886 struct sockaddr *addr;
887 int ret = 0;
888
889 if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
890 return -ENOSPC;
891
892 memset(&resp, 0, sizeof resp);
893
894 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
895 resp.src_size = rdma_addr_size(addr);
896 memcpy(&resp.src_addr, addr, resp.src_size);
897
898 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
899 resp.dst_size = rdma_addr_size(addr);
900 memcpy(&resp.dst_addr, addr, resp.dst_size);
901
902 ucma_query_device_addr(ctx->cm_id, &resp);
903
904 if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
905 ret = -EFAULT;
906
907 return ret;
908}
909
910static ssize_t ucma_query_path(struct ucma_context *ctx,
911 void __user *response, int out_len)
912{
913 struct rdma_ucm_query_path_resp *resp;
914 int i, ret = 0;
915
916 if (out_len < sizeof(*resp))
917 return -ENOSPC;
918
919 resp = kzalloc(out_len, GFP_KERNEL);
920 if (!resp)
921 return -ENOMEM;
922
923 resp->num_paths = ctx->cm_id->route.num_pri_alt_paths;
924 for (i = 0, out_len -= sizeof(*resp);
925 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
926 i++, out_len -= sizeof(struct ib_path_rec_data)) {
927 struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
928
929 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
930 IB_PATH_BIDIRECTIONAL;
931 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
932 struct sa_path_rec ib;
933
934 sa_convert_path_opa_to_ib(&ib, rec);
935 ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
936
937 } else {
938 ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
939 }
940 }
941
942 if (copy_to_user(response, resp, struct_size(resp, path_data, i)))
943 ret = -EFAULT;
944
945 kfree(resp);
946 return ret;
947}
948
949static ssize_t ucma_query_gid(struct ucma_context *ctx,
950 void __user *response, int out_len)
951{
952 struct rdma_ucm_query_addr_resp resp;
953 struct sockaddr_ib *addr;
954 int ret = 0;
955
956 if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
957 return -ENOSPC;
958
959 memset(&resp, 0, sizeof resp);
960
961 ucma_query_device_addr(ctx->cm_id, &resp);
962
963 addr = (struct sockaddr_ib *) &resp.src_addr;
964 resp.src_size = sizeof(*addr);
965 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
966 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
967 } else {
968 addr->sib_family = AF_IB;
969 addr->sib_pkey = (__force __be16) resp.pkey;
970 rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
971 NULL);
972 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
973 &ctx->cm_id->route.addr.src_addr);
974 }
975
976 addr = (struct sockaddr_ib *) &resp.dst_addr;
977 resp.dst_size = sizeof(*addr);
978 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
979 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
980 } else {
981 addr->sib_family = AF_IB;
982 addr->sib_pkey = (__force __be16) resp.pkey;
983 rdma_read_gids(ctx->cm_id, NULL,
984 (union ib_gid *)&addr->sib_addr);
985 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
986 &ctx->cm_id->route.addr.dst_addr);
987 }
988
989 if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
990 ret = -EFAULT;
991
992 return ret;
993}
994
995static ssize_t ucma_query(struct ucma_file *file,
996 const char __user *inbuf,
997 int in_len, int out_len)
998{
999 struct rdma_ucm_query cmd;
1000 struct ucma_context *ctx;
1001 void __user *response;
1002 int ret;
1003
1004 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1005 return -EFAULT;
1006
1007 response = u64_to_user_ptr(cmd.response);
1008 ctx = ucma_get_ctx(file, cmd.id);
1009 if (IS_ERR(ctx))
1010 return PTR_ERR(ctx);
1011
1012 mutex_lock(&ctx->mutex);
1013 switch (cmd.option) {
1014 case RDMA_USER_CM_QUERY_ADDR:
1015 ret = ucma_query_addr(ctx, response, out_len);
1016 break;
1017 case RDMA_USER_CM_QUERY_PATH:
1018 ret = ucma_query_path(ctx, response, out_len);
1019 break;
1020 case RDMA_USER_CM_QUERY_GID:
1021 ret = ucma_query_gid(ctx, response, out_len);
1022 break;
1023 default:
1024 ret = -ENOSYS;
1025 break;
1026 }
1027 mutex_unlock(&ctx->mutex);
1028
1029 ucma_put_ctx(ctx);
1030 return ret;
1031}
1032
1033static void ucma_copy_conn_param(struct rdma_cm_id *id,
1034 struct rdma_conn_param *dst,
1035 struct rdma_ucm_conn_param *src)
1036{
1037 dst->private_data = src->private_data;
1038 dst->private_data_len = src->private_data_len;
1039 dst->responder_resources = src->responder_resources;
1040 dst->initiator_depth = src->initiator_depth;
1041 dst->flow_control = src->flow_control;
1042 dst->retry_count = src->retry_count;
1043 dst->rnr_retry_count = src->rnr_retry_count;
1044 dst->srq = src->srq;
1045 dst->qp_num = src->qp_num & 0xFFFFFF;
1046 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1047}
1048
1049static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1050 int in_len, int out_len)
1051{
1052 struct rdma_conn_param conn_param;
1053 struct rdma_ucm_ece ece = {};
1054 struct rdma_ucm_connect cmd;
1055 struct ucma_context *ctx;
1056 size_t in_size;
1057 int ret;
1058
1059 if (in_len < offsetofend(typeof(cmd), reserved))
1060 return -EINVAL;
1061 in_size = min_t(size_t, in_len, sizeof(cmd));
1062 if (copy_from_user(&cmd, inbuf, in_size))
1063 return -EFAULT;
1064
1065 if (!cmd.conn_param.valid)
1066 return -EINVAL;
1067
1068 ctx = ucma_get_ctx_dev(file, cmd.id);
1069 if (IS_ERR(ctx))
1070 return PTR_ERR(ctx);
1071
1072 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1073 if (offsetofend(typeof(cmd), ece) <= in_size) {
1074 ece.vendor_id = cmd.ece.vendor_id;
1075 ece.attr_mod = cmd.ece.attr_mod;
1076 }
1077
1078 mutex_lock(&ctx->mutex);
1079 ret = rdma_connect_ece(ctx->cm_id, &conn_param, &ece);
1080 mutex_unlock(&ctx->mutex);
1081 ucma_put_ctx(ctx);
1082 return ret;
1083}
1084
1085static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1086 int in_len, int out_len)
1087{
1088 struct rdma_ucm_listen cmd;
1089 struct ucma_context *ctx;
1090 int ret;
1091
1092 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1093 return -EFAULT;
1094
1095 ctx = ucma_get_ctx(file, cmd.id);
1096 if (IS_ERR(ctx))
1097 return PTR_ERR(ctx);
1098
1099 if (cmd.backlog <= 0 || cmd.backlog > max_backlog)
1100 cmd.backlog = max_backlog;
1101 atomic_set(&ctx->backlog, cmd.backlog);
1102
1103 mutex_lock(&ctx->mutex);
1104 ret = rdma_listen(ctx->cm_id, cmd.backlog);
1105 mutex_unlock(&ctx->mutex);
1106 ucma_put_ctx(ctx);
1107 return ret;
1108}
1109
1110static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1111 int in_len, int out_len)
1112{
1113 struct rdma_ucm_accept cmd;
1114 struct rdma_conn_param conn_param;
1115 struct rdma_ucm_ece ece = {};
1116 struct ucma_context *ctx;
1117 size_t in_size;
1118 int ret;
1119
1120 if (in_len < offsetofend(typeof(cmd), reserved))
1121 return -EINVAL;
1122 in_size = min_t(size_t, in_len, sizeof(cmd));
1123 if (copy_from_user(&cmd, inbuf, in_size))
1124 return -EFAULT;
1125
1126 ctx = ucma_get_ctx_dev(file, cmd.id);
1127 if (IS_ERR(ctx))
1128 return PTR_ERR(ctx);
1129
1130 if (offsetofend(typeof(cmd), ece) <= in_size) {
1131 ece.vendor_id = cmd.ece.vendor_id;
1132 ece.attr_mod = cmd.ece.attr_mod;
1133 }
1134
1135 if (cmd.conn_param.valid) {
1136 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1137 mutex_lock(&ctx->mutex);
1138 rdma_lock_handler(ctx->cm_id);
1139 ret = rdma_accept_ece(ctx->cm_id, &conn_param, &ece);
1140 if (!ret) {
1141 /* The uid must be set atomically with the handler */
1142 ctx->uid = cmd.uid;
1143 }
1144 rdma_unlock_handler(ctx->cm_id);
1145 mutex_unlock(&ctx->mutex);
1146 } else {
1147 mutex_lock(&ctx->mutex);
1148 rdma_lock_handler(ctx->cm_id);
1149 ret = rdma_accept_ece(ctx->cm_id, NULL, &ece);
1150 rdma_unlock_handler(ctx->cm_id);
1151 mutex_unlock(&ctx->mutex);
1152 }
1153 ucma_put_ctx(ctx);
1154 return ret;
1155}
1156
1157static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1158 int in_len, int out_len)
1159{
1160 struct rdma_ucm_reject cmd;
1161 struct ucma_context *ctx;
1162 int ret;
1163
1164 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1165 return -EFAULT;
1166
1167 if (!cmd.reason)
1168 cmd.reason = IB_CM_REJ_CONSUMER_DEFINED;
1169
1170 switch (cmd.reason) {
1171 case IB_CM_REJ_CONSUMER_DEFINED:
1172 case IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED:
1173 break;
1174 default:
1175 return -EINVAL;
1176 }
1177
1178 ctx = ucma_get_ctx_dev(file, cmd.id);
1179 if (IS_ERR(ctx))
1180 return PTR_ERR(ctx);
1181
1182 mutex_lock(&ctx->mutex);
1183 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len,
1184 cmd.reason);
1185 mutex_unlock(&ctx->mutex);
1186 ucma_put_ctx(ctx);
1187 return ret;
1188}
1189
1190static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1191 int in_len, int out_len)
1192{
1193 struct rdma_ucm_disconnect cmd;
1194 struct ucma_context *ctx;
1195 int ret;
1196
1197 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1198 return -EFAULT;
1199
1200 ctx = ucma_get_ctx_dev(file, cmd.id);
1201 if (IS_ERR(ctx))
1202 return PTR_ERR(ctx);
1203
1204 mutex_lock(&ctx->mutex);
1205 ret = rdma_disconnect(ctx->cm_id);
1206 mutex_unlock(&ctx->mutex);
1207 ucma_put_ctx(ctx);
1208 return ret;
1209}
1210
1211static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1212 const char __user *inbuf,
1213 int in_len, int out_len)
1214{
1215 struct rdma_ucm_init_qp_attr cmd;
1216 struct ib_uverbs_qp_attr resp;
1217 struct ucma_context *ctx;
1218 struct ib_qp_attr qp_attr;
1219 int ret;
1220
1221 if (out_len < sizeof(resp))
1222 return -ENOSPC;
1223
1224 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1225 return -EFAULT;
1226
1227 if (cmd.qp_state > IB_QPS_ERR)
1228 return -EINVAL;
1229
1230 ctx = ucma_get_ctx_dev(file, cmd.id);
1231 if (IS_ERR(ctx))
1232 return PTR_ERR(ctx);
1233
1234 resp.qp_attr_mask = 0;
1235 memset(&qp_attr, 0, sizeof qp_attr);
1236 qp_attr.qp_state = cmd.qp_state;
1237 mutex_lock(&ctx->mutex);
1238 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1239 mutex_unlock(&ctx->mutex);
1240 if (ret)
1241 goto out;
1242
1243 ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1244 if (copy_to_user(u64_to_user_ptr(cmd.response),
1245 &resp, sizeof(resp)))
1246 ret = -EFAULT;
1247
1248out:
1249 ucma_put_ctx(ctx);
1250 return ret;
1251}
1252
