?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
WIP FPC-III support
[linux/fpc-iii.git] / drivers / infiniband / core / ucma.c
blob7dab9a27a145a8eee04150c3caa11f2988bae584
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
59 MODULE_AUTHOR("Sean Hefty");
60 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
61 MODULE_LICENSE("Dual BSD/GPL");
62
63 static unsigned int max_backlog = 1024;
64
65 static struct ctl_table_header *ucma_ctl_table_hdr;
66 static 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
77 struct ucma_file {
78 struct mutex mut;
79 struct file *filp;
80 struct list_head ctx_list;
81 struct list_head event_list;
82 wait_queue_head_t poll_wait;
83 };
84
85 struct ucma_context {
86 u32 id;
87 struct completion comp;
88 refcount_t ref;
89 int events_reported;
90 atomic_t backlog;
91
92 struct ucma_file *file;
93 struct rdma_cm_id *cm_id;
94 struct mutex mutex;
95 u64 uid;
96
97 struct list_head list;
98 /* sync between removal event and id destroy, protected by file mut */
99 int destroying;
100 struct work_struct close_work;
101 };
102
103 struct ucma_multicast {
104 struct ucma_context *ctx;
105 u32 id;
106 int events_reported;
107
108 u64 uid;
109 u8 join_state;
110 struct sockaddr_storage addr;
111 };
112
113 struct ucma_event {
114 struct ucma_context *ctx;
115 struct ucma_context *conn_req_ctx;
116 struct ucma_multicast *mc;
117 struct list_head list;
118 struct rdma_ucm_event_resp resp;
119 };
120
121 static DEFINE_XARRAY_ALLOC(ctx_table);
122 static DEFINE_XARRAY_ALLOC(multicast_table);
123
124 static const struct file_operations ucma_fops;
125 static int __destroy_id(struct ucma_context *ctx);
126
127 static inline struct ucma_context *_ucma_find_context(int id,
128 struct ucma_file *file)
129 {
130 struct ucma_context *ctx;
131
132 ctx = xa_load(&ctx_table, id);
133 if (!ctx)
134 ctx = ERR_PTR(-ENOENT);
135 else if (ctx->file != file)
136 ctx = ERR_PTR(-EINVAL);
137 return ctx;
138 }
139
140 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
141 {
142 struct ucma_context *ctx;
143
144 xa_lock(&ctx_table);
145 ctx = _ucma_find_context(id, file);
146 if (!IS_ERR(ctx))
147 if (!refcount_inc_not_zero(&ctx->ref))
148 ctx = ERR_PTR(-ENXIO);
149 xa_unlock(&ctx_table);
150 return ctx;
151 }
152
153 static void ucma_put_ctx(struct ucma_context *ctx)
154 {
155 if (refcount_dec_and_test(&ctx->ref))
156 complete(&ctx->comp);
157 }
158
159 /*
160 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
161 * CM_ID is bound.
162 */
163 static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
164 {
165 struct ucma_context *ctx = ucma_get_ctx(file, id);
166
167 if (IS_ERR(ctx))
168 return ctx;
169 if (!ctx->cm_id->device) {
170 ucma_put_ctx(ctx);
171 return ERR_PTR(-EINVAL);
172 }
173 return ctx;
174 }
175
176 static void ucma_close_id(struct work_struct *work)
177 {
178 struct ucma_context *ctx = container_of(work, struct ucma_context, close_work);
179
180 /* once all inflight tasks are finished, we close all underlying
181 * resources. The context is still alive till its explicit destryoing
182 * by its creator.
183 */
184 ucma_put_ctx(ctx);
185 wait_for_completion(&ctx->comp);
186 /* No new events will be generated after destroying the id. */
187 rdma_destroy_id(ctx->cm_id);
188
189 /*
190 * At this point ctx->ref is zero so the only place the ctx can be is in
191 * a uevent or in __destroy_id(). Since the former doesn't touch
192 * ctx->cm_id and the latter sync cancels this, there is no races with
193 * this store.
194 */
195 ctx->cm_id = NULL;
196 }
197
198 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
199 {
200 struct ucma_context *ctx;
201
202 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
203 if (!ctx)
204 return NULL;
205
206 INIT_WORK(&ctx->close_work, ucma_close_id);
207 refcount_set(&ctx->ref, 1);
208 init_completion(&ctx->comp);
209 /* So list_del() will work if we don't do ucma_finish_ctx() */
210 INIT_LIST_HEAD(&ctx->list);
211 ctx->file = file;
212 mutex_init(&ctx->mutex);
213
214 if (xa_alloc(&ctx_table, &ctx->id, NULL, xa_limit_32b, GFP_KERNEL)) {
215 kfree(ctx);
216 return NULL;
217 }
218 return ctx;
219 }
220
221 static void ucma_finish_ctx(struct ucma_context *ctx)
222 {
223 lockdep_assert_held(&ctx->file->mut);
224 list_add_tail(&ctx->list, &ctx->file->ctx_list);
225 xa_store(&ctx_table, ctx->id, ctx, GFP_KERNEL);
226 }
227
228 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
229 struct rdma_conn_param *src)
230 {
231 if (src->private_data_len)
232 memcpy(dst->private_data, src->private_data,
233 src->private_data_len);
234 dst->private_data_len = src->private_data_len;
235 dst->responder_resources =src->responder_resources;
236 dst->initiator_depth = src->initiator_depth;
237 dst->flow_control = src->flow_control;
238 dst->retry_count = src->retry_count;
239 dst->rnr_retry_count = src->rnr_retry_count;
240 dst->srq = src->srq;
241 dst->qp_num = src->qp_num;
242 }
243
244 static void ucma_copy_ud_event(struct ib_device *device,
245 struct rdma_ucm_ud_param *dst,
246 struct rdma_ud_param *src)
247 {
248 if (src->private_data_len)
249 memcpy(dst->private_data, src->private_data,
250 src->private_data_len);
251 dst->private_data_len = src->private_data_len;
252 ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
253 dst->qp_num = src->qp_num;
254 dst->qkey = src->qkey;
255 }
256
257 static struct ucma_event *ucma_create_uevent(struct ucma_context *ctx,
258 struct rdma_cm_event *event)
259 {
260 struct ucma_event *uevent;
261
262 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
263 if (!uevent)
264 return NULL;
265
266 uevent->ctx = ctx;
267 switch (event->event) {
268 case RDMA_CM_EVENT_MULTICAST_JOIN:
269 case RDMA_CM_EVENT_MULTICAST_ERROR:
270 uevent->mc = (struct ucma_multicast *)
271 event->param.ud.private_data;
272 uevent->resp.uid = uevent->mc->uid;
