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 struct workqueue_struct *close_wq;
84};
85
86struct ucma_context {
87 u32 id;
88 struct completion comp;
89 refcount_t ref;
90 int events_reported;
91 int backlog;
92
93 struct ucma_file *file;
94 struct rdma_cm_id *cm_id;
95 struct mutex mutex;
96 u64 uid;
97
98 struct list_head list;
99 struct list_head mc_list;
100 /* mark that device is in process of destroying the internal HW
101 * resources, protected by the ctx_table lock
102 */
103 int closing;
104 /* sync between removal event and id destroy, protected by file mut */
105 int destroying;
106 struct work_struct close_work;
107};
108
109struct ucma_multicast {
110 struct ucma_context *ctx;
111 u32 id;
112 int events_reported;
113
114 u64 uid;
115 u8 join_state;
116 struct list_head list;
117 struct sockaddr_storage addr;
118};
119
120struct ucma_event {
121 struct ucma_context *ctx;
122 struct ucma_multicast *mc;
123 struct list_head list;
124 struct rdma_cm_id *cm_id;
125 struct rdma_ucm_event_resp resp;
126 struct work_struct close_work;
127};
128
129static DEFINE_XARRAY_ALLOC(ctx_table);
130static DEFINE_XARRAY_ALLOC(multicast_table);
131
132static const struct file_operations ucma_fops;
133
134static inline struct ucma_context *_ucma_find_context(int id,
135 struct ucma_file *file)
136{
137 struct ucma_context *ctx;
138
139 ctx = xa_load(&ctx_table, id);
140 if (!ctx)
141 ctx = ERR_PTR(-ENOENT);
142 else if (ctx->file != file || !ctx->cm_id)
143 ctx = ERR_PTR(-EINVAL);
144 return ctx;
145}
146
147static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
148{
149 struct ucma_context *ctx;
150
151 xa_lock(&ctx_table);
152 ctx = _ucma_find_context(id, file);
153 if (!IS_ERR(ctx)) {
154 if (ctx->closing)
155 ctx = ERR_PTR(-EIO);
156 else
157 refcount_inc(&ctx->ref);
158 }
159 xa_unlock(&ctx_table);
160 return ctx;
161}
162
163static void ucma_put_ctx(struct ucma_context *ctx)
164{
165 if (refcount_dec_and_test(&ctx->ref))
166 complete(&ctx->comp);
167}
168
169/*
170 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
171 * CM_ID is bound.
172 */
173static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
174{
175 struct ucma_context *ctx = ucma_get_ctx(file, id);
176
177 if (IS_ERR(ctx))
178 return ctx;
179 if (!ctx->cm_id->device) {
180 ucma_put_ctx(ctx);
181 return ERR_PTR(-EINVAL);
182 }
183 return ctx;
184}
185
186static void ucma_close_event_id(struct work_struct *work)
187{
188 struct ucma_event *uevent_close = container_of(work, struct ucma_event, close_work);
189
190 rdma_destroy_id(uevent_close->cm_id);
191 kfree(uevent_close);
192}
193
194static void ucma_close_id(struct work_struct *work)
195{
196 struct ucma_context *ctx = container_of(work, struct ucma_context, close_work);
197
198 /* once all inflight tasks are finished, we close all underlying
199 * resources. The context is still alive till its explicit destryoing
200 * by its creator.
201 */
202 ucma_put_ctx(ctx);
203 wait_for_completion(&ctx->comp);
204 /* No new events will be generated after destroying the id. */
205 rdma_destroy_id(ctx->cm_id);
206}
207
208static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
209{
210 struct ucma_context *ctx;
211
212 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
213 if (!ctx)
214 return NULL;
215
216 INIT_WORK(&ctx->close_work, ucma_close_id);
217 refcount_set(&ctx->ref, 1);
218 init_completion(&ctx->comp);
219 INIT_LIST_HEAD(&ctx->mc_list);
220 ctx->file = file;
221 mutex_init(&ctx->mutex);
222
223 if (xa_alloc(&ctx_table, &ctx->id, ctx, xa_limit_32b, GFP_KERNEL))
224 goto error;
225
226 list_add_tail(&ctx->list, &file->ctx_list);
227 return ctx;
228
229error:
230 kfree(ctx);
231 return NULL;
232}
233
234static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
235{
236 struct ucma_multicast *mc;
237
238 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
239 if (!mc)
240 return NULL;
241
242 mc->ctx = ctx;
243 if (xa_alloc(&multicast_table, &mc->id, NULL, xa_limit_32b, GFP_KERNEL))
244 goto error;
245
246 list_add_tail(&mc->list, &ctx->mc_list);
247 return mc;
248
249error:
250 kfree(mc);
251 return NULL;
252}
253
254static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
255 struct rdma_conn_param *src)
256{
257 if (src->private_data_len)
258 memcpy(dst->private_data, src->private_data,
259 src->private_data_len);
260 dst->private_data_len = src->private_data_len;
261 dst->responder_resources =src->responder_resources;
262 dst->initiator_depth = src->initiator_depth;
263 dst->flow_control = src->flow_control;
264 dst->retry_count = src->retry_count;
265 dst->rnr_retry_count = src->rnr_retry_count;
266 dst->srq = src->srq;
267 dst->qp_num = src->qp_num;
268}
269
270static void ucma_copy_ud_event(struct ib_device *device,
271 struct rdma_ucm_ud_param *dst,
272 struct rdma_ud_param *src)
273{
274 if (src->private_data_len)
275 memcpy(dst->private_data, src->private_data,
276 src->private_data_len);
277 dst->private_data_len = src->private_data_len;
278 ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
279 dst->qp_num = src->qp_num;
280 dst->qkey = src->qkey;
281}
282
283static void ucma_set_event_context(struct ucma_context *ctx,
284 struct rdma_cm_event *event,
285 struct ucma_event *uevent)
286{
287 uevent->ctx = ctx;
288 switch (event->event) {
289 case RDMA_CM_EVENT_MULTICAST_JOIN:
290 case RDMA_CM_EVENT_MULTICAST_ERROR:
291 uevent->mc = (struct ucma_multicast *)
292 event->param.ud.private_data;
293 uevent->resp.uid = uevent->mc->uid;
294 uevent->resp.id = uevent->mc->id;
295 break;
296 default:
297 uevent->resp.uid = ctx->uid;
298 uevent->resp.id = ctx->id;
299 break;
300 }
301}
302
303/* Called with file->mut locked for the relevant context. */
304static void ucma_removal_event_handler(struct rdma_cm_id *cm_id)
305{
306 struct ucma_context *ctx = cm_id->context;
307 struct ucma_event *con_req_eve;
308 int event_found = 0;
309
310 if (ctx->destroying)
311 return;
312
313 /* only if context is pointing to cm_id that it owns it and can be
314 * queued to be closed, otherwise that cm_id is an inflight one that
315 * is part of that context event list pending to be detached and
316 * reattached to its new context as part of ucma_get_event,
317 * handled separately below.