1253static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1254 void *optval, size_t optlen)
1255{
1256 int ret = 0;
1257
1258 switch (optname) {
1259 case RDMA_OPTION_ID_TOS:
1260 if (optlen != sizeof(u8)) {
1261 ret = -EINVAL;
1262 break;
1263 }
1264 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1265 break;
1266 case RDMA_OPTION_ID_REUSEADDR:
1267 if (optlen != sizeof(int)) {
1268 ret = -EINVAL;
1269 break;
1270 }
1271 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1272 break;
1273 case RDMA_OPTION_ID_AFONLY:
1274 if (optlen != sizeof(int)) {
1275 ret = -EINVAL;
1276 break;
1277 }
1278 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1279 break;
1280 case RDMA_OPTION_ID_ACK_TIMEOUT:
1281 if (optlen != sizeof(u8)) {
1282 ret = -EINVAL;
1283 break;
1284 }
1285 ret = rdma_set_ack_timeout(ctx->cm_id, *((u8 *)optval));
1286 break;
1287 default:
1288 ret = -ENOSYS;
1289 }
1290
1291 return ret;
1292}
1293
1294static int ucma_set_ib_path(struct ucma_context *ctx,
1295 struct ib_path_rec_data *path_data, size_t optlen)
1296{
1297 struct sa_path_rec sa_path;
1298 struct rdma_cm_event event;
1299 int ret;
1300
1301 if (optlen % sizeof(*path_data))
1302 return -EINVAL;
1303
1304 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1305 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1306 IB_PATH_BIDIRECTIONAL))
1307 break;
1308 }
1309
1310 if (!optlen)
1311 return -EINVAL;
1312
1313 if (!ctx->cm_id->device)
1314 return -EINVAL;
1315
1316 memset(&sa_path, 0, sizeof(sa_path));
1317
1318 sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1319 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1320
1321 if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1322 struct sa_path_rec opa;
1323
1324 sa_convert_path_ib_to_opa(&opa, &sa_path);
1325 mutex_lock(&ctx->mutex);
1326 ret = rdma_set_ib_path(ctx->cm_id, &opa);
1327 mutex_unlock(&ctx->mutex);
1328 } else {
1329 mutex_lock(&ctx->mutex);
1330 ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
1331 mutex_unlock(&ctx->mutex);
1332 }
1333 if (ret)
1334 return ret;
1335
1336 memset(&event, 0, sizeof event);
1337 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1338 return ucma_event_handler(ctx->cm_id, &event);
1339}
1340
1341static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1342 void *optval, size_t optlen)
1343{
1344 int ret;
1345
1346 switch (optname) {
1347 case RDMA_OPTION_IB_PATH:
1348 ret = ucma_set_ib_path(ctx, optval, optlen);
1349 break;
1350 default:
1351 ret = -ENOSYS;
1352 }
1353
1354 return ret;
1355}
1356
1357static int ucma_set_option_level(struct ucma_context *ctx, int level,
1358 int optname, void *optval, size_t optlen)
1359{
1360 int ret;
1361
1362 switch (level) {
1363 case RDMA_OPTION_ID:
1364 mutex_lock(&ctx->mutex);
1365 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1366 mutex_unlock(&ctx->mutex);
1367 break;
1368 case RDMA_OPTION_IB:
1369 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1370 break;
1371 default:
1372 ret = -ENOSYS;
1373 }
1374
1375 return ret;
1376}
1377
1378static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1379 int in_len, int out_len)
1380{
1381 struct rdma_ucm_set_option cmd;
1382 struct ucma_context *ctx;
1383 void *optval;
1384 int ret;
1385
1386 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1387 return -EFAULT;
1388
1389 if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1390 return -EINVAL;
1391
1392 ctx = ucma_get_ctx(file, cmd.id);
1393 if (IS_ERR(ctx))
1394 return PTR_ERR(ctx);
1395
1396 optval = memdup_user(u64_to_user_ptr(cmd.optval),
1397 cmd.optlen);
1398 if (IS_ERR(optval)) {
1399 ret = PTR_ERR(optval);
1400 goto out;
1401 }
1402
1403 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1404 cmd.optlen);
1405 kfree(optval);
1406
1407out:
1408 ucma_put_ctx(ctx);
1409 return ret;
1410}
1411
1412static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1413 int in_len, int out_len)
1414{
1415 struct rdma_ucm_notify cmd;
1416 struct ucma_context *ctx;
1417 int ret = -EINVAL;
1418
1419 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1420 return -EFAULT;
1421
1422 ctx = ucma_get_ctx(file, cmd.id);
1423 if (IS_ERR(ctx))
1424 return PTR_ERR(ctx);
1425
1426 mutex_lock(&ctx->mutex);
1427 if (ctx->cm_id->device)
1428 ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
1429 mutex_unlock(&ctx->mutex);
1430
1431 ucma_put_ctx(ctx);
1432 return ret;
1433}
1434
1435static ssize_t ucma_process_join(struct ucma_file *file,
1436 struct rdma_ucm_join_mcast *cmd, int out_len)
1437{
1438 struct rdma_ucm_create_id_resp resp;
1439 struct ucma_context *ctx;
1440 struct ucma_multicast *mc;
1441 struct sockaddr *addr;
1442 int ret;
1443 u8 join_state;
1444
1445 if (out_len < sizeof(resp))
1446 return -ENOSPC;
1447
1448 addr = (struct sockaddr *) &cmd->addr;
1449 if (cmd->addr_size != rdma_addr_size(addr))
1450 return -EINVAL;
1451
1452 if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1453 join_state = BIT(FULLMEMBER_JOIN);
1454 else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1455 join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1456 else
1457 return -EINVAL;
1458
1459 ctx = ucma_get_ctx_dev(file, cmd->id);
1460 if (IS_ERR(ctx))
1461 return PTR_ERR(ctx);
1462
1463 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
1464 if (!mc) {
1465 ret = -ENOMEM;
1466 goto err_put_ctx;
1467 }
1468
1469 mc->ctx = ctx;
1470 mc->join_state = join_state;
1471 mc->uid = cmd->uid;
1472 memcpy(&mc->addr, addr, cmd->addr_size);
1473
1474 xa_lock(&multicast_table);
1475 if (__xa_alloc(&multicast_table, &mc->id, NULL, xa_limit_32b,
1476 GFP_KERNEL)) {
1477 ret = -ENOMEM;
1478 goto err_free_mc;
1479 }
1480
1481 list_add_tail(&mc->list, &ctx->mc_list);
1482 xa_unlock(&multicast_table);
1483
1484 mutex_lock(&ctx->mutex);
1485 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1486 join_state, mc);
1487 mutex_unlock(&ctx->mutex);
1488 if (ret)
1489 goto err_xa_erase;
1490
1491 resp.id = mc->id;
1492 if (copy_to_user(u64_to_user_ptr(cmd->response),
1493 &resp, sizeof(resp))) {
1494 ret = -EFAULT;
1495 goto err_leave_multicast;
1496 }
1497
1498 xa_store(&multicast_table, mc->id, mc, 0);
1499
1500 ucma_put_ctx(ctx);
1501 return 0;
1502
1503err_leave_multicast:
1504 mutex_lock(&ctx->mutex);
1505 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1506 mutex_unlock(&ctx->mutex);
1507 ucma_cleanup_mc_events(mc);
1508err_xa_erase:
1509 xa_lock(&multicast_table);
1510 list_del(&mc->list);
1511 __xa_erase(&multicast_table, mc->id);
1512err_free_mc:
1513 xa_unlock(&multicast_table);
1514 kfree(mc);
1515err_put_ctx:
1516 ucma_put_ctx(ctx);
1517 return ret;
1518}
1519
1520static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1521 const char __user *inbuf,
1522 int in_len, int out_len)
1523{
1524 struct rdma_ucm_join_ip_mcast cmd;
1525 struct rdma_ucm_join_mcast join_cmd;
1526
1527 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1528 return -EFAULT;
1529
1530 join_cmd.response = cmd.response;
1531 join_cmd.uid = cmd.uid;
1532 join_cmd.id = cmd.id;
1533 join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1534 if (!join_cmd.addr_size)
1535 return -EINVAL;
1536
1537 join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1538 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1539
1540 return ucma_process_join(file, &join_cmd, out_len);
1541}
1542
1543static ssize_t ucma_join_multicast(struct ucma_file *file,
1544 const char __user *inbuf,
1545 int in_len, int out_len)
1546{
1547 struct rdma_ucm_join_mcast cmd;
1548
1549 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1550 return -EFAULT;
1551
1552 if (!rdma_addr_size_kss(&cmd.addr))
1553 return -EINVAL;
1554
1555 return ucma_process_join(file, &cmd, out_len);
1556}
1557
1558static ssize_t ucma_leave_multicast(struct ucma_file *file,
1559 const char __user *inbuf,
1560 int in_len, int out_len)
1561{
1562 struct rdma_ucm_destroy_id cmd;
1563 struct rdma_ucm_destroy_id_resp resp;
1564 struct ucma_multicast *mc;
1565 int ret = 0;
1566
1567 if (out_len < sizeof(resp))
1568 return -ENOSPC;
1569
1570 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1571 return -EFAULT;
1572
1573 xa_lock(&multicast_table);
1574 mc = xa_load(&multicast_table, cmd.id);
1575 if (!mc)
1576 mc = ERR_PTR(-ENOENT);
1577 else if (READ_ONCE(mc->ctx->file) != file)
1578 mc = ERR_PTR(-EINVAL);
1579 else if (!refcount_inc_not_zero(&mc->ctx->ref))
1580 mc = ERR_PTR(-ENXIO);
1581
1582 if (IS_ERR(mc)) {
1583 xa_unlock(&multicast_table);
1584 ret = PTR_ERR(mc);
1585 goto out;
1586 }
1587
1588 list_del(&mc->list);
1589 __xa_erase(&multicast_table, mc->id);
1590 xa_unlock(&multicast_table);
1591
1592 mutex_lock(&mc->ctx->mutex);
1593 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1594 mutex_unlock(&mc->ctx->mutex);
1595
1596 ucma_cleanup_mc_events(mc);
1597
1598 ucma_put_ctx(mc->ctx);
1599 resp.events_reported = mc->events_reported;
1600 kfree(mc);
1601
1602 if (copy_to_user(u64_to_user_ptr(cmd.response),
1603 &resp, sizeof(resp)))
1604 ret = -EFAULT;
1605out:
1606 return ret;
1607}
1608
1609static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1610 const char __user *inbuf,
1611 int in_len, int out_len)
1612{
1613 struct rdma_ucm_migrate_id cmd;
1614 struct rdma_ucm_migrate_resp resp;
1615 struct ucma_event *uevent, *tmp;
1616 struct ucma_context *ctx;
1617 LIST_HEAD(event_list);
1618 struct fd f;
1619 struct ucma_file *cur_file;
1620 int ret = 0;
1621
1622 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1623 return -EFAULT;
1624
1625 /* Get current fd to protect against it being closed */
1626 f = fdget(cmd.fd);
1627 if (!f.file)
1628 return -ENOENT;
1629 if (f.file->f_op != &ucma_fops) {
1630 ret = -EINVAL;
1631 goto file_put;
1632 }
1633 cur_file = f.file->private_data;
1634
1635 /* Validate current fd and prevent destruction of id. */
1636 ctx = ucma_get_ctx(cur_file, cmd.id);
1637 if (IS_ERR(ctx)) {
1638 ret = PTR_ERR(ctx);
1639 goto file_put;
1640 }
1641
1642 rdma_lock_handler(ctx->cm_id);
1643 /*
1644 * ctx->file can only be changed under the handler & xa_lock. xa_load()
1645 * must be checked again to ensure the ctx hasn't begun destruction
1646 * since the ucma_get_ctx().