273 uevent->resp.id = uevent->mc->id;
274 break;
275 default:
276 uevent->resp.uid = ctx->uid;
277 uevent->resp.id = ctx->id;
278 break;
279 }
280 uevent->resp.event = event->event;
281 uevent->resp.status = event->status;
282 if (ctx->cm_id->qp_type == IB_QPT_UD)
283 ucma_copy_ud_event(ctx->cm_id->device, &uevent->resp.param.ud,
284 &event->param.ud);
285 else
286 ucma_copy_conn_event(&uevent->resp.param.conn,
287 &event->param.conn);
288
289 uevent->resp.ece.vendor_id = event->ece.vendor_id;
290 uevent->resp.ece.attr_mod = event->ece.attr_mod;
291 return uevent;
292 }
293
294 static int ucma_connect_event_handler(struct rdma_cm_id *cm_id,
295 struct rdma_cm_event *event)
296 {
297 struct ucma_context *listen_ctx = cm_id->context;
298 struct ucma_context *ctx;
299 struct ucma_event *uevent;
300
301 if (!atomic_add_unless(&listen_ctx->backlog, -1, 0))
302 return -ENOMEM;
303 ctx = ucma_alloc_ctx(listen_ctx->file);
304 if (!ctx)
305 goto err_backlog;
306 ctx->cm_id = cm_id;
307
308 uevent = ucma_create_uevent(listen_ctx, event);
309 if (!uevent)
310 goto err_alloc;
311 uevent->conn_req_ctx = ctx;
312 uevent->resp.id = ctx->id;
313
314 ctx->cm_id->context = ctx;
315
316 mutex_lock(&ctx->file->mut);
317 ucma_finish_ctx(ctx);
318 list_add_tail(&uevent->list, &ctx->file->event_list);
319 mutex_unlock(&ctx->file->mut);
320 wake_up_interruptible(&ctx->file->poll_wait);
321 return 0;
322
323 err_alloc:
324 xa_erase(&ctx_table, ctx->id);
325 kfree(ctx);
326 err_backlog:
327 atomic_inc(&listen_ctx->backlog);
328 /* Returning error causes the new ID to be destroyed */
329 return -ENOMEM;
330 }
331
332 static 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 && !ctx->destroying)
360 queue_work(system_unbound_wq, &ctx->close_work);
361 return 0;
362 }
363
364 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
365 int in_len, int out_len)
366 {
367 struct rdma_ucm_get_event cmd;
368 struct ucma_event *uevent;
369
370 /*
371 * Old 32 bit user space does not send the 4 byte padding in the
372 * reserved field. We don't care, allow it to keep working.
373 */
374 if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved) -
375 sizeof(uevent->resp.ece))
376 return -ENOSPC;
377
378 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
379 return -EFAULT;
380
381 mutex_lock(&file->mut);
382 while (list_empty(&file->event_list)) {
383 mutex_unlock(&file->mut);
384
385 if (file->filp->f_flags & O_NONBLOCK)
386 return -EAGAIN;
387
388 if (wait_event_interruptible(file->poll_wait,
389 !list_empty(&file->event_list)))
390 return -ERESTARTSYS;
391
392 mutex_lock(&file->mut);
393 }
394
395 uevent = list_first_entry(&file->event_list, struct ucma_event, list);
396
397 if (copy_to_user(u64_to_user_ptr(cmd.response),
398 &uevent->resp,
399 min_t(size_t, out_len, sizeof(uevent->resp)))) {
400 mutex_unlock(&file->mut);
401 return -EFAULT;
402 }
403
404 list_del(&uevent->list);
405 uevent->ctx->events_reported++;
406 if (uevent->mc)
407 uevent->mc->events_reported++;
408 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
409 atomic_inc(&uevent->ctx->backlog);
410 mutex_unlock(&file->mut);
411
412 kfree(uevent);
413 return 0;
414 }
415
416 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
417 {
418 switch (cmd->ps) {
419 case RDMA_PS_TCP:
420 *qp_type = IB_QPT_RC;
421 return 0;
422 case RDMA_PS_UDP:
423 case RDMA_PS_IPOIB:
424 *qp_type = IB_QPT_UD;
425 return 0;
426 case RDMA_PS_IB:
427 *qp_type = cmd->qp_type;
428 return 0;
429 default:
430 return -EINVAL;
431 }
432 }
433
434 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
435 int in_len, int out_len)
436 {
437 struct rdma_ucm_create_id cmd;
438 struct rdma_ucm_create_id_resp resp;
439 struct ucma_context *ctx;
440 struct rdma_cm_id *cm_id;
441 enum ib_qp_type qp_type;
442 int ret;
443
444 if (out_len < sizeof(resp))
445 return -ENOSPC;
446
447 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
448 return -EFAULT;
449
450 ret = ucma_get_qp_type(&cmd, &qp_type);
451 if (ret)
452 return ret;
453
454 ctx = ucma_alloc_ctx(file);
455 if (!ctx)
456 return -ENOMEM;
457
458 ctx->uid = cmd.uid;
459 cm_id = rdma_create_user_id(ucma_event_handler, ctx, cmd.ps, qp_type);
460 if (IS_ERR(cm_id)) {
461 ret = PTR_ERR(cm_id);
462 goto err1;
463 }
464 ctx->cm_id = cm_id;
465
466 resp.id = ctx->id;
467 if (copy_to_user(u64_to_user_ptr(cmd.response),
468 &resp, sizeof(resp))) {
469 xa_erase(&ctx_table, ctx->id);
470 __destroy_id(ctx);
471 return -EFAULT;
472 }
473
474 mutex_lock(&file->mut);
475 ucma_finish_ctx(ctx);
476 mutex_unlock(&file->mut);
477 return 0;
478
479 err1:
480 xa_erase(&ctx_table, ctx->id);
481 kfree(ctx);
482 return ret;
483 }
484
485 static void ucma_cleanup_multicast(struct ucma_context *ctx)
486 {
487 struct ucma_multicast *mc;
488 unsigned long index;
489
490 xa_for_each(&multicast_table, index, mc) {
491 if (mc->ctx != ctx)
492 continue;
493 /*
494 * At this point mc->ctx->ref is 0 so the mc cannot leave the
495 * lock on the reader and this is enough serialization
496 */
497 xa_erase(&multicast_table, index);
498 kfree(mc);
499 }
500 }
501
502 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
503 {
504 struct ucma_event *uevent, *tmp;
505
506 rdma_lock_handler(mc->ctx->cm_id);
507 mutex_lock(&mc->ctx->file->mut);
508 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
509 if (uevent->mc != mc)
510 continue;
511
512 list_del(&uevent->list);
513 kfree(uevent);
514 }
515 mutex_unlock(&mc->ctx->file->mut);
516 rdma_unlock_handler(mc->ctx->cm_id);
517 }
518
519 /*
520 * ucma_free_ctx is called after the underlying rdma CM-ID is destroyed. At
521 * this point, no new events will be reported from the hardware. However, we
522 * still need to cleanup the UCMA context for this ID. Specifically, there
523 * might be events that have not yet been consumed by the user space software.