318 */
319 if (ctx->cm_id == cm_id) {
320 xa_lock(&ctx_table);
321 ctx->closing = 1;
322 xa_unlock(&ctx_table);
323 queue_work(ctx->file->close_wq, &ctx->close_work);
324 return;
325 }
326
327 list_for_each_entry(con_req_eve, &ctx->file->event_list, list) {
328 if (con_req_eve->cm_id == cm_id &&
329 con_req_eve->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
330 list_del(&con_req_eve->list);
331 INIT_WORK(&con_req_eve->close_work, ucma_close_event_id);
332 queue_work(ctx->file->close_wq, &con_req_eve->close_work);
333 event_found = 1;
334 break;
335 }
336 }
337 if (!event_found)
338 pr_err("ucma_removal_event_handler: warning: connect request event wasn't found\n");
339}
340
341static int ucma_event_handler(struct rdma_cm_id *cm_id,
342 struct rdma_cm_event *event)
343{
344 struct ucma_event *uevent;
345 struct ucma_context *ctx = cm_id->context;
346 int ret = 0;
347
348 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
349 if (!uevent)
350 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
351
352 mutex_lock(&ctx->file->mut);
353 uevent->cm_id = cm_id;
354 ucma_set_event_context(ctx, event, uevent);
355 uevent->resp.event = event->event;
356 uevent->resp.status = event->status;
357 if (cm_id->qp_type == IB_QPT_UD)
358 ucma_copy_ud_event(cm_id->device, &uevent->resp.param.ud,
359 &event->param.ud);
360 else
361 ucma_copy_conn_event(&uevent->resp.param.conn,
362 &event->param.conn);
363
364 uevent->resp.ece.vendor_id = event->ece.vendor_id;
365 uevent->resp.ece.attr_mod = event->ece.attr_mod;
366
367 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
368 if (!ctx->backlog) {
369 ret = -ENOMEM;
370 kfree(uevent);
371 goto out;
372 }
373 ctx->backlog--;
374 } else if (!ctx->uid || ctx->cm_id != cm_id) {
375 /*
376 * We ignore events for new connections until userspace has set
377 * their context. This can only happen if an error occurs on a
378 * new connection before the user accepts it. This is okay,
379 * since the accept will just fail later. However, we do need
380 * to release the underlying HW resources in case of a device
381 * removal event.
382 */
383 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
384 ucma_removal_event_handler(cm_id);
385
386 kfree(uevent);
387 goto out;
388 }
389
390 list_add_tail(&uevent->list, &ctx->file->event_list);
391 wake_up_interruptible(&ctx->file->poll_wait);
392 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
393 ucma_removal_event_handler(cm_id);
394out:
395 mutex_unlock(&ctx->file->mut);
396 return ret;
397}
398
399static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
400 int in_len, int out_len)
401{
402 struct ucma_context *ctx;
403 struct rdma_ucm_get_event cmd;
404 struct ucma_event *uevent;
405 int ret = 0;
406
407 /*
408 * Old 32 bit user space does not send the 4 byte padding in the
409 * reserved field. We don't care, allow it to keep working.
410 */
411 if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved) -
412 sizeof(uevent->resp.ece))
413 return -ENOSPC;
414
415 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
416 return -EFAULT;
417
418 mutex_lock(&file->mut);
419 while (list_empty(&file->event_list)) {
420 mutex_unlock(&file->mut);
421
422 if (file->filp->f_flags & O_NONBLOCK)
423 return -EAGAIN;
424
425 if (wait_event_interruptible(file->poll_wait,
426 !list_empty(&file->event_list)))
427 return -ERESTARTSYS;
428
429 mutex_lock(&file->mut);
430 }
431
432 uevent = list_entry(file->event_list.next, struct ucma_event, list);
433
434 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
435 ctx = ucma_alloc_ctx(file);
436 if (!ctx) {
437 ret = -ENOMEM;
438 goto done;
439 }
440 uevent->ctx->backlog++;
441 ctx->cm_id = uevent->cm_id;
442 ctx->cm_id->context = ctx;
443 uevent->resp.id = ctx->id;
444 }
445
446 if (copy_to_user(u64_to_user_ptr(cmd.response),
447 &uevent->resp,
448 min_t(size_t, out_len, sizeof(uevent->resp)))) {
449 ret = -EFAULT;
450 goto done;
451 }
452
453 list_del(&uevent->list);
454 uevent->ctx->events_reported++;
455 if (uevent->mc)
456 uevent->mc->events_reported++;
457 kfree(uevent);
458done:
459 mutex_unlock(&file->mut);
460 return ret;
461}
462
463static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
464{
465 switch (cmd->ps) {
466 case RDMA_PS_TCP:
467 *qp_type = IB_QPT_RC;
468 return 0;
469 case RDMA_PS_UDP:
470 case RDMA_PS_IPOIB:
471 *qp_type = IB_QPT_UD;
472 return 0;
473 case RDMA_PS_IB:
474 *qp_type = cmd->qp_type;
475 return 0;
476 default:
477 return -EINVAL;
478 }
479}
480
481static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
482 int in_len, int out_len)
483{
484 struct rdma_ucm_create_id cmd;
485 struct rdma_ucm_create_id_resp resp;
486 struct ucma_context *ctx;
487 struct rdma_cm_id *cm_id;
488 enum ib_qp_type qp_type;
489 int ret;
490
491 if (out_len < sizeof(resp))
492 return -ENOSPC;
493
494 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
495 return -EFAULT;
496
497 ret = ucma_get_qp_type(&cmd, &qp_type);
498 if (ret)
499 return ret;
500
501 mutex_lock(&file->mut);
502 ctx = ucma_alloc_ctx(file);
503 mutex_unlock(&file->mut);
504 if (!ctx)
505 return -ENOMEM;
506
507 ctx->uid = cmd.uid;
508 cm_id = __rdma_create_id(current->nsproxy->net_ns,
509 ucma_event_handler, ctx, cmd.ps, qp_type, NULL);
510 if (IS_ERR(cm_id)) {
511 ret = PTR_ERR(cm_id);
512 goto err1;
513 }
514
515 resp.id = ctx->id;
516 if (copy_to_user(u64_to_user_ptr(cmd.response),
517 &resp, sizeof(resp))) {
518 ret = -EFAULT;
519 goto err2;
520 }
521
522 ctx->cm_id = cm_id;
523 return 0;
524
525err2:
526 rdma_destroy_id(cm_id);
527err1:
528 xa_erase(&ctx_table, ctx->id);
529 mutex_lock(&file->mut);
530 list_del(&ctx->list);
531 mutex_unlock(&file->mut);
532 kfree(ctx);
533 return ret;
534}
535
536static void ucma_cleanup_multicast(struct ucma_context *ctx)
537{
538 struct ucma_multicast *mc, *tmp;
539
540 mutex_lock(&ctx->file->mut);
541 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
542 list_del(&mc->list);
543 xa_erase(&multicast_table, mc->id);
544 kfree(mc);
545 }
546 mutex_unlock(&ctx->file->mut);
547}
548
549static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
550{
551 struct ucma_event *uevent, *tmp;
552
553 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
554 if (uevent->mc != mc)
555 continue;
556
557 list_del(&uevent->list);
558 kfree(uevent);
559 }
560}
561
562/*
563 * ucma_free_ctx is called after the underlying rdma CM-ID is destroyed. At
564 * this point, no new events will be reported from the hardware. However, we
565 * still need to cleanup the UCMA context for this ID. Specifically, there
566 * might be events that have not yet been consumed by the user space software.