1647 */
1648 xa_lock(&ctx_table);
1649 if (_ucma_find_context(cmd.id, cur_file) != ctx) {
1650 xa_unlock(&ctx_table);
1651 ret = -ENOENT;
1652 goto err_unlock;
1653 }
1654 ctx->file = new_file;
1655 xa_unlock(&ctx_table);
1656
1657 mutex_lock(&cur_file->mut);
1658 list_del(&ctx->list);
1659 /*
1660 * At this point lock_handler() prevents addition of new uevents for
1661 * this ctx.
1662 */
1663 list_for_each_entry_safe(uevent, tmp, &cur_file->event_list, list)
1664 if (uevent->ctx == ctx)
1665 list_move_tail(&uevent->list, &event_list);
1666 resp.events_reported = ctx->events_reported;
1667 mutex_unlock(&cur_file->mut);
1668
1669 mutex_lock(&new_file->mut);
1670 list_add_tail(&ctx->list, &new_file->ctx_list);
1671 list_splice_tail(&event_list, &new_file->event_list);
1672 mutex_unlock(&new_file->mut);
1673
1674 if (copy_to_user(u64_to_user_ptr(cmd.response),
1675 &resp, sizeof(resp)))
1676 ret = -EFAULT;
1677
1678err_unlock:
1679 rdma_unlock_handler(ctx->cm_id);
1680 ucma_put_ctx(ctx);
1681file_put:
1682 fdput(f);
1683 return ret;
1684}
1685
1686static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1687 const char __user *inbuf,
1688 int in_len, int out_len) = {
1689 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1690 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1691 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1692 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1693 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1694 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1695 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1696 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1697 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1698 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1699 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1700 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1701 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1702 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1703 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1704 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1705 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1706 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1707 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1708 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1709 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1710 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1711 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1712};
1713
1714static ssize_t ucma_write(struct file *filp, const char __user *buf,
1715 size_t len, loff_t *pos)
1716{
1717 struct ucma_file *file = filp->private_data;
1718 struct rdma_ucm_cmd_hdr hdr;
1719 ssize_t ret;
1720
1721 if (!ib_safe_file_access(filp)) {
1722 pr_err_once("%s: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1723 __func__, task_tgid_vnr(current), current->comm);
1724 return -EACCES;
1725 }
1726
1727 if (len < sizeof(hdr))
1728 return -EINVAL;
1729
1730 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1731 return -EFAULT;
1732
1733 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1734 return -EINVAL;
1735 hdr.cmd = array_index_nospec(hdr.cmd, ARRAY_SIZE(ucma_cmd_table));
1736
1737 if (hdr.in + sizeof(hdr) > len)
1738 return -EINVAL;
1739
1740 if (!ucma_cmd_table[hdr.cmd])
1741 return -ENOSYS;
1742
1743 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1744 if (!ret)
1745 ret = len;
1746
1747 return ret;
1748}
1749
1750static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1751{
1752 struct ucma_file *file = filp->private_data;
1753 __poll_t mask = 0;
1754
1755 poll_wait(filp, &file->poll_wait, wait);
1756
1757 if (!list_empty(&file->event_list))
1758 mask = EPOLLIN | EPOLLRDNORM;
1759
1760 return mask;
1761}
1762
1763/*
1764 * ucma_open() does not need the BKL:
1765 *
1766 * - no global state is referred to;
1767 * - there is no ioctl method to race against;
1768 * - no further module initialization is required for open to work
1769 * after the device is registered.
1770 */
1771static int ucma_open(struct inode *inode, struct file *filp)
1772{
1773 struct ucma_file *file;
1774
1775 file = kmalloc(sizeof *file, GFP_KERNEL);
1776 if (!file)
1777 return -ENOMEM;
1778
1779 INIT_LIST_HEAD(&file->event_list);
1780 INIT_LIST_HEAD(&file->ctx_list);
1781 init_waitqueue_head(&file->poll_wait);
1782 mutex_init(&file->mut);
1783
1784 filp->private_data = file;
1785 file->filp = filp;
1786
1787 return stream_open(inode, filp);
1788}
1789
1790static int ucma_close(struct inode *inode, struct file *filp)
1791{
1792 struct ucma_file *file = filp->private_data;
1793
1794 /*
1795 * All paths that touch ctx_list or ctx_list starting from write() are
1796 * prevented by this being a FD release function. The list_add_tail() in
1797 * ucma_connect_event_handler() can run concurrently, however it only
1798 * adds to the list *after* a listening ID. By only reading the first of
1799 * the list, and relying on ucma_destroy_private_ctx() to block
1800 * ucma_connect_event_handler(), no additional locking is needed.
1801 */
1802 while (!list_empty(&file->ctx_list)) {
1803 struct ucma_context *ctx = list_first_entry(
1804 &file->ctx_list, struct ucma_context, list);
1805
1806 WARN_ON(xa_cmpxchg(&ctx_table, ctx->id, ctx, XA_ZERO_ENTRY,
1807 GFP_KERNEL) != ctx);
1808 ucma_destroy_private_ctx(ctx);
1809 }
1810 kfree(file);
1811 return 0;
1812}
1813
1814static const struct file_operations ucma_fops = {
1815 .owner = THIS_MODULE,
1816 .open = ucma_open,
1817 .release = ucma_close,
1818 .write = ucma_write,
1819 .poll = ucma_poll,
1820 .llseek = no_llseek,
1821};
1822
1823static struct miscdevice ucma_misc = {
1824 .minor = MISC_DYNAMIC_MINOR,
1825 .name = "rdma_cm",
1826 .nodename = "infiniband/rdma_cm",
1827 .mode = 0666,
1828 .fops = &ucma_fops,
1829};
1830
1831static int ucma_get_global_nl_info(struct ib_client_nl_info *res)
1832{
1833 res->abi = RDMA_USER_CM_ABI_VERSION;
1834 res->cdev = ucma_misc.this_device;
1835 return 0;
1836}
1837
1838static struct ib_client rdma_cma_client = {
1839 .name = "rdma_cm",
1840 .get_global_nl_info = ucma_get_global_nl_info,
1841};
1842MODULE_ALIAS_RDMA_CLIENT("rdma_cm");
1843
1844static ssize_t abi_version_show(struct device *dev,
1845 struct device_attribute *attr, char *buf)
1846{
1847 return sysfs_emit(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1848}
1849static DEVICE_ATTR_RO(abi_version);
1850
1851static int __init ucma_init(void)
1852{
1853 int ret;
1854
1855 ret = misc_register(&ucma_misc);
1856 if (ret)
1857 return ret;
1858
1859 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1860 if (ret) {
1861 pr_err("rdma_ucm: couldn't create abi_version attr\n");
1862 goto err1;
1863 }
1864
1865 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1866 if (!ucma_ctl_table_hdr) {
1867 pr_err("rdma_ucm: couldn't register sysctl paths\n");
1868 ret = -ENOMEM;
1869 goto err2;
1870 }
1871
1872 ret = ib_register_client(&rdma_cma_client);
1873 if (ret)
1874 goto err3;
1875
1876 return 0;
1877err3:
1878 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1879err2:
1880 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1881err1:
1882 misc_deregister(&ucma_misc);
1883 return ret;
1884}
1885
1886static void __exit ucma_cleanup(void)
1887{
1888 ib_unregister_client(&rdma_cma_client);
1889 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1890 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1891 misc_deregister(&ucma_misc);
1892}
1893
1894module_init(ucma_init);
1895module_exit(ucma_cleanup);
1/*
2 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33#include <linux/completion.h>
34#include <linux/file.h>
35#include <linux/mutex.h>
36#include <linux/poll.h>
37#include <linux/sched.h>
38#include <linux/idr.h>
39#include <linux/in.h>
40#include <linux/in6.h>
41#include <linux/miscdevice.h>
42#include <linux/slab.h>
43#include <linux/sysctl.h>
44#include <linux/module.h>
45#include <linux/nsproxy.h>
46
47#include <linux/nospec.h>
48
49#include <rdma/rdma_user_cm.h>
50#include <rdma/ib_marshall.h>
51#include <rdma/rdma_cm.h>
52#include <rdma/rdma_cm_ib.h>
53#include <rdma/ib_addr.h>
54#include <rdma/ib.h>
55#include <rdma/ib_cm.h>
56#include <rdma/rdma_netlink.h>
57#include "core_priv.h"
58
59MODULE_AUTHOR("Sean Hefty");
60MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
61MODULE_LICENSE("Dual BSD/GPL");
62
63static unsigned int max_backlog = 1024;
64
65static struct ctl_table_header *ucma_ctl_table_hdr;
66static struct ctl_table ucma_ctl_table[] = {
67 {
68 .procname = "max_backlog",
69 .data = &max_backlog,
70 .maxlen = sizeof max_backlog,
71 .mode = 0644,
72 .proc_handler = proc_dointvec,
73 },
74 { }
75};
76
77struct ucma_file {
78 struct mutex mut;
79 struct file *filp;
80 struct list_head ctx_list;
81 struct list_head event_list;
82 wait_queue_head_t poll_wait;
83};
84
85struct ucma_context {
86 u32 id;
87 struct completion comp;
88 refcount_t ref;
89 int events_reported;
90 atomic_t backlog;
91
92 struct ucma_file *file;
93 struct rdma_cm_id *cm_id;
94 struct mutex mutex;
95 u64 uid;
96
97 struct list_head list;
98 struct list_head mc_list;
99 struct work_struct close_work;
100};
101
102struct ucma_multicast {
103 struct ucma_context *ctx;
104 u32 id;
105 int events_reported;
106
107 u64 uid;
108 u8 join_state;
109 struct list_head list;
110 struct sockaddr_storage addr;
111};
112
113struct ucma_event {
114 struct ucma_context *ctx;
115 struct ucma_context *conn_req_ctx;
116 struct ucma_multicast *mc;
117 struct list_head list;
118 struct rdma_ucm_event_resp resp;
119};
120
121static DEFINE_XARRAY_ALLOC(ctx_table);
122static DEFINE_XARRAY_ALLOC(multicast_table);
123
124static const struct file_operations ucma_fops;
125static int ucma_destroy_private_ctx(struct ucma_context *ctx);
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 = xa_load(&ctx_table, 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 xa_lock(&ctx_table);
145 ctx = _ucma_find_context(id, file);
146 if (!IS_ERR(ctx))
147 if (!refcount_inc_not_zero(&ctx->ref))
148 ctx = ERR_PTR(-ENXIO);
149 xa_unlock(&ctx_table);
150 return ctx;
151}
152
153static void ucma_put_ctx(struct ucma_context *ctx)
154{
155 if (refcount_dec_and_test(&ctx->ref))
156 complete(&ctx->comp);
157}
158
159/*
160 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
161 * CM_ID is bound.