524 * mutex. After that we release them as needed.
525 */
526 static int ucma_free_ctx(struct ucma_context *ctx)
527 {
528 int events_reported;
529 struct ucma_event *uevent, *tmp;
530 LIST_HEAD(list);
531
532 ucma_cleanup_multicast(ctx);
533
534 /* Cleanup events not yet reported to the user. */
535 mutex_lock(&ctx->file->mut);
536 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
537 if (uevent->ctx == ctx || uevent->conn_req_ctx == ctx)
538 list_move_tail(&uevent->list, &list);
539 }
540 list_del(&ctx->list);
541 events_reported = ctx->events_reported;
542 mutex_unlock(&ctx->file->mut);
543
544 /*
545 * If this was a listening ID then any connections spawned from it
546 * that have not been delivered to userspace are cleaned up too.
547 * Must be done outside any locks.
548 */
549 list_for_each_entry_safe(uevent, tmp, &list, list) {
550 list_del(&uevent->list);
551 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST &&
552 uevent->conn_req_ctx != ctx)
553 __destroy_id(uevent->conn_req_ctx);
554 kfree(uevent);
555 }
556
557 mutex_destroy(&ctx->mutex);
558 kfree(ctx);
559 return events_reported;
560 }
561
562 static int __destroy_id(struct ucma_context *ctx)
563 {
564 /*
565 * If the refcount is already 0 then ucma_close_id() has already
566 * destroyed the cm_id, otherwise holding the refcount keeps cm_id
567 * valid. Prevent queue_work() from being called.
568 */
569 if (refcount_inc_not_zero(&ctx->ref)) {
570 rdma_lock_handler(ctx->cm_id);
571 ctx->destroying = 1;
572 rdma_unlock_handler(ctx->cm_id);
573 ucma_put_ctx(ctx);
574 }
575
576 cancel_work_sync(&ctx->close_work);
577 /* At this point it's guaranteed that there is no inflight closing task */
578 if (ctx->cm_id)
579 ucma_close_id(&ctx->close_work);
580 return ucma_free_ctx(ctx);
581 }
582
583 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
584 int in_len, int out_len)
585 {
586 struct rdma_ucm_destroy_id cmd;
587 struct rdma_ucm_destroy_id_resp resp;
588 struct ucma_context *ctx;
589 int ret = 0;
590
591 if (out_len < sizeof(resp))
592 return -ENOSPC;
593
594 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
595 return -EFAULT;
596
597 xa_lock(&ctx_table);
598 ctx = _ucma_find_context(cmd.id, file);
599 if (!IS_ERR(ctx))
600 __xa_erase(&ctx_table, ctx->id);
601 xa_unlock(&ctx_table);
602
603 if (IS_ERR(ctx))
604 return PTR_ERR(ctx);
605
606 resp.events_reported = __destroy_id(ctx);
607 if (copy_to_user(u64_to_user_ptr(cmd.response),
608 &resp, sizeof(resp)))
609 ret = -EFAULT;
610
611 return ret;
612 }
613
614 static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
615 int in_len, int out_len)
616 {
617 struct rdma_ucm_bind_ip cmd;
618 struct ucma_context *ctx;
619 int ret;
620
621 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
622 return -EFAULT;
623
624 if (!rdma_addr_size_in6(&cmd.addr))
625 return -EINVAL;
626
627 ctx = ucma_get_ctx(file, cmd.id);
628 if (IS_ERR(ctx))
629 return PTR_ERR(ctx);
630
631 mutex_lock(&ctx->mutex);
632 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
633 mutex_unlock(&ctx->mutex);
634
635 ucma_put_ctx(ctx);
636 return ret;
637 }
638
639 static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
640 int in_len, int out_len)
641 {
642 struct rdma_ucm_bind cmd;
643 struct ucma_context *ctx;
644 int ret;
645
646 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
647 return -EFAULT;
648
649 if (cmd.reserved || !cmd.addr_size ||
650 cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
651 return -EINVAL;
652
653 ctx = ucma_get_ctx(file, cmd.id);
654 if (IS_ERR(ctx))
655 return PTR_ERR(ctx);
656
657 mutex_lock(&ctx->mutex);
658 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
659 mutex_unlock(&ctx->mutex);
660 ucma_put_ctx(ctx);
661 return ret;
662 }
663
664 static ssize_t ucma_resolve_ip(struct ucma_file *file,
665 const char __user *inbuf,
666 int in_len, int out_len)
667 {
668 struct rdma_ucm_resolve_ip cmd;
669 struct ucma_context *ctx;
670 int ret;
671
672 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
673 return -EFAULT;
674
675 if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
676 !rdma_addr_size_in6(&cmd.dst_addr))
677 return -EINVAL;
678
679 ctx = ucma_get_ctx(file, cmd.id);
680 if (IS_ERR(ctx))
681 return PTR_ERR(ctx);
682
683 mutex_lock(&ctx->mutex);
684 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
685 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
686 mutex_unlock(&ctx->mutex);
687 ucma_put_ctx(ctx);
688 return ret;
689 }
690
691 static ssize_t ucma_resolve_addr(struct ucma_file *file,
692 const char __user *inbuf,
693 int in_len, int out_len)
694 {
695 struct rdma_ucm_resolve_addr cmd;
696 struct ucma_context *ctx;
697 int ret;
698
699 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
700 return -EFAULT;
701
702 if (cmd.reserved ||
703 (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
704 !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
705 return -EINVAL;
706
707 ctx = ucma_get_ctx(file, cmd.id);
708 if (IS_ERR(ctx))
709 return PTR_ERR(ctx);
710
711 mutex_lock(&ctx->mutex);
712 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
713 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
714 mutex_unlock(&ctx->mutex);
715 ucma_put_ctx(ctx);
716 return ret;
717 }
718
719 static ssize_t ucma_resolve_route(struct ucma_file *file,
720 const char __user *inbuf,
721 int in_len, int out_len)
722 {
723 struct rdma_ucm_resolve_route cmd;
724 struct ucma_context *ctx;
725 int ret;
726
727 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
728 return -EFAULT;
729
730 ctx = ucma_get_ctx_dev(file, cmd.id);
731 if (IS_ERR(ctx))
732 return PTR_ERR(ctx);
733
734 mutex_lock(&ctx->mutex);
735 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
736 mutex_unlock(&ctx->mutex);
737 ucma_put_ctx(ctx);
738 return ret;
739 }
740
741 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
742 struct rdma_route *route)
743 {
744 struct rdma_dev_addr *dev_addr;
745
746 resp->num_paths = route->num_paths;