567 * These might include pending connect requests which we have not completed
568 * processing. We cannot call rdma_destroy_id while holding the lock of the
569 * context (file->mut), as it might cause a deadlock. We therefore extract all
570 * relevant events from the context pending events list while holding the
571 * mutex. After that we release them as needed.
572 */
573static int ucma_free_ctx(struct ucma_context *ctx)
574{
575 int events_reported;
576 struct ucma_event *uevent, *tmp;
577 LIST_HEAD(list);
578
579
580 ucma_cleanup_multicast(ctx);
581
582 /* Cleanup events not yet reported to the user. */
583 mutex_lock(&ctx->file->mut);
584 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
585 if (uevent->ctx == ctx)
586 list_move_tail(&uevent->list, &list);
587 }
588 list_del(&ctx->list);
589 mutex_unlock(&ctx->file->mut);
590
591 list_for_each_entry_safe(uevent, tmp, &list, list) {
592 list_del(&uevent->list);
593 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
594 rdma_destroy_id(uevent->cm_id);
595 kfree(uevent);
596 }
597
598 events_reported = ctx->events_reported;
599 mutex_destroy(&ctx->mutex);
600 kfree(ctx);
601 return events_reported;
602}
603
604static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
605 int in_len, int out_len)
606{
607 struct rdma_ucm_destroy_id cmd;
608 struct rdma_ucm_destroy_id_resp resp;
609 struct ucma_context *ctx;
610 int ret = 0;
611
612 if (out_len < sizeof(resp))
613 return -ENOSPC;
614
615 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
616 return -EFAULT;
617
618 xa_lock(&ctx_table);
619 ctx = _ucma_find_context(cmd.id, file);
620 if (!IS_ERR(ctx))
621 __xa_erase(&ctx_table, ctx->id);
622 xa_unlock(&ctx_table);
623
624 if (IS_ERR(ctx))
625 return PTR_ERR(ctx);
626
627 mutex_lock(&ctx->file->mut);
628 ctx->destroying = 1;
629 mutex_unlock(&ctx->file->mut);
630
631 flush_workqueue(ctx->file->close_wq);
632 /* At this point it's guaranteed that there is no inflight
633 * closing task */
634 xa_lock(&ctx_table);
635 if (!ctx->closing) {
636 xa_unlock(&ctx_table);
637 ucma_put_ctx(ctx);
638 wait_for_completion(&ctx->comp);
639 rdma_destroy_id(ctx->cm_id);
640 } else {
641 xa_unlock(&ctx_table);
642 }
643
644 resp.events_reported = ucma_free_ctx(ctx);
645 if (copy_to_user(u64_to_user_ptr(cmd.response),
646 &resp, sizeof(resp)))
647 ret = -EFAULT;
648
649 return ret;
650}
651
652static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
653 int in_len, int out_len)
654{
655 struct rdma_ucm_bind_ip cmd;
656 struct ucma_context *ctx;
657 int ret;
658
659 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
660 return -EFAULT;
661
662 if (!rdma_addr_size_in6(&cmd.addr))
663 return -EINVAL;
664
665 ctx = ucma_get_ctx(file, cmd.id);
666 if (IS_ERR(ctx))
667 return PTR_ERR(ctx);
668
669 mutex_lock(&ctx->mutex);
670 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
671 mutex_unlock(&ctx->mutex);
672
673 ucma_put_ctx(ctx);
674 return ret;
675}
676
677static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
678 int in_len, int out_len)
679{
680 struct rdma_ucm_bind cmd;
681 struct ucma_context *ctx;
682 int ret;
683
684 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
685 return -EFAULT;
686
687 if (cmd.reserved || !cmd.addr_size ||
688 cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
689 return -EINVAL;
690
691 ctx = ucma_get_ctx(file, cmd.id);
692 if (IS_ERR(ctx))
693 return PTR_ERR(ctx);
694
695 mutex_lock(&ctx->mutex);
696 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
697 mutex_unlock(&ctx->mutex);
698 ucma_put_ctx(ctx);
699 return ret;
700}
701
702static ssize_t ucma_resolve_ip(struct ucma_file *file,
703 const char __user *inbuf,
704 int in_len, int out_len)
705{
706 struct rdma_ucm_resolve_ip 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.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
714 !rdma_addr_size_in6(&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_addr(struct ucma_file *file,
730 const char __user *inbuf,
731 int in_len, int out_len)
732{
733 struct rdma_ucm_resolve_addr cmd;
734 struct ucma_context *ctx;
735 int ret;
736
737 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
738 return -EFAULT;
739
740 if (cmd.reserved ||
741 (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
742 !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
743 return -EINVAL;
744
745 ctx = ucma_get_ctx(file, cmd.id);
746 if (IS_ERR(ctx))
747 return PTR_ERR(ctx);
748
749 mutex_lock(&ctx->mutex);
750 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
751 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
752 mutex_unlock(&ctx->mutex);
753 ucma_put_ctx(ctx);
754 return ret;
755}
756
757static ssize_t ucma_resolve_route(struct ucma_file *file,
758 const char __user *inbuf,
759 int in_len, int out_len)
760{
761 struct rdma_ucm_resolve_route cmd;
762 struct ucma_context *ctx;
763 int ret;
764
765 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
766 return -EFAULT;
767
768 ctx = ucma_get_ctx_dev(file, cmd.id);
769 if (IS_ERR(ctx))
770 return PTR_ERR(ctx);
771
772 mutex_lock(&ctx->mutex);
773 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
774 mutex_unlock(&ctx->mutex);
775 ucma_put_ctx(ctx);
776 return ret;
777}
778
779static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
780 struct rdma_route *route)
781{
782 struct rdma_dev_addr *dev_addr;
783
784 resp->num_paths = route->num_paths;
785 switch (route->num_paths) {
786 case 0:
787 dev_addr = &route->addr.dev_addr;
788 rdma_addr_get_dgid(dev_addr,
789 (union ib_gid *) &resp->ib_route[0].dgid);
790 rdma_addr_get_sgid(dev_addr,
791 (union ib_gid *) &resp->ib_route[0].sgid);
792 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
793 break;
794 case 2:
795 ib_copy_path_rec_to_user(&resp->ib_route[1],
796 &route->path_rec[1]);
797 fallthrough;
798 case 1:
799 ib_copy_path_rec_to_user(&resp->ib_route[0],
800 &route->path_rec[0]);
801 break;
802 default:
803 break;
804 }
805}
806