162 */
163static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
164{
165 struct ucma_context *ctx = ucma_get_ctx(file, id);
166
167 if (IS_ERR(ctx))
168 return ctx;
169 if (!ctx->cm_id->device) {
170 ucma_put_ctx(ctx);
171 return ERR_PTR(-EINVAL);
172 }
173 return ctx;
174}
175
176static void ucma_close_id(struct work_struct *work)
177{
178 struct ucma_context *ctx = container_of(work, struct ucma_context, close_work);
179
180 /* once all inflight tasks are finished, we close all underlying
181 * resources. The context is still alive till its explicit destryoing
182 * by its creator. This puts back the xarray's reference.
183 */
184 ucma_put_ctx(ctx);
185 wait_for_completion(&ctx->comp);
186 /* No new events will be generated after destroying the id. */
187 rdma_destroy_id(ctx->cm_id);
188
189 /* Reading the cm_id without holding a positive ref is not allowed */
190 ctx->cm_id = NULL;
191}
192
193static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
194{
195 struct ucma_context *ctx;
196
197 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
198 if (!ctx)
199 return NULL;
200
201 INIT_WORK(&ctx->close_work, ucma_close_id);
202 init_completion(&ctx->comp);
203 INIT_LIST_HEAD(&ctx->mc_list);
204 /* So list_del() will work if we don't do ucma_finish_ctx() */
205 INIT_LIST_HEAD(&ctx->list);
206 ctx->file = file;
207 mutex_init(&ctx->mutex);
208
209 if (xa_alloc(&ctx_table, &ctx->id, NULL, xa_limit_32b, GFP_KERNEL)) {
210 kfree(ctx);
211 return NULL;
212 }
213 return ctx;
214}
215
216static void ucma_set_ctx_cm_id(struct ucma_context *ctx,
217 struct rdma_cm_id *cm_id)
218{
219 refcount_set(&ctx->ref, 1);
220 ctx->cm_id = cm_id;
221}
222
223static void ucma_finish_ctx(struct ucma_context *ctx)
224{
225 lockdep_assert_held(&ctx->file->mut);
226 list_add_tail(&ctx->list, &ctx->file->ctx_list);
227 xa_store(&ctx_table, ctx->id, ctx, GFP_KERNEL);
228}
229
230static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
231 struct rdma_conn_param *src)
232{
233 if (src->private_data_len)
234 memcpy(dst->private_data, src->private_data,
235 src->private_data_len);
236 dst->private_data_len = src->private_data_len;
237 dst->responder_resources = src->responder_resources;
238 dst->initiator_depth = src->initiator_depth;
239 dst->flow_control = src->flow_control;
240 dst->retry_count = src->retry_count;
241 dst->rnr_retry_count = src->rnr_retry_count;
242 dst->srq = src->srq;
243 dst->qp_num = src->qp_num;
244}
245
246static void ucma_copy_ud_event(struct ib_device *device,
247 struct rdma_ucm_ud_param *dst,
248 struct rdma_ud_param *src)
249{
250 if (src->private_data_len)
251 memcpy(dst->private_data, src->private_data,
252 src->private_data_len);
253 dst->private_data_len = src->private_data_len;
254 ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
255 dst->qp_num = src->qp_num;
256 dst->qkey = src->qkey;
257}
258
259static struct ucma_event *ucma_create_uevent(struct ucma_context *ctx,
260 struct rdma_cm_event *event)
261{
262 struct ucma_event *uevent;
263
264 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
265 if (!uevent)
266 return NULL;
267
268 uevent->ctx = ctx;
269 switch (event->event) {
270 case RDMA_CM_EVENT_MULTICAST_JOIN:
271 case RDMA_CM_EVENT_MULTICAST_ERROR:
272 uevent->mc = (struct ucma_multicast *)
273 event->param.ud.private_data;
274 uevent->resp.uid = uevent->mc->uid;
275 uevent->resp.id = uevent->mc->id;
276 break;
277 default:
278 uevent->resp.uid = ctx->uid;
279 uevent->resp.id = ctx->id;
280 break;
281 }
282 uevent->resp.event = event->event;
283 uevent->resp.status = event->status;
284 if (ctx->cm_id->qp_type == IB_QPT_UD)
285 ucma_copy_ud_event(ctx->cm_id->device, &uevent->resp.param.ud,
286 &event->param.ud);
287 else
288 ucma_copy_conn_event(&uevent->resp.param.conn,
289 &event->param.conn);
290
291 uevent->resp.ece.vendor_id = event->ece.vendor_id;
292 uevent->resp.ece.attr_mod = event->ece.attr_mod;
293 return uevent;
294}
295
296static int ucma_connect_event_handler(struct rdma_cm_id *cm_id,
297 struct rdma_cm_event *event)
298{
299 struct ucma_context *listen_ctx = cm_id->context;
300 struct ucma_context *ctx;
301 struct ucma_event *uevent;
302
303 if (!atomic_add_unless(&listen_ctx->backlog, -1, 0))
304 return -ENOMEM;
305 ctx = ucma_alloc_ctx(listen_ctx->file);
306 if (!ctx)
307 goto err_backlog;
308 ucma_set_ctx_cm_id(ctx, cm_id);
309
310 uevent = ucma_create_uevent(listen_ctx, event);
311 if (!uevent)
312 goto err_alloc;
313 uevent->conn_req_ctx = ctx;
314 uevent->resp.id = ctx->id;
315
316 ctx->cm_id->context = ctx;
317
318 mutex_lock(&ctx->file->mut);
319 ucma_finish_ctx(ctx);
320 list_add_tail(&uevent->list, &ctx->file->event_list);
321 mutex_unlock(&ctx->file->mut);
322 wake_up_interruptible(&ctx->file->poll_wait);
323 return 0;
324
325err_alloc:
326 ucma_destroy_private_ctx(ctx);
327err_backlog:
328 atomic_inc(&listen_ctx->backlog);
329 /* Returning error causes the new ID to be destroyed */
330 return -ENOMEM;
331}
332
333static int ucma_event_handler(struct rdma_cm_id *cm_id,
334 struct rdma_cm_event *event)
335{
336 struct ucma_event *uevent;
337 struct ucma_context *ctx = cm_id->context;
338
339 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST)
340 return ucma_connect_event_handler(cm_id, event);
341
342 /*
343 * We ignore events for new connections until userspace has set their
344 * context. This can only happen if an error occurs on a new connection
345 * before the user accepts it. This is okay, since the accept will just
346 * fail later. However, we do need to release the underlying HW
347 * resources in case of a device removal event.
348 */
349 if (ctx->uid) {
350 uevent = ucma_create_uevent(ctx, event);
351 if (!uevent)
352 return 0;
353
354 mutex_lock(&ctx->file->mut);
355 list_add_tail(&uevent->list, &ctx->file->event_list);
356 mutex_unlock(&ctx->file->mut);
357 wake_up_interruptible(&ctx->file->poll_wait);
358 }
359
360 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL) {
361 xa_lock(&ctx_table);
362 if (xa_load(&ctx_table, ctx->id) == ctx)
363 queue_work(system_unbound_wq, &ctx->close_work);
364 xa_unlock(&ctx_table);
365 }
366 return 0;
367}
368
369static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
370 int in_len, int out_len)
371{
372 struct rdma_ucm_get_event cmd;
373 struct ucma_event *uevent;
374
375 /*
376 * Old 32 bit user space does not send the 4 byte padding in the
377 * reserved field. We don't care, allow it to keep working.