747 switch (route->num_paths) {
748 case 0:
749 dev_addr = &route->addr.dev_addr;
750 rdma_addr_get_dgid(dev_addr,
751 (union ib_gid *) &resp->ib_route[0].dgid);
752 rdma_addr_get_sgid(dev_addr,
753 (union ib_gid *) &resp->ib_route[0].sgid);
754 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
755 break;
756 case 2:
757 ib_copy_path_rec_to_user(&resp->ib_route[1],
758 &route->path_rec[1]);
759 fallthrough;
760 case 1:
761 ib_copy_path_rec_to_user(&resp->ib_route[0],
762 &route->path_rec[0]);
763 break;
764 default:
765 break;
766 }
767 }
768
769 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
770 struct rdma_route *route)
771 {
772
773 resp->num_paths = route->num_paths;
774 switch (route->num_paths) {
775 case 0:
776 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
777 (union ib_gid *)&resp->ib_route[0].dgid);
778 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
779 (union ib_gid *)&resp->ib_route[0].sgid);
780 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
781 break;
782 case 2:
783 ib_copy_path_rec_to_user(&resp->ib_route[1],
784 &route->path_rec[1]);
785 fallthrough;
786 case 1:
787 ib_copy_path_rec_to_user(&resp->ib_route[0],
788 &route->path_rec[0]);
789 break;
790 default:
791 break;
792 }
793 }
794
795 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
796 struct rdma_route *route)
797 {
798 struct rdma_dev_addr *dev_addr;
799
800 dev_addr = &route->addr.dev_addr;
801 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
802 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
803 }
804
805 static ssize_t ucma_query_route(struct ucma_file *file,
806 const char __user *inbuf,
807 int in_len, int out_len)
808 {
809 struct rdma_ucm_query cmd;
810 struct rdma_ucm_query_route_resp resp;
811 struct ucma_context *ctx;
812 struct sockaddr *addr;
813 int ret = 0;
814
815 if (out_len < offsetof(struct rdma_ucm_query_route_resp, ibdev_index))
816 return -ENOSPC;
817
818 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
819 return -EFAULT;
820
821 ctx = ucma_get_ctx(file, cmd.id);
822 if (IS_ERR(ctx))
823 return PTR_ERR(ctx);
824
825 mutex_lock(&ctx->mutex);
826 memset(&resp, 0, sizeof resp);
827 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
828 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
829 sizeof(struct sockaddr_in) :
830 sizeof(struct sockaddr_in6));
831 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
832 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
833 sizeof(struct sockaddr_in) :
834 sizeof(struct sockaddr_in6));
835 if (!ctx->cm_id->device)
836 goto out;
837
838 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
839 resp.ibdev_index = ctx->cm_id->device->index;
840 resp.port_num = ctx->cm_id->port_num;
841
842 if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
843 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
844 else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
845 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
846 else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
847 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
848
849 out:
850 mutex_unlock(&ctx->mutex);
851 if (copy_to_user(u64_to_user_ptr(cmd.response), &resp,
852 min_t(size_t, out_len, sizeof(resp))))
853 ret = -EFAULT;
854
855 ucma_put_ctx(ctx);
856 return ret;
857 }
858
859 static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
860 struct rdma_ucm_query_addr_resp *resp)
861 {
862 if (!cm_id->device)
863 return;
864
865 resp->node_guid = (__force __u64) cm_id->device->node_guid;
866 resp->ibdev_index = cm_id->device->index;
867 resp->port_num = cm_id->port_num;
868 resp->pkey = (__force __u16) cpu_to_be16(
869 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
870 }
871
872 static ssize_t ucma_query_addr(struct ucma_context *ctx,
873 void __user *response, int out_len)
874 {
875 struct rdma_ucm_query_addr_resp resp;
876 struct sockaddr *addr;
877 int ret = 0;
878
879 if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
880 return -ENOSPC;
881
882 memset(&resp, 0, sizeof resp);
883
884 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
885 resp.src_size = rdma_addr_size(addr);
886 memcpy(&resp.src_addr, addr, resp.src_size);
887
888 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
889 resp.dst_size = rdma_addr_size(addr);
890 memcpy(&resp.dst_addr, addr, resp.dst_size);
891
892 ucma_query_device_addr(ctx->cm_id, &resp);
893
894 if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
895 ret = -EFAULT;
896
897 return ret;
898 }
899
900 static ssize_t ucma_query_path(struct ucma_context *ctx,
901 void __user *response, int out_len)
902 {
903 struct rdma_ucm_query_path_resp *resp;
904 int i, ret = 0;
905
906 if (out_len < sizeof(*resp))
907 return -ENOSPC;
908
909 resp = kzalloc(out_len, GFP_KERNEL);
910 if (!resp)
911 return -ENOMEM;
912
913 resp->num_paths = ctx->cm_id->route.num_paths;
914 for (i = 0, out_len -= sizeof(*resp);
915 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
916 i++, out_len -= sizeof(struct ib_path_rec_data)) {
917 struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
918
919 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
920 IB_PATH_BIDIRECTIONAL;
921 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
922 struct sa_path_rec ib;
923
924 sa_convert_path_opa_to_ib(&ib, rec);
925 ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
926
927 } else {
928 ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
929 }
930 }
931
932 if (copy_to_user(response, resp, struct_size(resp, path_data, i)))
933 ret = -EFAULT;
934
935 kfree(resp);
936 return ret;
937 }
938
939 static ssize_t ucma_query_gid(struct ucma_context *ctx,
940 void __user *response, int out_len)
941 {
942 struct rdma_ucm_query_addr_resp resp;
943 struct sockaddr_ib *addr;
944 int ret = 0;
945
946 if (out_len < offsetof(struct rdma_ucm_query_addr_resp, ibdev_index))
947 return -ENOSPC;
948
949 memset(&resp, 0, sizeof resp);
950
951 ucma_query_device_addr(ctx->cm_id, &resp);
952