807static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
808 struct rdma_route *route)
809{
810
811 resp->num_paths = route->num_paths;
812 switch (route->num_paths) {
813 case 0:
814 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
815 (union ib_gid *)&resp->ib_route[0].dgid);
816 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
817 (union ib_gid *)&resp->ib_route[0].sgid);
818 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
819 break;
820 case 2:
821 ib_copy_path_rec_to_user(&resp->ib_route[1],
822 &route->path_rec[1]);
823 fallthrough;
824 case 1:
825 ib_copy_path_rec_to_user(&resp->ib_route[0],
826 &route->path_rec[0]);
827 break;
828 default:
829 break;
830 }
831}
832
833static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
834 struct rdma_route *route)
835{
836 struct rdma_dev_addr *dev_addr;
837
838 dev_addr = &route->addr.dev_addr;
839 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
840 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
841}
842
843static ssize_t ucma_query_route(struct ucma_file *file,
844 const char __user *inbuf,
845 int in_len, int out_len)
846{
847 struct rdma_ucm_query cmd;
848 struct rdma_ucm_query_route_resp resp;
849 struct ucma_context *ctx;
850 struct sockaddr *addr;
851 int ret = 0;
852
853 if (out_len < offsetof(struct rdma_ucm_query_route_resp, ibdev_index))
854 return -ENOSPC;
855
856 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
857 return -EFAULT;
858
859 ctx = ucma_get_ctx(file, cmd.id);
860 if (IS_ERR(ctx))
861 return PTR_ERR(ctx);
862
863 mutex_lock(&ctx->mutex);
864 memset(&resp, 0, sizeof resp);
865 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
866 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
867 sizeof(struct sockaddr_in) :
868 sizeof(struct sockaddr_in6));
869 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
870 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
871 sizeof(struct sockaddr_in) :
872 sizeof(struct sockaddr_in6));
873 if (!ctx->cm_id->device)
874 goto out;
875
876 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
877 resp.ibdev_index = ctx->cm_id->device->index;
878 resp.port_num = ctx->cm_id->port_num;
879
880 if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
881 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
882 else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
883 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
884 else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
885 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
886
887out:
888 mutex_unlock(&ctx->mutex);
889 if (copy_to_user(u64_to_user_ptr(cmd.response), &resp,
890 min_t(size_t, out_len, sizeof(resp))))
891 ret = -EFAULT;
892
893 ucma_put_ctx(ctx);
894 return ret;
895}
896
897static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
898 struct rdma_ucm_query_addr_resp *resp)
899{
900 if (!cm_id->device)
901 return;
902
903 resp->node_guid = (__force __u64) cm_id->device->node_guid;
904 resp->ibdev_index = cm_id->device->index;
905 resp->port_num = cm_id->port_num;
906 resp->pkey = (__force __u16) cpu_to_be16(
907 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
908}
909
910static ssize_t ucma_query_addr(struct ucma_context *ctx,
911 void __user *response, int out_len)
912{
913 struct rdma_ucm_query_addr_resp resp;
914 struct sockaddr *addr;
915 int ret = 0;
916
917 if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
918 return -ENOSPC;
919
920 memset(&resp, 0, sizeof resp);
921
922 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
923 resp.src_size = rdma_addr_size(addr);
924 memcpy(&resp.src_addr, addr, resp.src_size);
925
926 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
927 resp.dst_size = rdma_addr_size(addr);
928 memcpy(&resp.dst_addr, addr, resp.dst_size);
929
930 ucma_query_device_addr(ctx->cm_id, &resp);
931
932 if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
933 ret = -EFAULT;
934
935 return ret;
936}
937
938static ssize_t ucma_query_path(struct ucma_context *ctx,
939 void __user *response, int out_len)
940{
941 struct rdma_ucm_query_path_resp *resp;
942 int i, ret = 0;
943
944 if (out_len < sizeof(*resp))
945 return -ENOSPC;
946
947 resp = kzalloc(out_len, GFP_KERNEL);
948 if (!resp)
949 return -ENOMEM;
950
951 resp->num_paths = ctx->cm_id->route.num_paths;
952 for (i = 0, out_len -= sizeof(*resp);
953 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
954 i++, out_len -= sizeof(struct ib_path_rec_data)) {
955 struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
956
957 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
958 IB_PATH_BIDIRECTIONAL;
959 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
960 struct sa_path_rec ib;
961
962 sa_convert_path_opa_to_ib(&ib, rec);
963 ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
964
965 } else {
966 ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
967 }
968 }
969
970 if (copy_to_user(response, resp, struct_size(resp, path_data, i)))
971 ret = -EFAULT;
972
973 kfree(resp);
974 return ret;
975}
976
977static ssize_t ucma_query_gid(struct ucma_context *ctx,
978 void __user *response, int out_len)
979{
980 struct rdma_ucm_query_addr_resp resp;
981 struct sockaddr_ib *addr;
982 int ret = 0;
983
984 if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
985 return -ENOSPC;
986
987 memset(&resp, 0, sizeof resp);
988
989 ucma_query_device_addr(ctx->cm_id, &resp);
990
991 addr = (struct sockaddr_ib *) &resp.src_addr;
992 resp.src_size = sizeof(*addr);
993 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
994 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
995 } else {
996 addr->sib_family = AF_IB;
997 addr->sib_pkey = (__force __be16) resp.pkey;
998 rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
999 NULL);
1000 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
1001 &ctx->cm_id->route.addr.src_addr);
1002 }
1003
1004 addr = (struct sockaddr_ib *) &resp.dst_addr;
1005 resp.dst_size = sizeof(*addr);