378 */
379 if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved) -
380 sizeof(uevent->resp.ece))
381 return -ENOSPC;
382
383 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
384 return -EFAULT;
385
386 mutex_lock(&file->mut);
387 while (list_empty(&file->event_list)) {
388 mutex_unlock(&file->mut);
389
390 if (file->filp->f_flags & O_NONBLOCK)
391 return -EAGAIN;
392
393 if (wait_event_interruptible(file->poll_wait,
394 !list_empty(&file->event_list)))
395 return -ERESTARTSYS;
396
397 mutex_lock(&file->mut);
398 }
399
400 uevent = list_first_entry(&file->event_list, struct ucma_event, list);
401
402 if (copy_to_user(u64_to_user_ptr(cmd.response),
403 &uevent->resp,
404 min_t(size_t, out_len, sizeof(uevent->resp)))) {
405 mutex_unlock(&file->mut);
406 return -EFAULT;
407 }
408
409 list_del(&uevent->list);
410 uevent->ctx->events_reported++;
411 if (uevent->mc)
412 uevent->mc->events_reported++;
413 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
414 atomic_inc(&uevent->ctx->backlog);
415 mutex_unlock(&file->mut);
416
417 kfree(uevent);
418 return 0;
419}
420
421static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
422{
423 switch (cmd->ps) {
424 case RDMA_PS_TCP:
425 *qp_type = IB_QPT_RC;
426 return 0;
427 case RDMA_PS_UDP:
428 case RDMA_PS_IPOIB:
429 *qp_type = IB_QPT_UD;
430 return 0;
431 case RDMA_PS_IB:
432 *qp_type = cmd->qp_type;
433 return 0;
434 default:
435 return -EINVAL;
436 }
437}
438
439static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
440 int in_len, int out_len)
441{
442 struct rdma_ucm_create_id cmd;
443 struct rdma_ucm_create_id_resp resp;
444 struct ucma_context *ctx;
445 struct rdma_cm_id *cm_id;
446 enum ib_qp_type qp_type;
447 int ret;
448
449 if (out_len < sizeof(resp))
450 return -ENOSPC;
451
452 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
453 return -EFAULT;
454
455 ret = ucma_get_qp_type(&cmd, &qp_type);
456 if (ret)
457 return ret;
458
459 ctx = ucma_alloc_ctx(file);
460 if (!ctx)
461 return -ENOMEM;
462
463 ctx->uid = cmd.uid;
464 cm_id = rdma_create_user_id(ucma_event_handler, ctx, cmd.ps, qp_type);
465 if (IS_ERR(cm_id)) {
466 ret = PTR_ERR(cm_id);
467 goto err1;
468 }
469 ucma_set_ctx_cm_id(ctx, cm_id);
470
471 resp.id = ctx->id;
472 if (copy_to_user(u64_to_user_ptr(cmd.response),
473 &resp, sizeof(resp))) {
474 ret = -EFAULT;
475 goto err1;
476 }
477
478 mutex_lock(&file->mut);
479 ucma_finish_ctx(ctx);
480 mutex_unlock(&file->mut);
481 return 0;
482
483err1:
484 ucma_destroy_private_ctx(ctx);
485 return ret;
486}
487
488static void ucma_cleanup_multicast(struct ucma_context *ctx)
489{
490 struct ucma_multicast *mc, *tmp;
491
492 xa_lock(&multicast_table);
493 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
494 list_del(&mc->list);
495 /*
496 * At this point mc->ctx->ref is 0 so the mc cannot leave the
497 * lock on the reader and this is enough serialization
498 */
499 __xa_erase(&multicast_table, mc->id);
500 kfree(mc);
501 }
502 xa_unlock(&multicast_table);
503}
504
505static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
506{
507 struct ucma_event *uevent, *tmp;
508
509 rdma_lock_handler(mc->ctx->cm_id);
510 mutex_lock(&mc->ctx->file->mut);
511 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
512 if (uevent->mc != mc)
513 continue;
514
515 list_del(&uevent->list);
516 kfree(uevent);
517 }
518 mutex_unlock(&mc->ctx->file->mut);
519 rdma_unlock_handler(mc->ctx->cm_id);
520}
521
522static int ucma_cleanup_ctx_events(struct ucma_context *ctx)
523{
524 int events_reported;
525 struct ucma_event *uevent, *tmp;
526 LIST_HEAD(list);
527
528 /* Cleanup events not yet reported to the user.*/
529 mutex_lock(&ctx->file->mut);
530 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
531 if (uevent->ctx != ctx)
532 continue;
533
534 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST &&
535 xa_cmpxchg(&ctx_table, uevent->conn_req_ctx->id,
536 uevent->conn_req_ctx, XA_ZERO_ENTRY,
537 GFP_KERNEL) == uevent->conn_req_ctx) {
538 list_move_tail(&uevent->list, &list);
539 continue;
540 }
541 list_del(&uevent->list);
542 kfree(uevent);
543 }
544 list_del(&ctx->list);
545 events_reported = ctx->events_reported;
546 mutex_unlock(&ctx->file->mut);
547
548 /*
549 * If this was a listening ID then any connections spawned from it that
550 * have not been delivered to userspace are cleaned up too. Must be done
551 * outside any locks.
552 */
553 list_for_each_entry_safe(uevent, tmp, &list, list) {
554 ucma_destroy_private_ctx(uevent->conn_req_ctx);
555 kfree(uevent);
556 }
557 return events_reported;
558}
559
560/*
561 * When this is called the xarray must have a XA_ZERO_ENTRY in the ctx->id (ie
562 * the ctx is not public to the user). This either because:
563 * - ucma_finish_ctx() hasn't been called
564 * - xa_cmpxchg() succeed to remove the entry (only one thread can succeed)
565 */
566static int ucma_destroy_private_ctx(struct ucma_context *ctx)
567{
568 int events_reported;
569
570 /*
571 * Destroy the underlying cm_id. New work queuing is prevented now by
572 * the removal from the xarray. Once the work is cancled ref will either
573 * be 0 because the work ran to completion and consumed the ref from the
574 * xarray, or it will be positive because we still have the ref from the
575 * xarray. This can also be 0 in cases where cm_id was never set
576 */
577 cancel_work_sync(&ctx->close_work);
578 if (refcount_read(&ctx->ref))
579 ucma_close_id(&ctx->close_work);
580
581 events_reported = ucma_cleanup_ctx_events(ctx);
582 ucma_cleanup_multicast(ctx);
583
584 WARN_ON(xa_cmpxchg(&ctx_table, ctx->id, XA_ZERO_ENTRY, NULL,
585 GFP_KERNEL) != NULL);
586 mutex_destroy(&ctx->mutex);
587 kfree(ctx);
588 return events_reported;
589}
590
591static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
592 int in_len, int out_len)
593{
594 struct rdma_ucm_destroy_id cmd;
595 struct rdma_ucm_destroy_id_resp resp;
596 struct ucma_context *ctx;
597 int ret = 0;
598
599 if (out_len < sizeof(resp))
600 return -ENOSPC;
601
602 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
603 return -EFAULT;
604
605 xa_lock(&ctx_table);
606 ctx = _ucma_find_context(cmd.id, file);
607 if (!IS_ERR(ctx)) {
608 if (__xa_cmpxchg(&ctx_table, ctx->id, ctx, XA_ZERO_ENTRY,
609 GFP_KERNEL) != ctx)
610 ctx = ERR_PTR(-ENOENT);
611 }
612 xa_unlock(&ctx_table);
613
614 if (IS_ERR(ctx))
615 return PTR_ERR(ctx);
616
617 resp.events_reported = ucma_destroy_private_ctx(ctx);
618 if (copy_to_user(u64_to_user_ptr(cmd.response),
619 &resp, sizeof(resp)))
620 ret = -EFAULT;
621
622 return ret;
623}
624
625static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
626 int in_len, int out_len)
627{
628 struct rdma_ucm_bind_ip cmd;
629 struct ucma_context *ctx;
630 int ret;
631
632 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
633 return -EFAULT;
634
635 if (!rdma_addr_size_in6(&cmd.addr))
636 return -EINVAL;
637
638 ctx = ucma_get_ctx(file, cmd.id);
639 if (IS_ERR(ctx))
640 return PTR_ERR(ctx);
641
642 mutex_lock(&ctx->mutex);
643 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
644 mutex_unlock(&ctx->mutex);
645
646 ucma_put_ctx(ctx);
647 return ret;
648}
649
650static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
651 int in_len, int out_len)
652{
653 struct rdma_ucm_bind cmd;
654 struct ucma_context *ctx;
655 int ret;
656
657 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
658 return -EFAULT;
659
660 if (cmd.reserved || !cmd.addr_size ||
661 cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
662 return -EINVAL;
663
664 ctx = ucma_get_ctx(file, cmd.id);
665 if (IS_ERR(ctx))
666 return PTR_ERR(ctx);
667
668 mutex_lock(&ctx->mutex);
669 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
670 mutex_unlock(&ctx->mutex);
671 ucma_put_ctx(ctx);
672 return ret;
673}
674
675static ssize_t ucma_resolve_ip(struct ucma_file *file,
676 const char __user *inbuf,
677 int in_len, int out_len)
678{
679 struct rdma_ucm_resolve_ip cmd;
680 struct ucma_context *ctx;
681 int ret;
682
683 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
684 return -EFAULT;
685
686 if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
687 !rdma_addr_size_in6(&cmd.dst_addr))
688 return -EINVAL;
689
690 ctx = ucma_get_ctx(file, cmd.id);
691 if (IS_ERR(ctx))
692 return PTR_ERR(ctx);
693
694 mutex_lock(&ctx->mutex);
695 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
696 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
697 mutex_unlock(&ctx->mutex);
698 ucma_put_ctx(ctx);
699 return ret;
700}
701
702static ssize_t ucma_resolve_addr(struct ucma_file *file,
703 const char __user *inbuf,
704 int in_len, int out_len)
705{
706 struct rdma_ucm_resolve_addr cmd;
707 struct ucma_context *ctx;
708 int ret;
709
710 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
711 return -EFAULT;
712
713 if (cmd.reserved ||
714 (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
715 !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
716 return -EINVAL;
717
718 ctx = ucma_get_ctx(file, cmd.id);
719 if (IS_ERR(ctx))
720 return PTR_ERR(ctx);
721
722 mutex_lock(&ctx->mutex);
723 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
724 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
725 mutex_unlock(&ctx->mutex);
726 ucma_put_ctx(ctx);
727 return ret;
728}
729
730static ssize_t ucma_resolve_route(struct ucma_file *file,
731 const char __user *inbuf,
732 int in_len, int out_len)
733{
734 struct rdma_ucm_resolve_route cmd;
735 struct ucma_context *ctx;
736 int ret;
737
738 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
739 return -EFAULT;
740
741 ctx = ucma_get_ctx_dev(file, cmd.id);
742 if (IS_ERR(ctx))
743 return PTR_ERR(ctx);
744
745 mutex_lock(&ctx->mutex);
746 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
747 mutex_unlock(&ctx->mutex);
748 ucma_put_ctx(ctx);
749 return ret;
750}
751
752static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
753 struct rdma_route *route)
754{
755 struct rdma_dev_addr *dev_addr;
756
757 resp->num_paths = route->num_pri_alt_paths;
758 switch (route->num_pri_alt_paths) {
759 case 0:
760 dev_addr = &route->addr.dev_addr;
761 rdma_addr_get_dgid(dev_addr,
762 (union ib_gid *) &resp->ib_route[0].dgid);
763 rdma_addr_get_sgid(dev_addr,
764 (union ib_gid *) &resp->ib_route[0].sgid);
765 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
766 break;
767 case 2:
768 ib_copy_path_rec_to_user(&resp->ib_route[1],
769 &route->path_rec[1]);
770 fallthrough;
771 case 1:
772 ib_copy_path_rec_to_user(&resp->ib_route[0],
773 &route->path_rec[0]);
774 break;
775 default:
776 break;
777 }
778}
779
780static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
781 struct rdma_route *route)
782{
783
784 resp->num_paths = route->num_pri_alt_paths;
785 switch (route->num_pri_alt_paths) {
786 case 0:
787 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
788 (union ib_gid *)&resp->ib_route[0].dgid);
789 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
790 (union ib_gid *)&resp->ib_route[0].sgid);
791 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
792 break;
793 case 2:
794 ib_copy_path_rec_to_user(&resp->ib_route[1],
795 &route->path_rec[1]);
796 fallthrough;
797 case 1:
798 ib_copy_path_rec_to_user(&resp->ib_route[0],
799 &route->path_rec[0]);
800 break;
801 default:
802 break;
803 }
804}
805
806static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
807 struct rdma_route *route)
808{
809 struct rdma_dev_addr *dev_addr;
810
811 dev_addr = &route->addr.dev_addr;
812 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
813 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
814}
815
816static ssize_t ucma_query_route(struct ucma_file *file,
817 const char __user *inbuf,
818 int in_len, int out_len)
819{
820 struct rdma_ucm_query cmd;
821 struct rdma_ucm_query_route_resp resp;
822 struct ucma_context *ctx;
823 struct sockaddr *addr;
824 int ret = 0;
825
826 if (out_len < offsetof(struct rdma_ucm_query_route_resp, ibdev_index))
827 return -ENOSPC;
828
829 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
830 return -EFAULT;
831
832 ctx = ucma_get_ctx(file, cmd.id);
833 if (IS_ERR(ctx))
834 return PTR_ERR(ctx);
835
836 mutex_lock(&ctx->mutex);
837 memset(&resp, 0, sizeof resp);
838 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
839 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
840 sizeof(struct sockaddr_in) :
841 sizeof(struct sockaddr_in6));
842 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
843 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
844 sizeof(struct sockaddr_in) :
845 sizeof(struct sockaddr_in6));
846 if (!ctx->cm_id->device)
847 goto out;
848
849 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
850 resp.ibdev_index = ctx->cm_id->device->index;
851 resp.port_num = ctx->cm_id->port_num;
852
853 if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
854 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
855 else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
856 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
857 else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
858 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
859
860out:
861 mutex_unlock(&ctx->mutex);
862 if (copy_to_user(u64_to_user_ptr(cmd.response), &resp,
863 min_t(size_t, out_len, sizeof(resp))))
864 ret = -EFAULT;
865
866 ucma_put_ctx(ctx);
867 return ret;
868}
869
870static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
871 struct rdma_ucm_query_addr_resp *resp)
872{
873 if (!cm_id->device)
874 return;
875
876 resp->node_guid = (__force __u64) cm_id->device->node_guid;
877 resp->ibdev_index = cm_id->device->index;
878 resp->port_num = cm_id->port_num;
879 resp->pkey = (__force __u16) cpu_to_be16(
880 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
881}
882
883static ssize_t ucma_query_addr(struct ucma_context *ctx,
884 void __user *response, int out_len)
885{
886 struct rdma_ucm_query_addr_resp resp;
887 struct sockaddr *addr;
888 int ret = 0;
889
890 if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
891 return -ENOSPC;
892
893 memset(&resp, 0, sizeof resp);
894
895 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
896 resp.src_size = rdma_addr_size(addr);
897 memcpy(&resp.src_addr, addr, resp.src_size);
898
899 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
900 resp.dst_size = rdma_addr_size(addr);
901 memcpy(&resp.dst_addr, addr, resp.dst_size);
902
903 ucma_query_device_addr(ctx->cm_id, &resp);
904
905 if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
906 ret = -EFAULT;
907
908 return ret;
909}
910
911static ssize_t ucma_query_path(struct ucma_context *ctx,
912 void __user *response, int out_len)
913{
914 struct rdma_ucm_query_path_resp *resp;
915 int i, ret = 0;
916
917 if (out_len < sizeof(*resp))
918 return -ENOSPC;
919
920 resp = kzalloc(out_len, GFP_KERNEL);
921 if (!resp)
922 return -ENOMEM;
923
924 resp->num_paths = ctx->cm_id->route.num_pri_alt_paths;
925 for (i = 0, out_len -= sizeof(*resp);
926 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
927 i++, out_len -= sizeof(struct ib_path_rec_data)) {
928 struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
929
930 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
931 IB_PATH_BIDIRECTIONAL;
932 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
933 struct sa_path_rec ib;
934
935 sa_convert_path_opa_to_ib(&ib, rec);
936 ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
937
938 } else {
939 ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
940 }
941 }
942
943 if (copy_to_user(response, resp, struct_size(resp, path_data, i)))
944 ret = -EFAULT;
945
946 kfree(resp);
947 return ret;
948}
949
950static ssize_t ucma_query_gid(struct ucma_context *ctx,
951 void __user *response, int out_len)
952{
953 struct rdma_ucm_query_addr_resp resp;
954 struct sockaddr_ib *addr;
955 int ret = 0;
956
957 if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
958 return -ENOSPC;
959
960 memset(&resp, 0, sizeof resp);
961
962 ucma_query_device_addr(ctx->cm_id, &resp);
963
964 addr = (struct sockaddr_ib *) &resp.src_addr;
965 resp.src_size = sizeof(*addr);
966 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
967 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
968 } else {
969 addr->sib_family = AF_IB;
970 addr->sib_pkey = (__force __be16) resp.pkey;
971 rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
972 NULL);
973 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
974 &ctx->cm_id->route.addr.src_addr);
975 }
976
977 addr = (struct sockaddr_ib *) &resp.dst_addr;
978 resp.dst_size = sizeof(*addr);
979 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
980 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
981 } else {
982 addr->sib_family = AF_IB;
983 addr->sib_pkey = (__force __be16) resp.pkey;
984 rdma_read_gids(ctx->cm_id, NULL,
985 (union ib_gid *)&addr->sib_addr);
986 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
987 &ctx->cm_id->route.addr.dst_addr);
988 }
989
990 if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
991 ret = -EFAULT;
992
993 return ret;
994}
995
996static ssize_t ucma_query(struct ucma_file *file,
997 const char __user *inbuf,
998 int in_len, int out_len)
999{
1000 struct rdma_ucm_query cmd;
1001 struct ucma_context *ctx;
1002 void __user *response;
1003 int ret;
1004
1005 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1006 return -EFAULT;
1007
1008 response = u64_to_user_ptr(cmd.response);
1009 ctx = ucma_get_ctx(file, cmd.id);
1010 if (IS_ERR(ctx))
1011 return PTR_ERR(ctx);
1012
1013 mutex_lock(&ctx->mutex);
1014 switch (cmd.option) {
1015 case RDMA_USER_CM_QUERY_ADDR:
1016 ret = ucma_query_addr(ctx, response, out_len);
1017 break;
1018 case RDMA_USER_CM_QUERY_PATH:
1019 ret = ucma_query_path(ctx, response, out_len);
1020 break;
1021 case RDMA_USER_CM_QUERY_GID:
1022 ret = ucma_query_gid(ctx, response, out_len);
1023 break;
1024 default:
1025 ret = -ENOSYS;
1026 break;
1027 }
1028 mutex_unlock(&ctx->mutex);
1029
1030 ucma_put_ctx(ctx);
1031 return ret;
1032}
1033
1034static void ucma_copy_conn_param(struct rdma_cm_id *id,
1035 struct rdma_conn_param *dst,
1036 struct rdma_ucm_conn_param *src)
1037{
1038 dst->private_data = src->private_data;
1039 dst->private_data_len = src->private_data_len;
1040 dst->responder_resources = src->responder_resources;
1041 dst->initiator_depth = src->initiator_depth;
1042 dst->flow_control = src->flow_control;
1043 dst->retry_count = src->retry_count;
1044 dst->rnr_retry_count = src->rnr_retry_count;
1045 dst->srq = src->srq;
1046 dst->qp_num = src->qp_num & 0xFFFFFF;
1047 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1048}
1049
1050static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1051 int in_len, int out_len)
1052{
1053 struct rdma_conn_param conn_param;
1054 struct rdma_ucm_ece ece = {};
1055 struct rdma_ucm_connect cmd;
1056 struct ucma_context *ctx;
1057 size_t in_size;
1058 int ret;
1059
1060 if (in_len < offsetofend(typeof(cmd), reserved))
1061 return -EINVAL;
1062 in_size = min_t(size_t, in_len, sizeof(cmd));
1063 if (copy_from_user(&cmd, inbuf, in_size))
1064 return -EFAULT;
1065
1066 if (!cmd.conn_param.valid)
1067 return -EINVAL;
1068
1069 ctx = ucma_get_ctx_dev(file, cmd.id);
1070 if (IS_ERR(ctx))
1071 return PTR_ERR(ctx);
1072
1073 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1074 if (offsetofend(typeof(cmd), ece) <= in_size) {
1075 ece.vendor_id = cmd.ece.vendor_id;
1076 ece.attr_mod = cmd.ece.attr_mod;
1077 }
1078
1079 mutex_lock(&ctx->mutex);
1080 ret = rdma_connect_ece(ctx->cm_id, &conn_param, &ece);
1081 mutex_unlock(&ctx->mutex);
1082 ucma_put_ctx(ctx);
1083 return ret;
1084}
1085
1086static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1087 int in_len, int out_len)
1088{
1089 struct rdma_ucm_listen cmd;
1090 struct ucma_context *ctx;
1091 int ret;
1092
1093 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1094 return -EFAULT;
1095
1096 ctx = ucma_get_ctx(file, cmd.id);
1097 if (IS_ERR(ctx))
1098 return PTR_ERR(ctx);
1099
1100 if (cmd.backlog <= 0 || cmd.backlog > max_backlog)
1101 cmd.backlog = max_backlog;
1102 atomic_set(&ctx->backlog, cmd.backlog);
1103
1104 mutex_lock(&ctx->mutex);
1105 ret = rdma_listen(ctx->cm_id, cmd.backlog);
1106 mutex_unlock(&ctx->mutex);
1107 ucma_put_ctx(ctx);
1108 return ret;
1109}
1110
1111static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1112 int in_len, int out_len)
1113{
1114 struct rdma_ucm_accept cmd;
1115 struct rdma_conn_param conn_param;
1116 struct rdma_ucm_ece ece = {};
1117 struct ucma_context *ctx;
1118 size_t in_size;
1119 int ret;
1120
1121 if (in_len < offsetofend(typeof(cmd), reserved))
1122 return -EINVAL;
1123 in_size = min_t(size_t, in_len, sizeof(cmd));
1124 if (copy_from_user(&cmd, inbuf, in_size))
1125 return -EFAULT;
1126
1127 ctx = ucma_get_ctx_dev(file, cmd.id);
1128 if (IS_ERR(ctx))
1129 return PTR_ERR(ctx);
1130
1131 if (offsetofend(typeof(cmd), ece) <= in_size) {