953 addr = (struct sockaddr_ib *) &resp.src_addr;
954 resp.src_size = sizeof(*addr);
955 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
956 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
957 } else {
958 addr->sib_family = AF_IB;
959 addr->sib_pkey = (__force __be16) resp.pkey;
960 rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
961 NULL);
962 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
963 &ctx->cm_id->route.addr.src_addr);
964 }
965
966 addr = (struct sockaddr_ib *) &resp.dst_addr;
967 resp.dst_size = sizeof(*addr);
968 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
969 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
970 } else {
971 addr->sib_family = AF_IB;
972 addr->sib_pkey = (__force __be16) resp.pkey;
973 rdma_read_gids(ctx->cm_id, NULL,
974 (union ib_gid *)&addr->sib_addr);
975 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
976 &ctx->cm_id->route.addr.dst_addr);
977 }
978
979 if (copy_to_user(response, &resp, min_t(size_t, out_len, sizeof(resp))))
980 ret = -EFAULT;
981
982 return ret;
983 }
984
985 static ssize_t ucma_query(struct ucma_file *file,
986 const char __user *inbuf,
987 int in_len, int out_len)
988 {
989 struct rdma_ucm_query cmd;
990 struct ucma_context *ctx;
991 void __user *response;
992 int ret;
993
994 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
995 return -EFAULT;
996
997 response = u64_to_user_ptr(cmd.response);
998 ctx = ucma_get_ctx(file, cmd.id);
999 if (IS_ERR(ctx))
1000 return PTR_ERR(ctx);
1001
1002 mutex_lock(&ctx->mutex);
1003 switch (cmd.option) {
1004 case RDMA_USER_CM_QUERY_ADDR:
1005 ret = ucma_query_addr(ctx, response, out_len);
1006 break;
1007 case RDMA_USER_CM_QUERY_PATH:
1008 ret = ucma_query_path(ctx, response, out_len);
1009 break;
1010 case RDMA_USER_CM_QUERY_GID:
1011 ret = ucma_query_gid(ctx, response, out_len);
1012 break;
1013 default:
1014 ret = -ENOSYS;
1015 break;
1016 }
1017 mutex_unlock(&ctx->mutex);
1018
1019 ucma_put_ctx(ctx);
1020 return ret;
1021 }
1022
1023 static void ucma_copy_conn_param(struct rdma_cm_id *id,
1024 struct rdma_conn_param *dst,
1025 struct rdma_ucm_conn_param *src)
1026 {
1027 dst->private_data = src->private_data;
1028 dst->private_data_len = src->private_data_len;
1029 dst->responder_resources =src->responder_resources;
1030 dst->initiator_depth = src->initiator_depth;
1031 dst->flow_control = src->flow_control;
1032 dst->retry_count = src->retry_count;
1033 dst->rnr_retry_count = src->rnr_retry_count;
1034 dst->srq = src->srq;
1035 dst->qp_num = src->qp_num & 0xFFFFFF;
1036 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1037 }
1038
1039 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1040 int in_len, int out_len)
1041 {
1042 struct rdma_conn_param conn_param;
1043 struct rdma_ucm_ece ece = {};
1044 struct rdma_ucm_connect cmd;
1045 struct ucma_context *ctx;
1046 size_t in_size;
1047 int ret;
1048
1049 if (in_len < offsetofend(typeof(cmd), reserved))
1050 return -EINVAL;
1051 in_size = min_t(size_t, in_len, sizeof(cmd));
1052 if (copy_from_user(&cmd, inbuf, in_size))
1053 return -EFAULT;
1054
1055 if (!cmd.conn_param.valid)
1056 return -EINVAL;
1057
1058 ctx = ucma_get_ctx_dev(file, cmd.id);
1059 if (IS_ERR(ctx))
1060 return PTR_ERR(ctx);
1061
1062 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1063 if (offsetofend(typeof(cmd), ece) <= in_size) {
1064 ece.vendor_id = cmd.ece.vendor_id;
1065 ece.attr_mod = cmd.ece.attr_mod;
1066 }
1067
1068 mutex_lock(&ctx->mutex);
1069 ret = rdma_connect_ece(ctx->cm_id, &conn_param, &ece);
1070 mutex_unlock(&ctx->mutex);
1071 ucma_put_ctx(ctx);
1072 return ret;
1073 }
1074
1075 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1076 int in_len, int out_len)
1077 {
1078 struct rdma_ucm_listen cmd;
1079 struct ucma_context *ctx;
1080 int ret;
1081
1082 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1083 return -EFAULT;
1084
1085 ctx = ucma_get_ctx(file, cmd.id);
1086 if (IS_ERR(ctx))
1087 return PTR_ERR(ctx);
1088
1089 if (cmd.backlog <= 0 || cmd.backlog > max_backlog)
1090 cmd.backlog = max_backlog;
1091 atomic_set(&ctx->backlog, cmd.backlog);
1092
1093 mutex_lock(&ctx->mutex);
1094 ret = rdma_listen(ctx->cm_id, cmd.backlog);
1095 mutex_unlock(&ctx->mutex);
1096 ucma_put_ctx(ctx);
1097 return ret;
1098 }
1099
1100 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1101 int in_len, int out_len)
1102 {
1103 struct rdma_ucm_accept cmd;
1104 struct rdma_conn_param conn_param;
1105 struct rdma_ucm_ece ece = {};
1106 struct ucma_context *ctx;
1107 size_t in_size;
1108 int ret;
1109
1110 if (in_len < offsetofend(typeof(cmd), reserved))
1111 return -EINVAL;
1112 in_size = min_t(size_t, in_len, sizeof(cmd));
1113 if (copy_from_user(&cmd, inbuf, in_size))
1114 return -EFAULT;
1115
1116 ctx = ucma_get_ctx_dev(file, cmd.id);
1117 if (IS_ERR(ctx))
1118 return PTR_ERR(ctx);
1119
1120 if (offsetofend(typeof(cmd), ece) <= in_size) {
1121 ece.vendor_id = cmd.ece.vendor_id;
1122 ece.attr_mod = cmd.ece.attr_mod;
1123 }
1124
1125 if (cmd.conn_param.valid) {
1126 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1127 mutex_lock(&ctx->mutex);
1128 rdma_lock_handler(ctx->cm_id);
1129 ret = rdma_accept_ece(ctx->cm_id, &conn_param, &ece);
1130 if (!ret) {
1131 /* The uid must be set atomically with the handler */
1132 ctx->uid = cmd.uid;
1133 }
1134 rdma_unlock_handler(ctx->cm_id);
1135 mutex_unlock(&ctx->mutex);
1136 } else {
1137 mutex_lock(&ctx->mutex);
1138 rdma_lock_handler(ctx->cm_id);
1139 ret = rdma_accept_ece(ctx->cm_id, NULL, &ece);
1140 rdma_unlock_handler(ctx->cm_id);
1141 mutex_unlock(&ctx->mutex);
1142 }
1143 ucma_put_ctx(ctx);
1144 return ret;
1145 }
1146
1147 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1148 int in_len, int out_len)
1149 {
1150 struct rdma_ucm_reject cmd;
1151 struct ucma_context *ctx;
1152 int ret;
1153
1154 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1155 return -EFAULT;
1156
1157 if (!cmd.reason)
1158 cmd.reason = IB_CM_REJ_CONSUMER_DEFINED;
1159