1006 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
1007 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
1008 } else {
1009 addr->sib_family = AF_IB;
1010 addr->sib_pkey = (__force __be16) resp.pkey;
1011 rdma_read_gids(ctx->cm_id, NULL,
1012 (union ib_gid *)&addr->sib_addr);
1013 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
1014 &ctx->cm_id->route.addr.dst_addr);
1015 }
1016
1017 if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
1018 ret = -EFAULT;
1019
1020 return ret;
1021}
1022
1023static ssize_t ucma_query(struct ucma_file *file,
1024 const char __user *inbuf,
1025 int in_len, int out_len)
1026{
1027 struct rdma_ucm_query cmd;
1028 struct ucma_context *ctx;
1029 void __user *response;
1030 int ret;
1031
1032 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1033 return -EFAULT;
1034
1035 response = u64_to_user_ptr(cmd.response);
1036 ctx = ucma_get_ctx(file, cmd.id);
1037 if (IS_ERR(ctx))
1038 return PTR_ERR(ctx);
1039
1040 mutex_lock(&ctx->mutex);
1041 switch (cmd.option) {
1042 case RDMA_USER_CM_QUERY_ADDR:
1043 ret = ucma_query_addr(ctx, response, out_len);
1044 break;
1045 case RDMA_USER_CM_QUERY_PATH:
1046 ret = ucma_query_path(ctx, response, out_len);
1047 break;
1048 case RDMA_USER_CM_QUERY_GID:
1049 ret = ucma_query_gid(ctx, response, out_len);
1050 break;
1051 default:
1052 ret = -ENOSYS;
1053 break;
1054 }
1055 mutex_unlock(&ctx->mutex);
1056
1057 ucma_put_ctx(ctx);
1058 return ret;
1059}
1060
1061static void ucma_copy_conn_param(struct rdma_cm_id *id,
1062 struct rdma_conn_param *dst,
1063 struct rdma_ucm_conn_param *src)
1064{
1065 dst->private_data = src->private_data;
1066 dst->private_data_len = src->private_data_len;
1067 dst->responder_resources =src->responder_resources;
1068 dst->initiator_depth = src->initiator_depth;
1069 dst->flow_control = src->flow_control;
1070 dst->retry_count = src->retry_count;
1071 dst->rnr_retry_count = src->rnr_retry_count;
1072 dst->srq = src->srq;
1073 dst->qp_num = src->qp_num & 0xFFFFFF;
1074 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1075}
1076
1077static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1078 int in_len, int out_len)
1079{
1080 struct rdma_conn_param conn_param;
1081 struct rdma_ucm_ece ece = {};
1082 struct rdma_ucm_connect cmd;
1083 struct ucma_context *ctx;
1084 size_t in_size;
1085 int ret;
1086
1087 if (in_len < offsetofend(typeof(cmd), reserved))
1088 return -EINVAL;
1089 in_size = min_t(size_t, in_len, sizeof(cmd));
1090 if (copy_from_user(&cmd, inbuf, in_size))
1091 return -EFAULT;
1092
1093 if (!cmd.conn_param.valid)
1094 return -EINVAL;
1095
1096 ctx = ucma_get_ctx_dev(file, cmd.id);
1097 if (IS_ERR(ctx))
1098 return PTR_ERR(ctx);
1099
1100 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1101 if (offsetofend(typeof(cmd), ece) <= in_size) {
1102 ece.vendor_id = cmd.ece.vendor_id;
1103 ece.attr_mod = cmd.ece.attr_mod;
1104 }
1105
1106 mutex_lock(&ctx->mutex);
1107 ret = rdma_connect_ece(ctx->cm_id, &conn_param, &ece);
1108 mutex_unlock(&ctx->mutex);
1109 ucma_put_ctx(ctx);
1110 return ret;
1111}
1112
1113static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1114 int in_len, int out_len)
1115{
1116 struct rdma_ucm_listen cmd;
1117 struct ucma_context *ctx;
1118 int ret;
1119
1120 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1121 return -EFAULT;
1122
1123 ctx = ucma_get_ctx(file, cmd.id);
1124 if (IS_ERR(ctx))
1125 return PTR_ERR(ctx);
1126
1127 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
1128 cmd.backlog : max_backlog;
1129 mutex_lock(&ctx->mutex);
1130 ret = rdma_listen(ctx->cm_id, ctx->backlog);
1131 mutex_unlock(&ctx->mutex);
1132 ucma_put_ctx(ctx);
1133 return ret;
1134}
1135
1136static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1137 int in_len, int out_len)
1138{
1139 struct rdma_ucm_accept cmd;
1140 struct rdma_conn_param conn_param;
1141 struct rdma_ucm_ece ece = {};
1142 struct ucma_context *ctx;
1143 size_t in_size;
1144 int ret;
1145
1146 if (in_len < offsetofend(typeof(cmd), reserved))
1147 return -EINVAL;
1148 in_size = min_t(size_t, in_len, sizeof(cmd));
1149 if (copy_from_user(&cmd, inbuf, in_size))
1150 return -EFAULT;
1151
1152 ctx = ucma_get_ctx_dev(file, cmd.id);
1153 if (IS_ERR(ctx))
1154 return PTR_ERR(ctx);
1155
1156 if (offsetofend(typeof(cmd), ece) <= in_size) {
1157 ece.vendor_id = cmd.ece.vendor_id;
1158 ece.attr_mod = cmd.ece.attr_mod;
1159 }
1160
1161 if (cmd.conn_param.valid) {
1162 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1163 mutex_lock(&file->mut);
1164 mutex_lock(&ctx->mutex);
1165 ret = __rdma_accept_ece(ctx->cm_id, &conn_param, NULL, &ece);
1166 mutex_unlock(&ctx->mutex);
1167 if (!ret)
1168 ctx->uid = cmd.uid;
1169 mutex_unlock(&file->mut);
1170 } else {
1171 mutex_lock(&ctx->mutex);
1172 ret = __rdma_accept_ece(ctx->cm_id, NULL, NULL, &ece);
1173 mutex_unlock(&ctx->mutex);
1174 }
1175 ucma_put_ctx(ctx);
1176 return ret;
1177}
1178
1179static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1180 int in_len, int out_len)
1181{
1182 struct rdma_ucm_reject cmd;
1183 struct ucma_context *ctx;
1184 int ret;
1185
1186 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1187 return -EFAULT;
1188
1189 if (!cmd.reason)
1190 cmd.reason = IB_CM_REJ_CONSUMER_DEFINED;
1191
1192 switch (cmd.reason) {
1193 case IB_CM_REJ_CONSUMER_DEFINED:
1194 case IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED:
1195 break;
1196 default:
1197 return -EINVAL;
1198 }
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_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len,
1206 cmd.reason);
1207 mutex_unlock(&ctx->mutex);
1208 ucma_put_ctx(ctx);
1209 return ret;
1210}
1211
1212static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1213 int in_len, int out_len)
1214{
1215 struct rdma_ucm_disconnect cmd;
1216 struct ucma_context *ctx;
1217 int ret;
1218
1219 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1220 return -EFAULT;
1221