1132 ece.vendor_id = cmd.ece.vendor_id;
1133 ece.attr_mod = cmd.ece.attr_mod;
1134 }
1135
1136 if (cmd.conn_param.valid) {
1137 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1138 mutex_lock(&ctx->mutex);
1139 rdma_lock_handler(ctx->cm_id);
1140 ret = rdma_accept_ece(ctx->cm_id, &conn_param, &ece);
1141 if (!ret) {
1142 /* The uid must be set atomically with the handler */
1143 ctx->uid = cmd.uid;
1144 }
1145 rdma_unlock_handler(ctx->cm_id);
1146 mutex_unlock(&ctx->mutex);
1147 } else {
1148 mutex_lock(&ctx->mutex);
1149 rdma_lock_handler(ctx->cm_id);
1150 ret = rdma_accept_ece(ctx->cm_id, NULL, &ece);
1151 rdma_unlock_handler(ctx->cm_id);
1152 mutex_unlock(&ctx->mutex);
1153 }
1154 ucma_put_ctx(ctx);
1155 return ret;
1156}
1157
1158static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1159 int in_len, int out_len)
1160{
1161 struct rdma_ucm_reject cmd;
1162 struct ucma_context *ctx;
1163 int ret;
1164
1165 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1166 return -EFAULT;
1167
1168 if (!cmd.reason)
1169 cmd.reason = IB_CM_REJ_CONSUMER_DEFINED;
1170
1171 switch (cmd.reason) {
1172 case IB_CM_REJ_CONSUMER_DEFINED:
1173 case IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED:
1174 break;
1175 default:
1176 return -EINVAL;
1177 }
1178
1179 ctx = ucma_get_ctx_dev(file, cmd.id);
1180 if (IS_ERR(ctx))
1181 return PTR_ERR(ctx);
1182
1183 mutex_lock(&ctx->mutex);
1184 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len,
1185 cmd.reason);
1186 mutex_unlock(&ctx->mutex);
1187 ucma_put_ctx(ctx);
1188 return ret;
1189}
1190
1191static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1192 int in_len, int out_len)
1193{
1194 struct rdma_ucm_disconnect cmd;
1195 struct ucma_context *ctx;
1196 int ret;
1197
1198 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1199 return -EFAULT;
1200
1201 ctx = ucma_get_ctx_dev(file, cmd.id);
1202 if (IS_ERR(ctx))
1203 return PTR_ERR(ctx);
1204
1205 mutex_lock(&ctx->mutex);
1206 ret = rdma_disconnect(ctx->cm_id);
1207 mutex_unlock(&ctx->mutex);
1208 ucma_put_ctx(ctx);
1209 return ret;
1210}
1211
1212static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1213 const char __user *inbuf,
1214 int in_len, int out_len)
1215{
1216 struct rdma_ucm_init_qp_attr cmd;
1217 struct ib_uverbs_qp_attr resp;
1218 struct ucma_context *ctx;
1219 struct ib_qp_attr qp_attr;
1220 int ret;
1221
1222 if (out_len < sizeof(resp))
1223 return -ENOSPC;
1224
1225 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1226 return -EFAULT;
1227
1228 if (cmd.qp_state > IB_QPS_ERR)
1229 return -EINVAL;
1230
1231 ctx = ucma_get_ctx_dev(file, cmd.id);
1232 if (IS_ERR(ctx))
1233 return PTR_ERR(ctx);
1234
1235 resp.qp_attr_mask = 0;
1236 memset(&qp_attr, 0, sizeof qp_attr);
1237 qp_attr.qp_state = cmd.qp_state;
1238 mutex_lock(&ctx->mutex);
1239 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1240 mutex_unlock(&ctx->mutex);
1241 if (ret)
1242 goto out;
1243
1244 ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1245 if (copy_to_user(u64_to_user_ptr(cmd.response),
1246 &resp, sizeof(resp)))
1247 ret = -EFAULT;
1248
1249out:
1250 ucma_put_ctx(ctx);
1251 return ret;
1252}
1253
1254static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1255 void *optval, size_t optlen)
1256{
1257 int ret = 0;
1258
1259 switch (optname) {
1260 case RDMA_OPTION_ID_TOS:
1261 if (optlen != sizeof(u8)) {
1262 ret = -EINVAL;
1263 break;
1264 }
1265 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1266 break;
1267 case RDMA_OPTION_ID_REUSEADDR:
1268 if (optlen != sizeof(int)) {
1269 ret = -EINVAL;
1270 break;
1271 }
1272 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1273 break;
1274 case RDMA_OPTION_ID_AFONLY:
1275 if (optlen != sizeof(int)) {
1276 ret = -EINVAL;
1277 break;
1278 }
1279 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1280 break;
1281 case RDMA_OPTION_ID_ACK_TIMEOUT:
1282 if (optlen != sizeof(u8)) {
1283 ret = -EINVAL;
1284 break;
1285 }
1286 ret = rdma_set_ack_timeout(ctx->cm_id, *((u8 *)optval));
1287 break;
1288 default:
1289 ret = -ENOSYS;
1290 }
1291
1292 return ret;
1293}
1294
1295static int ucma_set_ib_path(struct ucma_context *ctx,
1296 struct ib_path_rec_data *path_data, size_t optlen)
1297{
1298 struct sa_path_rec sa_path;
1299 struct rdma_cm_event event;
1300 int ret;
1301
1302 if (optlen % sizeof(*path_data))
1303 return -EINVAL;
1304
1305 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1306 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1307 IB_PATH_BIDIRECTIONAL))
1308 break;
1309 }
1310
1311 if (!optlen)
1312 return -EINVAL;
1313
1314 if (!ctx->cm_id->device)
1315 return -EINVAL;
1316
1317 memset(&sa_path, 0, sizeof(sa_path));
1318
1319 sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1320 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1321
1322 if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1323 struct sa_path_rec opa;
1324
1325 sa_convert_path_ib_to_opa(&opa, &sa_path);
1326 mutex_lock(&ctx->mutex);
1327 ret = rdma_set_ib_path(ctx->cm_id, &opa);
1328 mutex_unlock(&ctx->mutex);
1329 } else {
1330 mutex_lock(&ctx->mutex);
1331 ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
1332 mutex_unlock(&ctx->mutex);
1333 }
1334 if (ret)
1335 return ret;
1336
1337 memset(&event, 0, sizeof event);
1338 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1339 return ucma_event_handler(ctx->cm_id, &event);
1340}
1341
1342static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1343 void *optval, size_t optlen)
1344{
1345 int ret;
1346
1347 switch (optname) {
1348 case RDMA_OPTION_IB_PATH:
1349 ret = ucma_set_ib_path(ctx, optval, optlen);
1350 break;
1351 default:
1352 ret = -ENOSYS;
1353 }
1354
1355 return ret;
1356}
1357
1358static int ucma_set_option_level(struct ucma_context *ctx, int level,
1359 int optname, void *optval, size_t optlen)
1360{
1361 int ret;
1362
1363 switch (level) {
1364 case RDMA_OPTION_ID:
1365 mutex_lock(&ctx->mutex);
1366 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1367 mutex_unlock(&ctx->mutex);
1368 break;
1369 case RDMA_OPTION_IB:
1370 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1371 break;
1372 default:
1373 ret = -ENOSYS;
1374 }
1375
1376 return ret;
1377}
1378
1379static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1380 int in_len, int out_len)
1381{
1382 struct rdma_ucm_set_option cmd;
1383 struct ucma_context *ctx;
1384 void *optval;
1385 int ret;
1386
1387 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1388 return -EFAULT;
1389
1390 if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1391 return -EINVAL;
1392
1393 ctx = ucma_get_ctx(file, cmd.id);
1394 if (IS_ERR(ctx))
1395 return PTR_ERR(ctx);
1396
1397 optval = memdup_user(u64_to_user_ptr(cmd.optval),
1398 cmd.optlen);
1399 if (IS_ERR(optval)) {
1400 ret = PTR_ERR(optval);
1401 goto out;
1402 }
1403
1404 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1405 cmd.optlen);
1406 kfree(optval);
1407
1408out:
1409 ucma_put_ctx(ctx);
1410 return ret;
1411}
1412
1413static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1414 int in_len, int out_len)
1415{
1416 struct rdma_ucm_notify cmd;
1417 struct ucma_context *ctx;
1418 int ret = -EINVAL;
1419
1420 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1421 return -EFAULT;
1422
1423 ctx = ucma_get_ctx(file, cmd.id);
1424 if (IS_ERR(ctx))
1425 return PTR_ERR(ctx);
1426
1427 mutex_lock(&ctx->mutex);
1428 if (ctx->cm_id->device)
1429 ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
1430 mutex_unlock(&ctx->mutex);
1431
1432 ucma_put_ctx(ctx);
1433 return ret;
1434}
1435
1436static ssize_t ucma_process_join(struct ucma_file *file,
1437 struct rdma_ucm_join_mcast *cmd, int out_len)
1438{
1439 struct rdma_ucm_create_id_resp resp;
1440 struct ucma_context *ctx;
1441 struct ucma_multicast *mc;
1442 struct sockaddr *addr;
1443 int ret;
1444 u8 join_state;
1445
1446 if (out_len < sizeof(resp))
1447 return -ENOSPC;
1448
1449 addr = (struct sockaddr *) &cmd->addr;
1450 if (cmd->addr_size != rdma_addr_size(addr))
1451 return -EINVAL;
1452
1453 if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1454 join_state = BIT(FULLMEMBER_JOIN);
1455 else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1456 join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1457 else
1458 return -EINVAL;
1459
1460 ctx = ucma_get_ctx_dev(file, cmd->id);
1461 if (IS_ERR(ctx))
1462 return PTR_ERR(ctx);
1463
1464 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
1465 if (!mc) {
1466 ret = -ENOMEM;
1467 goto err_put_ctx;
1468 }
1469
1470 mc->ctx = ctx;
1471 mc->join_state = join_state;
1472 mc->uid = cmd->uid;
1473 memcpy(&mc->addr, addr, cmd->addr_size);
1474
1475 xa_lock(&multicast_table);
1476 if (__xa_alloc(&multicast_table, &mc->id, NULL, xa_limit_32b,
1477 GFP_KERNEL)) {
1478 ret = -ENOMEM;
1479 goto err_free_mc;
1480 }
1481
1482 list_add_tail(&mc->list, &ctx->mc_list);
1483 xa_unlock(&multicast_table);
1484
1485 mutex_lock(&ctx->mutex);
1486 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1487 join_state, mc);
1488 mutex_unlock(&ctx->mutex);
1489 if (ret)
1490 goto err_xa_erase;
1491
1492 resp.id = mc->id;
1493 if (copy_to_user(u64_to_user_ptr(cmd->response),
1494 &resp, sizeof(resp))) {
1495 ret = -EFAULT;
1496 goto err_leave_multicast;
1497 }
1498
1499 xa_store(&multicast_table, mc->id, mc, 0);
1500
1501 ucma_put_ctx(ctx);
1502 return 0;
1503
1504err_leave_multicast:
1505 mutex_lock(&ctx->mutex);
1506 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1507 mutex_unlock(&ctx->mutex);
1508 ucma_cleanup_mc_events(mc);
1509err_xa_erase:
1510 xa_lock(&multicast_table);
1511 list_del(&mc->list);
1512 __xa_erase(&multicast_table, mc->id);
1513err_free_mc:
1514 xa_unlock(&multicast_table);
1515 kfree(mc);
1516err_put_ctx:
1517 ucma_put_ctx(ctx);
1518 return ret;
1519}
1520
1521static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1522 const char __user *inbuf,
1523 int in_len, int out_len)
1524{
1525 struct rdma_ucm_join_ip_mcast cmd;
1526 struct rdma_ucm_join_mcast join_cmd;
1527
1528 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1529 return -EFAULT;
1530
1531 join_cmd.response = cmd.response;
1532 join_cmd.uid = cmd.uid;
1533 join_cmd.id = cmd.id;
1534 join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1535 if (!join_cmd.addr_size)
1536 return -EINVAL;
1537
1538 join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1539 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1540