1160 switch (cmd.reason) {
1161 case IB_CM_REJ_CONSUMER_DEFINED:
1162 case IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED:
1163 break;
1164 default:
1165 return -EINVAL;
1166 }
1167
1168 ctx = ucma_get_ctx_dev(file, cmd.id);
1169 if (IS_ERR(ctx))
1170 return PTR_ERR(ctx);
1171
1172 mutex_lock(&ctx->mutex);
1173 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len,
1174 cmd.reason);
1175 mutex_unlock(&ctx->mutex);
1176 ucma_put_ctx(ctx);
1177 return ret;
1178 }
1179
1180 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1181 int in_len, int out_len)
1182 {
1183 struct rdma_ucm_disconnect cmd;
1184 struct ucma_context *ctx;
1185 int ret;
1186
1187 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1188 return -EFAULT;
1189
1190 ctx = ucma_get_ctx_dev(file, cmd.id);
1191 if (IS_ERR(ctx))
1192 return PTR_ERR(ctx);
1193
1194 mutex_lock(&ctx->mutex);
1195 ret = rdma_disconnect(ctx->cm_id);
1196 mutex_unlock(&ctx->mutex);
1197 ucma_put_ctx(ctx);
1198 return ret;
1199 }
1200
1201 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1202 const char __user *inbuf,
1203 int in_len, int out_len)
1204 {
1205 struct rdma_ucm_init_qp_attr cmd;
1206 struct ib_uverbs_qp_attr resp;
1207 struct ucma_context *ctx;
1208 struct ib_qp_attr qp_attr;
1209 int ret;
1210
1211 if (out_len < sizeof(resp))
1212 return -ENOSPC;
1213
1214 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1215 return -EFAULT;
1216
1217 if (cmd.qp_state > IB_QPS_ERR)
1218 return -EINVAL;
1219
1220 ctx = ucma_get_ctx_dev(file, cmd.id);
1221 if (IS_ERR(ctx))
1222 return PTR_ERR(ctx);
1223
1224 resp.qp_attr_mask = 0;
1225 memset(&qp_attr, 0, sizeof qp_attr);
1226 qp_attr.qp_state = cmd.qp_state;
1227 mutex_lock(&ctx->mutex);
1228 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1229 mutex_unlock(&ctx->mutex);
1230 if (ret)
1231 goto out;
1232
1233 ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1234 if (copy_to_user(u64_to_user_ptr(cmd.response),
1235 &resp, sizeof(resp)))
1236 ret = -EFAULT;
1237
1238 out:
1239 ucma_put_ctx(ctx);
1240 return ret;
1241 }
1242
1243 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1244 void *optval, size_t optlen)
1245 {
1246 int ret = 0;
1247
1248 switch (optname) {
1249 case RDMA_OPTION_ID_TOS:
1250 if (optlen != sizeof(u8)) {
1251 ret = -EINVAL;
1252 break;
1253 }
1254 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1255 break;
1256 case RDMA_OPTION_ID_REUSEADDR:
1257 if (optlen != sizeof(int)) {
1258 ret = -EINVAL;
1259 break;
1260 }
1261 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1262 break;
1263 case RDMA_OPTION_ID_AFONLY:
1264 if (optlen != sizeof(int)) {
1265 ret = -EINVAL;
1266 break;
1267 }
1268 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1269 break;
1270 case RDMA_OPTION_ID_ACK_TIMEOUT:
1271 if (optlen != sizeof(u8)) {
1272 ret = -EINVAL;
1273 break;
1274 }
1275 ret = rdma_set_ack_timeout(ctx->cm_id, *((u8 *)optval));
1276 break;
1277 default:
1278 ret = -ENOSYS;
1279 }
1280
1281 return ret;
1282 }
1283
1284 static int ucma_set_ib_path(struct ucma_context *ctx,
1285 struct ib_path_rec_data *path_data, size_t optlen)
1286 {
1287 struct sa_path_rec sa_path;
1288 struct rdma_cm_event event;
1289 int ret;
1290
1291 if (optlen % sizeof(*path_data))
1292 return -EINVAL;
1293
1294 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1295 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1296 IB_PATH_BIDIRECTIONAL))
1297 break;
1298 }
1299
1300 if (!optlen)
1301 return -EINVAL;
1302
1303 if (!ctx->cm_id->device)
1304 return -EINVAL;
1305
1306 memset(&sa_path, 0, sizeof(sa_path));
1307
1308 sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1309 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1310
1311 if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1312 struct sa_path_rec opa;
1313
1314 sa_convert_path_ib_to_opa(&opa, &sa_path);
1315 mutex_lock(&ctx->mutex);
1316 ret = rdma_set_ib_path(ctx->cm_id, &opa);
1317 mutex_unlock(&ctx->mutex);
1318 } else {
1319 mutex_lock(&ctx->mutex);
1320 ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
1321 mutex_unlock(&ctx->mutex);
1322 }
1323 if (ret)
1324 return ret;
1325
1326 memset(&event, 0, sizeof event);
1327 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1328 return ucma_event_handler(ctx->cm_id, &event);
1329 }
1330
1331 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1332 void *optval, size_t optlen)
1333 {
1334 int ret;
1335
1336 switch (optname) {
1337 case RDMA_OPTION_IB_PATH:
1338 ret = ucma_set_ib_path(ctx, optval, optlen);
1339 break;
1340 default:
1341 ret = -ENOSYS;
1342 }
1343
1344 return ret;
1345 }
1346
1347 static int ucma_set_option_level(struct ucma_context *ctx, int level,
1348 int optname, void *optval, size_t optlen)
1349 {
1350 int ret;
1351
1352 switch (level) {
1353 case RDMA_OPTION_ID:
1354 mutex_lock(&ctx->mutex);
1355 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1356 mutex_unlock(&ctx->mutex);
1357 break;
1358 case RDMA_OPTION_IB:
1359 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1360 break;
1361 default:
1362 ret = -ENOSYS;
1363 }
1364
1365 return ret;
1366 }
1367
1368 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1369 int in_len, int out_len)
1370 {
1371 struct rdma_ucm_set_option cmd;
1372 struct ucma_context *ctx;
1373 void *optval;
1374 int ret;
1375
1376 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1377 return -EFAULT;
1378
1379 if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1380 return -EINVAL;
1381
1382 ctx = ucma_get_ctx(file, cmd.id);
1383 if (IS_ERR(ctx))
1384 return PTR_ERR(ctx);
1385
1386 optval = memdup_user(u64_to_user_ptr(cmd.optval),
1387 cmd.optlen);
1388 if (IS_ERR(optval)) {
1389 ret = PTR_ERR(optval);
1390 goto out;
1391 }
1392
1393 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1394 cmd.optlen);
1395 kfree(optval);
1396
1397 out:
1398 ucma_put_ctx(ctx);
1399 return ret;
1400 }
1401
1402 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1403 int in_len, int out_len)