1222 ctx = ucma_get_ctx_dev(file, cmd.id);
1223 if (IS_ERR(ctx))
1224 return PTR_ERR(ctx);
1225
1226 mutex_lock(&ctx->mutex);
1227 ret = rdma_disconnect(ctx->cm_id);
1228 mutex_unlock(&ctx->mutex);
1229 ucma_put_ctx(ctx);
1230 return ret;
1231}
1232
1233static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1234 const char __user *inbuf,
1235 int in_len, int out_len)
1236{
1237 struct rdma_ucm_init_qp_attr cmd;
1238 struct ib_uverbs_qp_attr resp;
1239 struct ucma_context *ctx;
1240 struct ib_qp_attr qp_attr;
1241 int ret;
1242
1243 if (out_len < sizeof(resp))
1244 return -ENOSPC;
1245
1246 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1247 return -EFAULT;
1248
1249 if (cmd.qp_state > IB_QPS_ERR)
1250 return -EINVAL;
1251
1252 ctx = ucma_get_ctx_dev(file, cmd.id);
1253 if (IS_ERR(ctx))
1254 return PTR_ERR(ctx);
1255
1256 resp.qp_attr_mask = 0;
1257 memset(&qp_attr, 0, sizeof qp_attr);
1258 qp_attr.qp_state = cmd.qp_state;
1259 mutex_lock(&ctx->mutex);
1260 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1261 mutex_unlock(&ctx->mutex);
1262 if (ret)
1263 goto out;
1264
1265 ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1266 if (copy_to_user(u64_to_user_ptr(cmd.response),
1267 &resp, sizeof(resp)))
1268 ret = -EFAULT;
1269
1270out:
1271 ucma_put_ctx(ctx);
1272 return ret;
1273}
1274
1275static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1276 void *optval, size_t optlen)
1277{
1278 int ret = 0;
1279
1280 switch (optname) {
1281 case RDMA_OPTION_ID_TOS:
1282 if (optlen != sizeof(u8)) {
1283 ret = -EINVAL;
1284 break;
1285 }
1286 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1287 break;
1288 case RDMA_OPTION_ID_REUSEADDR:
1289 if (optlen != sizeof(int)) {
1290 ret = -EINVAL;
1291 break;
1292 }
1293 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1294 break;
1295 case RDMA_OPTION_ID_AFONLY:
1296 if (optlen != sizeof(int)) {
1297 ret = -EINVAL;
1298 break;
1299 }
1300 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1301 break;
1302 case RDMA_OPTION_ID_ACK_TIMEOUT:
1303 if (optlen != sizeof(u8)) {
1304 ret = -EINVAL;
1305 break;
1306 }
1307 ret = rdma_set_ack_timeout(ctx->cm_id, *((u8 *)optval));
1308 break;
1309 default:
1310 ret = -ENOSYS;
1311 }
1312
1313 return ret;
1314}
1315
1316static int ucma_set_ib_path(struct ucma_context *ctx,
1317 struct ib_path_rec_data *path_data, size_t optlen)
1318{
1319 struct sa_path_rec sa_path;
1320 struct rdma_cm_event event;
1321 int ret;
1322
1323 if (optlen % sizeof(*path_data))
1324 return -EINVAL;
1325
1326 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1327 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1328 IB_PATH_BIDIRECTIONAL))
1329 break;
1330 }
1331
1332 if (!optlen)
1333 return -EINVAL;
1334
1335 if (!ctx->cm_id->device)
1336 return -EINVAL;
1337
1338 memset(&sa_path, 0, sizeof(sa_path));
1339
1340 sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1341 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1342
1343 if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1344 struct sa_path_rec opa;
1345
1346 sa_convert_path_ib_to_opa(&opa, &sa_path);
1347 mutex_lock(&ctx->mutex);
1348 ret = rdma_set_ib_path(ctx->cm_id, &opa);
1349 mutex_unlock(&ctx->mutex);
1350 } else {
1351 mutex_lock(&ctx->mutex);
1352 ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
1353 mutex_unlock(&ctx->mutex);
1354 }
1355 if (ret)
1356 return ret;
1357
1358 memset(&event, 0, sizeof event);
1359 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1360 return ucma_event_handler(ctx->cm_id, &event);
1361}
1362
1363static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1364 void *optval, size_t optlen)
1365{
1366 int ret;
1367
1368 switch (optname) {
1369 case RDMA_OPTION_IB_PATH:
1370 ret = ucma_set_ib_path(ctx, optval, optlen);
1371 break;
1372 default:
1373 ret = -ENOSYS;
1374 }
1375
1376 return ret;
1377}
1378
1379static int ucma_set_option_level(struct ucma_context *ctx, int level,
1380 int optname, void *optval, size_t optlen)
1381{
1382 int ret;
1383
1384 switch (level) {
1385 case RDMA_OPTION_ID:
1386 mutex_lock(&ctx->mutex);
1387 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1388 mutex_unlock(&ctx->mutex);
1389 break;
1390 case RDMA_OPTION_IB:
1391 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1392 break;
1393 default:
1394 ret = -ENOSYS;
1395 }
1396
1397 return ret;
1398}
1399
1400static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1401 int in_len, int out_len)
1402{
1403 struct rdma_ucm_set_option cmd;
1404 struct ucma_context *ctx;
1405 void *optval;
1406 int ret;
1407
1408 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1409 return -EFAULT;
1410
1411 if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1412 return -EINVAL;
1413
1414 ctx = ucma_get_ctx(file, cmd.id);
1415 if (IS_ERR(ctx))
1416 return PTR_ERR(ctx);
1417
1418 optval = memdup_user(u64_to_user_ptr(cmd.optval),
1419 cmd.optlen);
1420 if (IS_ERR(optval)) {
1421 ret = PTR_ERR(optval);
1422 goto out;
1423 }
1424
1425 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1426 cmd.optlen);
1427 kfree(optval);
1428
1429out:
1430 ucma_put_ctx(ctx);
1431 return ret;
1432}
1433
1434static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1435 int in_len, int out_len)
1436{
1437 struct rdma_ucm_notify cmd;
1438 struct ucma_context *ctx;
1439 int ret = -EINVAL;
1440
1441 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1442 return -EFAULT;
1443
1444 ctx = ucma_get_ctx(file, cmd.id);
1445 if (IS_ERR(ctx))
1446 return PTR_ERR(ctx);
1447
1448 mutex_lock(&ctx->mutex);
1449 if (ctx->cm_id->device)
1450 ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
1451 mutex_unlock(&ctx->mutex);
1452
1453 ucma_put_ctx(ctx);
1454 return ret;
1455}
1456
1457static ssize_t ucma_process_join(struct ucma_file *file,
1458 struct rdma_ucm_join_mcast *cmd, int out_len)
1459{
1460 struct rdma_ucm_create_id_resp resp;
1461 struct ucma_context *ctx;
1462 struct ucma_multicast *mc;
1463 struct sockaddr *addr;