1541 return ucma_process_join(file, &join_cmd, out_len);
1542}
1543
1544static ssize_t ucma_join_multicast(struct ucma_file *file,
1545 const char __user *inbuf,
1546 int in_len, int out_len)
1547{
1548 struct rdma_ucm_join_mcast cmd;
1549
1550 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1551 return -EFAULT;
1552
1553 if (!rdma_addr_size_kss(&cmd.addr))
1554 return -EINVAL;
1555
1556 return ucma_process_join(file, &cmd, out_len);
1557}
1558
1559static ssize_t ucma_leave_multicast(struct ucma_file *file,
1560 const char __user *inbuf,
1561 int in_len, int out_len)
1562{
1563 struct rdma_ucm_destroy_id cmd;
1564 struct rdma_ucm_destroy_id_resp resp;
1565 struct ucma_multicast *mc;
1566 int ret = 0;
1567
1568 if (out_len < sizeof(resp))
1569 return -ENOSPC;
1570
1571 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1572 return -EFAULT;
1573
1574 xa_lock(&multicast_table);
1575 mc = xa_load(&multicast_table, cmd.id);
1576 if (!mc)
1577 mc = ERR_PTR(-ENOENT);
1578 else if (READ_ONCE(mc->ctx->file) != file)
1579 mc = ERR_PTR(-EINVAL);
1580 else if (!refcount_inc_not_zero(&mc->ctx->ref))
1581 mc = ERR_PTR(-ENXIO);
1582
1583 if (IS_ERR(mc)) {
1584 xa_unlock(&multicast_table);
1585 ret = PTR_ERR(mc);
1586 goto out;
1587 }
1588
1589 list_del(&mc->list);
1590 __xa_erase(&multicast_table, mc->id);
1591 xa_unlock(&multicast_table);
1592
1593 mutex_lock(&mc->ctx->mutex);
1594 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1595 mutex_unlock(&mc->ctx->mutex);
1596
1597 ucma_cleanup_mc_events(mc);
1598
1599 ucma_put_ctx(mc->ctx);
1600 resp.events_reported = mc->events_reported;
1601 kfree(mc);
1602
1603 if (copy_to_user(u64_to_user_ptr(cmd.response),
1604 &resp, sizeof(resp)))
1605 ret = -EFAULT;
1606out:
1607 return ret;
1608}
1609
1610static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1611 const char __user *inbuf,
1612 int in_len, int out_len)
1613{
1614 struct rdma_ucm_migrate_id cmd;
1615 struct rdma_ucm_migrate_resp resp;
1616 struct ucma_event *uevent, *tmp;
1617 struct ucma_context *ctx;
1618 LIST_HEAD(event_list);
1619 struct fd f;
1620 struct ucma_file *cur_file;
1621 int ret = 0;
1622
1623 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1624 return -EFAULT;
1625
1626 /* Get current fd to protect against it being closed */
1627 f = fdget(cmd.fd);
1628 if (!f.file)
1629 return -ENOENT;
1630 if (f.file->f_op != &ucma_fops) {
1631 ret = -EINVAL;
1632 goto file_put;
1633 }
1634 cur_file = f.file->private_data;
1635
1636 /* Validate current fd and prevent destruction of id. */
1637 ctx = ucma_get_ctx(cur_file, cmd.id);
1638 if (IS_ERR(ctx)) {
1639 ret = PTR_ERR(ctx);
1640 goto file_put;
1641 }
1642
1643 rdma_lock_handler(ctx->cm_id);
1644 /*
1645 * ctx->file can only be changed under the handler & xa_lock. xa_load()
1646 * must be checked again to ensure the ctx hasn't begun destruction
1647 * since the ucma_get_ctx().
1648 */
1649 xa_lock(&ctx_table);
1650 if (_ucma_find_context(cmd.id, cur_file) != ctx) {
1651 xa_unlock(&ctx_table);
1652 ret = -ENOENT;
1653 goto err_unlock;
1654 }
1655 ctx->file = new_file;
1656 xa_unlock(&ctx_table);
1657
1658 mutex_lock(&cur_file->mut);
1659 list_del(&ctx->list);
1660 /*
1661 * At this point lock_handler() prevents addition of new uevents for
1662 * this ctx.
1663 */
1664 list_for_each_entry_safe(uevent, tmp, &cur_file->event_list, list)
1665 if (uevent->ctx == ctx)
1666 list_move_tail(&uevent->list, &event_list);
1667 resp.events_reported = ctx->events_reported;
1668 mutex_unlock(&cur_file->mut);
1669
1670 mutex_lock(&new_file->mut);
1671 list_add_tail(&ctx->list, &new_file->ctx_list);
1672 list_splice_tail(&event_list, &new_file->event_list);
1673 mutex_unlock(&new_file->mut);
1674
1675 if (copy_to_user(u64_to_user_ptr(cmd.response),
1676 &resp, sizeof(resp)))
1677 ret = -EFAULT;
1678
1679err_unlock:
1680 rdma_unlock_handler(ctx->cm_id);
1681 ucma_put_ctx(ctx);
1682file_put:
1683 fdput(f);
1684 return ret;
1685}
1686
1687static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1688 const char __user *inbuf,
1689 int in_len, int out_len) = {
1690 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1691 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1692 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1693 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1694 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1695 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1696 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1697 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1698 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1699 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1700 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1701 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1702 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1703 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1704 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1705 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1706 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1707 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1708 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1709 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1710 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1711 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1712 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1713};
1714
1715static ssize_t ucma_write(struct file *filp, const char __user *buf,
1716 size_t len, loff_t *pos)
1717{
1718 struct ucma_file *file = filp->private_data;
1719 struct rdma_ucm_cmd_hdr hdr;
1720 ssize_t ret;
1721
1722 if (!ib_safe_file_access(filp)) {
1723 pr_err_once("%s: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1724 __func__, task_tgid_vnr(current), current->comm);
1725 return -EACCES;
1726 }
1727
1728 if (len < sizeof(hdr))
1729 return -EINVAL;
1730
1731 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1732 return -EFAULT;
1733
1734 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1735 return -EINVAL;
1736 hdr.cmd = array_index_nospec(hdr.cmd, ARRAY_SIZE(ucma_cmd_table));
1737
1738 if (hdr.in + sizeof(hdr) > len)
1739 return -EINVAL;
1740
1741 if (!ucma_cmd_table[hdr.cmd])
1742 return -ENOSYS;
1743
1744 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1745 if (!ret)
1746 ret = len;
1747
1748 return ret;
1749}
1750
1751static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1752{
1753 struct ucma_file *file = filp->private_data;
1754 __poll_t mask = 0;
1755
1756 poll_wait(filp, &file->poll_wait, wait);
1757
1758 if (!list_empty(&file->event_list))
1759 mask = EPOLLIN | EPOLLRDNORM;
1760
1761 return mask;
1762}
1763
1764/*
1765 * ucma_open() does not need the BKL:
1766 *
1767 * - no global state is referred to;
1768 * - there is no ioctl method to race against;
1769 * - no further module initialization is required for open to work
1770 * after the device is registered.
1771 */
1772static int ucma_open(struct inode *inode, struct file *filp)
1773{
1774 struct ucma_file *file;
1775
1776 file = kmalloc(sizeof *file, GFP_KERNEL);
1777 if (!file)
1778 return -ENOMEM;
1779
1780 INIT_LIST_HEAD(&file->event_list);
1781 INIT_LIST_HEAD(&file->ctx_list);
1782 init_waitqueue_head(&file->poll_wait);
1783 mutex_init(&file->mut);
1784
1785 filp->private_data = file;
1786 file->filp = filp;
1787
1788 return stream_open(inode, filp);
1789}
1790
1791static int ucma_close(struct inode *inode, struct file *filp)
1792{
1793 struct ucma_file *file = filp->private_data;
1794
1795 /*
1796 * All paths that touch ctx_list or ctx_list starting from write() are
1797 * prevented by this being a FD release function. The list_add_tail() in
1798 * ucma_connect_event_handler() can run concurrently, however it only
1799 * adds to the list *after* a listening ID. By only reading the first of
1800 * the list, and relying on ucma_destroy_private_ctx() to block
1801 * ucma_connect_event_handler(), no additional locking is needed.
1802 */
1803 while (!list_empty(&file->ctx_list)) {
1804 struct ucma_context *ctx = list_first_entry(
1805 &file->ctx_list, struct ucma_context, list);
1806
1807 WARN_ON(xa_cmpxchg(&ctx_table, ctx->id, ctx, XA_ZERO_ENTRY,
1808 GFP_KERNEL) != ctx);
1809 ucma_destroy_private_ctx(ctx);
1810 }
1811 kfree(file);
1812 return 0;
1813}
1814
1815static const struct file_operations ucma_fops = {
1816 .owner = THIS_MODULE,
1817 .open = ucma_open,
1818 .release = ucma_close,
1819 .write = ucma_write,
1820 .poll = ucma_poll,
1821 .llseek = no_llseek,
1822};
1823
1824static struct miscdevice ucma_misc = {
1825 .minor = MISC_DYNAMIC_MINOR,
1826 .name = "rdma_cm",
1827 .nodename = "infiniband/rdma_cm",
1828 .mode = 0666,
1829 .fops = &ucma_fops,
1830};
1831
1832static int ucma_get_global_nl_info(struct ib_client_nl_info *res)
1833{
1834 res->abi = RDMA_USER_CM_ABI_VERSION;
1835 res->cdev = ucma_misc.this_device;
1836 return 0;
1837}
1838
1839static struct ib_client rdma_cma_client = {
1840 .name = "rdma_cm",
1841 .get_global_nl_info = ucma_get_global_nl_info,
1842};
1843MODULE_ALIAS_RDMA_CLIENT("rdma_cm");
1844
1845static ssize_t abi_version_show(struct device *dev,
1846 struct device_attribute *attr, char *buf)
1847{
1848 return sysfs_emit(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1849}
1850static DEVICE_ATTR_RO(abi_version);
1851
1852static int __init ucma_init(void)
1853{
1854 int ret;
1855
1856 ret = misc_register(&ucma_misc);
1857 if (ret)
1858 return ret;
1859
1860 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1861 if (ret) {
1862 pr_err("rdma_ucm: couldn't create abi_version attr\n");
1863 goto err1;
1864 }
1865
1866 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1867 if (!ucma_ctl_table_hdr) {
1868 pr_err("rdma_ucm: couldn't register sysctl paths\n");
1869 ret = -ENOMEM;
1870 goto err2;
1871 }
1872
1873 ret = ib_register_client(&rdma_cma_client);
1874 if (ret)
1875 goto err3;
1876
1877 return 0;
1878err3:
1879 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1880err2:
1881 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1882err1:
1883 misc_deregister(&ucma_misc);
1884 return ret;
1885}
1886
1887static void __exit ucma_cleanup(void)
1888{
1889 ib_unregister_client(&rdma_cma_client);
1890 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1891 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1892 misc_deregister(&ucma_misc);
1893}
1894
1895module_init(ucma_init);
1896module_exit(ucma_cleanup);