1404 {
1405 struct rdma_ucm_notify cmd;
1406 struct ucma_context *ctx;
1407 int ret = -EINVAL;
1408
1409 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1410 return -EFAULT;
1411
1412 ctx = ucma_get_ctx(file, cmd.id);
1413 if (IS_ERR(ctx))
1414 return PTR_ERR(ctx);
1415
1416 mutex_lock(&ctx->mutex);
1417 if (ctx->cm_id->device)
1418 ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
1419 mutex_unlock(&ctx->mutex);
1420
1421 ucma_put_ctx(ctx);
1422 return ret;
1423 }
1424
1425 static ssize_t ucma_process_join(struct ucma_file *file,
1426 struct rdma_ucm_join_mcast *cmd, int out_len)
1427 {
1428 struct rdma_ucm_create_id_resp resp;
1429 struct ucma_context *ctx;
1430 struct ucma_multicast *mc;
1431 struct sockaddr *addr;
1432 int ret;
1433 u8 join_state;
1434
1435 if (out_len < sizeof(resp))
1436 return -ENOSPC;
1437
1438 addr = (struct sockaddr *) &cmd->addr;
1439 if (cmd->addr_size != rdma_addr_size(addr))
1440 return -EINVAL;
1441
1442 if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1443 join_state = BIT(FULLMEMBER_JOIN);
1444 else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1445 join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1446 else
1447 return -EINVAL;
1448
1449 ctx = ucma_get_ctx_dev(file, cmd->id);
1450 if (IS_ERR(ctx))
1451 return PTR_ERR(ctx);
1452
1453 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
1454 if (!mc) {
1455 ret = -ENOMEM;
1456 goto err_put_ctx;
1457 }
1458
1459 mc->ctx = ctx;
1460 mc->join_state = join_state;
1461 mc->uid = cmd->uid;
1462 memcpy(&mc->addr, addr, cmd->addr_size);
1463
1464 if (xa_alloc(&multicast_table, &mc->id, NULL, xa_limit_32b,
1465 GFP_KERNEL)) {
1466 ret = -ENOMEM;
1467 goto err_free_mc;
1468 }
1469
1470 mutex_lock(&ctx->mutex);
1471 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1472 join_state, mc);
1473 mutex_unlock(&ctx->mutex);
1474 if (ret)
1475 goto err_xa_erase;
1476
1477 resp.id = mc->id;
1478 if (copy_to_user(u64_to_user_ptr(cmd->response),
1479 &resp, sizeof(resp))) {
1480 ret = -EFAULT;
1481 goto err_leave_multicast;
1482 }
1483
1484 xa_store(&multicast_table, mc->id, mc, 0);
1485
1486 ucma_put_ctx(ctx);
1487 return 0;
1488
1489 err_leave_multicast:
1490 mutex_lock(&ctx->mutex);
1491 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1492 mutex_unlock(&ctx->mutex);
1493 ucma_cleanup_mc_events(mc);
1494 err_xa_erase:
1495 xa_erase(&multicast_table, mc->id);
1496 err_free_mc:
1497 kfree(mc);
1498 err_put_ctx:
1499 ucma_put_ctx(ctx);
1500 return ret;
1501 }
1502
1503 static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1504 const char __user *inbuf,
1505 int in_len, int out_len)
1506 {
1507 struct rdma_ucm_join_ip_mcast cmd;
1508 struct rdma_ucm_join_mcast join_cmd;
1509
1510 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1511 return -EFAULT;
1512
1513 join_cmd.response = cmd.response;
1514 join_cmd.uid = cmd.uid;
1515 join_cmd.id = cmd.id;
1516 join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1517 if (!join_cmd.addr_size)
1518 return -EINVAL;
1519
1520 join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1521 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1522
1523 return ucma_process_join(file, &join_cmd, out_len);
1524 }
1525
1526 static ssize_t ucma_join_multicast(struct ucma_file *file,
1527 const char __user *inbuf,
1528 int in_len, int out_len)
1529 {
1530 struct rdma_ucm_join_mcast cmd;
1531
1532 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1533 return -EFAULT;
1534
1535 if (!rdma_addr_size_kss(&cmd.addr))
1536 return -EINVAL;
1537
1538 return ucma_process_join(file, &cmd, out_len);
1539 }
1540
1541 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1542 const char __user *inbuf,
1543 int in_len, int out_len)
1544 {
1545 struct rdma_ucm_destroy_id cmd;
1546 struct rdma_ucm_destroy_id_resp resp;
1547 struct ucma_multicast *mc;
1548 int ret = 0;
1549
1550 if (out_len < sizeof(resp))
1551 return -ENOSPC;
1552
1553 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1554 return -EFAULT;
1555
1556 xa_lock(&multicast_table);
1557 mc = xa_load(&multicast_table, cmd.id);
1558 if (!mc)
1559 mc = ERR_PTR(-ENOENT);
1560 else if (READ_ONCE(mc->ctx->file) != file)
1561 mc = ERR_PTR(-EINVAL);
1562 else if (!refcount_inc_not_zero(&mc->ctx->ref))
1563 mc = ERR_PTR(-ENXIO);
1564 else
1565 __xa_erase(&multicast_table, mc->id);
1566 xa_unlock(&multicast_table);
1567
1568 if (IS_ERR(mc)) {
1569 ret = PTR_ERR(mc);
1570 goto out;
1571 }
1572
1573 mutex_lock(&mc->ctx->mutex);
1574 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1575 mutex_unlock(&mc->ctx->mutex);
1576
1577 ucma_cleanup_mc_events(mc);
1578
1579 ucma_put_ctx(mc->ctx);
1580 resp.events_reported = mc->events_reported;
1581 kfree(mc);
1582
1583 if (copy_to_user(u64_to_user_ptr(cmd.response),
1584 &resp, sizeof(resp)))
1585 ret = -EFAULT;
1586 out:
1587 return ret;
1588 }
1589
1590 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1591 const char __user *inbuf,
1592 int in_len, int out_len)
1593 {
1594 struct rdma_ucm_migrate_id cmd;
1595 struct rdma_ucm_migrate_resp resp;
1596 struct ucma_event *uevent, *tmp;
1597 struct ucma_context *ctx;
1598 LIST_HEAD(event_list);
1599 struct fd f;
1600 struct ucma_file *cur_file;
1601 int ret = 0;
1602
1603 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1604 return -EFAULT;
1605
1606 /* Get current fd to protect against it being closed */
1607 f = fdget(cmd.fd);
1608 if (!f.file)
1609 return -ENOENT;
1610 if (f.file->f_op != &ucma_fops) {
1611 ret = -EINVAL;
1612 goto file_put;
1613 }
1614 cur_file = f.file->private_data;
1615
1616 /* Validate current fd and prevent destruction of id. */
1617 ctx = ucma_get_ctx(cur_file, cmd.id);
1618 if (IS_ERR(ctx)) {
1619 ret = PTR_ERR(ctx);
1620 goto file_put;
1621 }
1622
1623 rdma_lock_handler(ctx->cm_id);
1624 /*
1625 * ctx->file can only be changed under the handler & xa_lock. xa_load()
1626 * must be checked again to ensure the ctx hasn't begun destruction
1627 * since the ucma_get_ctx().