1464 int ret;
1465 u8 join_state;
1466
1467 if (out_len < sizeof(resp))
1468 return -ENOSPC;
1469
1470 addr = (struct sockaddr *) &cmd->addr;
1471 if (cmd->addr_size != rdma_addr_size(addr))
1472 return -EINVAL;
1473
1474 if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1475 join_state = BIT(FULLMEMBER_JOIN);
1476 else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1477 join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1478 else
1479 return -EINVAL;
1480
1481 ctx = ucma_get_ctx_dev(file, cmd->id);
1482 if (IS_ERR(ctx))
1483 return PTR_ERR(ctx);
1484
1485 mutex_lock(&file->mut);
1486 mc = ucma_alloc_multicast(ctx);
1487 if (!mc) {
1488 ret = -ENOMEM;
1489 goto err1;
1490 }
1491 mc->join_state = join_state;
1492 mc->uid = cmd->uid;
1493 memcpy(&mc->addr, addr, cmd->addr_size);
1494 mutex_lock(&ctx->mutex);
1495 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1496 join_state, mc);
1497 mutex_unlock(&ctx->mutex);
1498 if (ret)
1499 goto err2;
1500
1501 resp.id = mc->id;
1502 if (copy_to_user(u64_to_user_ptr(cmd->response),
1503 &resp, sizeof(resp))) {
1504 ret = -EFAULT;
1505 goto err3;
1506 }
1507
1508 xa_store(&multicast_table, mc->id, mc, 0);
1509
1510 mutex_unlock(&file->mut);
1511 ucma_put_ctx(ctx);
1512 return 0;
1513
1514err3:
1515 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1516 ucma_cleanup_mc_events(mc);
1517err2:
1518 xa_erase(&multicast_table, mc->id);
1519 list_del(&mc->list);
1520 kfree(mc);
1521err1:
1522 mutex_unlock(&file->mut);
1523 ucma_put_ctx(ctx);
1524 return ret;
1525}
1526
1527static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1528 const char __user *inbuf,
1529 int in_len, int out_len)
1530{
1531 struct rdma_ucm_join_ip_mcast cmd;
1532 struct rdma_ucm_join_mcast join_cmd;
1533
1534 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1535 return -EFAULT;
1536
1537 join_cmd.response = cmd.response;
1538 join_cmd.uid = cmd.uid;
1539 join_cmd.id = cmd.id;
1540 join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1541 if (!join_cmd.addr_size)
1542 return -EINVAL;
1543
1544 join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1545 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1546
1547 return ucma_process_join(file, &join_cmd, out_len);
1548}
1549
1550static ssize_t ucma_join_multicast(struct ucma_file *file,
1551 const char __user *inbuf,
1552 int in_len, int out_len)
1553{
1554 struct rdma_ucm_join_mcast cmd;
1555
1556 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1557 return -EFAULT;
1558
1559 if (!rdma_addr_size_kss(&cmd.addr))
1560 return -EINVAL;
1561
1562 return ucma_process_join(file, &cmd, out_len);
1563}
1564
1565static ssize_t ucma_leave_multicast(struct ucma_file *file,
1566 const char __user *inbuf,
1567 int in_len, int out_len)
1568{
1569 struct rdma_ucm_destroy_id cmd;
1570 struct rdma_ucm_destroy_id_resp resp;
1571 struct ucma_multicast *mc;
1572 int ret = 0;
1573
1574 if (out_len < sizeof(resp))
1575 return -ENOSPC;
1576
1577 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1578 return -EFAULT;
1579
1580 xa_lock(&multicast_table);
1581 mc = xa_load(&multicast_table, cmd.id);
1582 if (!mc)
1583 mc = ERR_PTR(-ENOENT);
1584 else if (mc->ctx->file != file)
1585 mc = ERR_PTR(-EINVAL);
1586 else if (!refcount_inc_not_zero(&mc->ctx->ref))
1587 mc = ERR_PTR(-ENXIO);
1588 else
1589 __xa_erase(&multicast_table, mc->id);
1590 xa_unlock(&multicast_table);
1591
1592 if (IS_ERR(mc)) {
1593 ret = PTR_ERR(mc);
1594 goto out;
1595 }
1596
1597 mutex_lock(&mc->ctx->mutex);
1598 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1599 mutex_unlock(&mc->ctx->mutex);
1600
1601 mutex_lock(&mc->ctx->file->mut);
1602 ucma_cleanup_mc_events(mc);
1603 list_del(&mc->list);
1604 mutex_unlock(&mc->ctx->file->mut);
1605
1606 ucma_put_ctx(mc->ctx);
1607 resp.events_reported = mc->events_reported;
1608 kfree(mc);
1609
1610 if (copy_to_user(u64_to_user_ptr(cmd.response),
1611 &resp, sizeof(resp)))
1612 ret = -EFAULT;
1613out:
1614 return ret;
1615}
1616
1617static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1618{
1619 /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1620 if (file1 < file2) {
1621 mutex_lock(&file1->mut);
1622 mutex_lock_nested(&file2->mut, SINGLE_DEPTH_NESTING);
1623 } else {
1624 mutex_lock(&file2->mut);
1625 mutex_lock_nested(&file1->mut, SINGLE_DEPTH_NESTING);
1626 }
1627}
1628
1629static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1630{
1631 if (file1 < file2) {
1632 mutex_unlock(&file2->mut);
1633 mutex_unlock(&file1->mut);
1634 } else {
1635 mutex_unlock(&file1->mut);
1636 mutex_unlock(&file2->mut);
1637 }
1638}
1639
1640static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1641{
1642 struct ucma_event *uevent, *tmp;
1643
1644 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1645 if (uevent->ctx == ctx)
1646 list_move_tail(&uevent->list, &file->event_list);
1647}
1648
1649static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1650 const char __user *inbuf,
1651 int in_len, int out_len)
1652{
1653 struct rdma_ucm_migrate_id cmd;
1654 struct rdma_ucm_migrate_resp resp;
1655 struct ucma_context *ctx;
1656 struct fd f;
1657 struct ucma_file *cur_file;
1658 int ret = 0;
1659
1660 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1661 return -EFAULT;
1662
1663 /* Get current fd to protect against it being closed */
1664 f = fdget(cmd.fd);
1665 if (!f.file)
1666 return -ENOENT;
1667 if (f.file->f_op != &ucma_fops) {
1668 ret = -EINVAL;
1669 goto file_put;
1670 }
1671
1672 /* Validate current fd and prevent destruction of id. */
1673 ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1674 if (IS_ERR(ctx)) {
1675 ret = PTR_ERR(ctx);
1676 goto file_put;
1677 }
1678
1679 cur_file = ctx->file;
1680 if (cur_file == new_file) {
1681 resp.events_reported = ctx->events_reported;
1682 goto response;
1683 }
1684
1685 /*
1686 * Migrate events between fd's, maintaining order, and avoiding new
1687 * events being added before existing events.
1688 */
1689 ucma_lock_files(cur_file, new_file);
1690 xa_lock(&ctx_table);
1691
1692 list_move_tail(&ctx->list, &new_file->ctx_list);
1693 ucma_move_events(ctx, new_file);
1694 ctx->file = new_file;
1695 resp.events_reported = ctx->events_reported;
1696
1697 xa_unlock(&ctx_table);
1698 ucma_unlock_files(cur_file, new_file);
1699
1700response:
1701 if (copy_to_user(u64_to_user_ptr(cmd.response),
1702 &resp, sizeof(resp)))
1703 ret = -EFAULT;
1704
1705 ucma_put_ctx(ctx);
1706file_put:
1707 fdput(f);
1708 return ret;
1709}
1710
1711static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1712 const char __user *inbuf,
1713 int in_len, int out_len) = {
1714 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1715 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1716 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1717 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1718 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1719 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1720 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1721 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1722 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1723 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1724 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1725 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1726 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1727 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1728 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1729 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1730 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1731 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1732 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1733 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1734 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1735 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1736 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1737};
1738
1739static ssize_t ucma_write(struct file *filp, const char __user *buf,
1740 size_t len, loff_t *pos)
1741{
1742 struct ucma_file *file = filp->private_data;
1743 struct rdma_ucm_cmd_hdr hdr;
1744 ssize_t ret;
1745
1746 if (!ib_safe_file_access(filp)) {
1747 pr_err_once("ucma_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1748 task_tgid_vnr(current), current->comm);
1749 return -EACCES;
1750 }
1751
1752 if (len < sizeof(hdr))
1753 return -EINVAL;
1754
1755 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1756 return -EFAULT;
1757
1758 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1759 return -EINVAL;
1760 hdr.cmd = array_index_nospec(hdr.cmd, ARRAY_SIZE(ucma_cmd_table));
1761
1762 if (hdr.in + sizeof(hdr) > len)
1763 return -EINVAL;
1764
1765 if (!ucma_cmd_table[hdr.cmd])
1766 return -ENOSYS;
1767
1768 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1769 if (!ret)
1770 ret = len;
1771
1772 return ret;
1773}
1774
1775static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1776{
1777 struct ucma_file *file = filp->private_data;
1778 __poll_t mask = 0;
1779
1780 poll_wait(filp, &file->poll_wait, wait);
1781
1782 if (!list_empty(&file->event_list))
1783 mask = EPOLLIN | EPOLLRDNORM;
1784
1785 return mask;
1786}
1787
1788/*
1789 * ucma_open() does not need the BKL:
1790 *
1791 * - no global state is referred to;
1792 * - there is no ioctl method to race against;
1793 * - no further module initialization is required for open to work
1794 * after the device is registered.
1795 */
1796static int ucma_open(struct inode *inode, struct file *filp)
1797{
1798 struct ucma_file *file;
1799
1800 file = kmalloc(sizeof *file, GFP_KERNEL);
1801 if (!file)
1802 return -ENOMEM;
1803
1804 file->close_wq = alloc_ordered_workqueue("ucma_close_id",
1805 WQ_MEM_RECLAIM);
1806 if (!file->close_wq) {
1807 kfree(file);
1808 return -ENOMEM;
1809 }
1810
1811 INIT_LIST_HEAD(&file->event_list);
1812 INIT_LIST_HEAD(&file->ctx_list);
1813 init_waitqueue_head(&file->poll_wait);
1814 mutex_init(&file->mut);
1815
1816 filp->private_data = file;
1817 file->filp = filp;
1818
1819 return stream_open(inode, filp);
1820}
1821
1822static int ucma_close(struct inode *inode, struct file *filp)
1823{
1824 struct ucma_file *file = filp->private_data;
1825 struct ucma_context *ctx, *tmp;
1826
1827 mutex_lock(&file->mut);
1828 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1829 ctx->destroying = 1;
1830 mutex_unlock(&file->mut);
1831
1832 xa_erase(&ctx_table, ctx->id);
1833 flush_workqueue(file->close_wq);
1834 /* At that step once ctx was marked as destroying and workqueue
1835 * was flushed we are safe from any inflights handlers that
1836 * might put other closing task.
1837 */
1838 xa_lock(&ctx_table);
1839 if (!ctx->closing) {
1840 xa_unlock(&ctx_table);
1841 ucma_put_ctx(ctx);
1842 wait_for_completion(&ctx->comp);
1843 /* rdma_destroy_id ensures that no event handlers are
1844 * inflight for that id before releasing it.
1845 */
1846 rdma_destroy_id(ctx->cm_id);
1847 } else {
1848 xa_unlock(&ctx_table);
1849 }
1850
1851 ucma_free_ctx(ctx);
1852 mutex_lock(&file->mut);
1853 }
1854 mutex_unlock(&file->mut);
1855 destroy_workqueue(file->close_wq);
1856 kfree(file);
1857 return 0;
1858}
1859
1860static const struct file_operations ucma_fops = {
1861 .owner = THIS_MODULE,
1862 .open = ucma_open,
1863 .release = ucma_close,
1864 .write = ucma_write,
1865 .poll = ucma_poll,
1866 .llseek = no_llseek,
1867};
1868
1869static struct miscdevice ucma_misc = {
1870 .minor = MISC_DYNAMIC_MINOR,
1871 .name = "rdma_cm",
1872 .nodename = "infiniband/rdma_cm",
1873 .mode = 0666,
1874 .fops = &ucma_fops,
1875};
1876
1877static int ucma_get_global_nl_info(struct ib_client_nl_info *res)
1878{
1879 res->abi = RDMA_USER_CM_ABI_VERSION;
1880 res->cdev = ucma_misc.this_device;
1881 return 0;
1882}
1883
1884static struct ib_client rdma_cma_client = {
1885 .name = "rdma_cm",
1886 .get_global_nl_info = ucma_get_global_nl_info,
1887};
1888MODULE_ALIAS_RDMA_CLIENT("rdma_cm");
1889
1890static ssize_t show_abi_version(struct device *dev,
1891 struct device_attribute *attr,
1892 char *buf)
1893{
1894 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1895}
1896static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1897
1898static int __init ucma_init(void)
1899{
1900 int ret;
1901
1902 ret = misc_register(&ucma_misc);
1903 if (ret)
1904 return ret;
1905
1906 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1907 if (ret) {
1908 pr_err("rdma_ucm: couldn't create abi_version attr\n");
1909 goto err1;
1910 }
1911
1912 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1913 if (!ucma_ctl_table_hdr) {
1914 pr_err("rdma_ucm: couldn't register sysctl paths\n");
1915 ret = -ENOMEM;
1916 goto err2;
1917 }
1918
1919 ret = ib_register_client(&rdma_cma_client);
1920 if (ret)
1921 goto err3;
1922
1923 return 0;
1924err3:
1925 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1926err2:
1927 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1928err1:
1929 misc_deregister(&ucma_misc);
1930 return ret;
1931}
1932
1933static void __exit ucma_cleanup(void)
1934{
1935 ib_unregister_client(&rdma_cma_client);
1936 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1937 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1938 misc_deregister(&ucma_misc);
1939}
1940
1941module_init(ucma_init);
1942module_exit(ucma_cleanup);