1628 */
1629 xa_lock(&ctx_table);
1630 if (_ucma_find_context(cmd.id, cur_file) != ctx) {
1631 xa_unlock(&ctx_table);
1632 ret = -ENOENT;
1633 goto err_unlock;
1634 }
1635 ctx->file = new_file;
1636 xa_unlock(&ctx_table);
1637
1638 mutex_lock(&cur_file->mut);
1639 list_del(&ctx->list);
1640 /*
1641 * At this point lock_handler() prevents addition of new uevents for
1642 * this ctx.
1643 */
1644 list_for_each_entry_safe(uevent, tmp, &cur_file->event_list, list)
1645 if (uevent->ctx == ctx)
1646 list_move_tail(&uevent->list, &event_list);
1647 resp.events_reported = ctx->events_reported;
1648 mutex_unlock(&cur_file->mut);
1649
1650 mutex_lock(&new_file->mut);
1651 list_add_tail(&ctx->list, &new_file->ctx_list);
1652 list_splice_tail(&event_list, &new_file->event_list);
1653 mutex_unlock(&new_file->mut);
1654
1655 if (copy_to_user(u64_to_user_ptr(cmd.response),
1656 &resp, sizeof(resp)))
1657 ret = -EFAULT;
1658
1659 err_unlock:
1660 rdma_unlock_handler(ctx->cm_id);
1661 ucma_put_ctx(ctx);
1662 file_put:
1663 fdput(f);
1664 return ret;
1665 }
1666
1667 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1668 const char __user *inbuf,
1669 int in_len, int out_len) = {
1670 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1671 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1672 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1673 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1674 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1675 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1676 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1677 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1678 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1679 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1680 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1681 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1682 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1683 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1684 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1685 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1686 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1687 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1688 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1689 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1690 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1691 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1692 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1693 };
1694
1695 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1696 size_t len, loff_t *pos)
1697 {
1698 struct ucma_file *file = filp->private_data;
1699 struct rdma_ucm_cmd_hdr hdr;
1700 ssize_t ret;
1701
1702 if (!ib_safe_file_access(filp)) {
1703 pr_err_once("ucma_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1704 task_tgid_vnr(current), current->comm);
1705 return -EACCES;
1706 }
1707
1708 if (len < sizeof(hdr))
1709 return -EINVAL;
1710
1711 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1712 return -EFAULT;
1713
1714 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1715 return -EINVAL;
1716 hdr.cmd = array_index_nospec(hdr.cmd, ARRAY_SIZE(ucma_cmd_table));
1717
1718 if (hdr.in + sizeof(hdr) > len)
1719 return -EINVAL;
1720
1721 if (!ucma_cmd_table[hdr.cmd])
1722 return -ENOSYS;
1723
1724 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1725 if (!ret)
1726 ret = len;
1727
1728 return ret;
1729 }
1730
1731 static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1732 {
1733 struct ucma_file *file = filp->private_data;
1734 __poll_t mask = 0;
1735
1736 poll_wait(filp, &file->poll_wait, wait);
1737
1738 if (!list_empty(&file->event_list))
1739 mask = EPOLLIN | EPOLLRDNORM;
1740
1741 return mask;
1742 }
1743
1744 /*
1745 * ucma_open() does not need the BKL:
1746 *
1747 * - no global state is referred to;
1748 * - there is no ioctl method to race against;
1749 * - no further module initialization is required for open to work
1750 * after the device is registered.
1751 */
1752 static int ucma_open(struct inode *inode, struct file *filp)
1753 {
1754 struct ucma_file *file;
1755
1756 file = kmalloc(sizeof *file, GFP_KERNEL);
1757 if (!file)
1758 return -ENOMEM;
1759
1760 INIT_LIST_HEAD(&file->event_list);
1761 INIT_LIST_HEAD(&file->ctx_list);
1762 init_waitqueue_head(&file->poll_wait);
1763 mutex_init(&file->mut);
1764
1765 filp->private_data = file;
1766 file->filp = filp;
1767
1768 return stream_open(inode, filp);
1769 }
1770
1771 static int ucma_close(struct inode *inode, struct file *filp)
1772 {
1773 struct ucma_file *file = filp->private_data;
1774
1775 /*
1776 * All paths that touch ctx_list or ctx_list starting from write() are
1777 * prevented by this being a FD release function. The list_add_tail() in
1778 * ucma_connect_event_handler() can run concurrently, however it only
1779 * adds to the list *after* a listening ID. By only reading the first of
1780 * the list, and relying on __destroy_id() to block
1781 * ucma_connect_event_handler(), no additional locking is needed.
1782 */
1783 while (!list_empty(&file->ctx_list)) {
1784 struct ucma_context *ctx = list_first_entry(
1785 &file->ctx_list, struct ucma_context, list);
1786
1787 xa_erase(&ctx_table, ctx->id);
1788 __destroy_id(ctx);
1789 }
1790 kfree(file);
1791 return 0;
1792 }
1793
1794 static const struct file_operations ucma_fops = {
1795 .owner = THIS_MODULE,
1796 .open = ucma_open,
1797 .release = ucma_close,
1798 .write = ucma_write,
1799 .poll = ucma_poll,
1800 .llseek = no_llseek,
1801 };
1802
1803 static struct miscdevice ucma_misc = {
1804 .minor = MISC_DYNAMIC_MINOR,
1805 .name = "rdma_cm",
1806 .nodename = "infiniband/rdma_cm",
1807 .mode = 0666,
1808 .fops = &ucma_fops,
1809 };
1810
1811 static int ucma_get_global_nl_info(struct ib_client_nl_info *res)
1812 {
1813 res->abi = RDMA_USER_CM_ABI_VERSION;
1814 res->cdev = ucma_misc.this_device;
1815 return 0;
1816 }
1817
1818 static struct ib_client rdma_cma_client = {
1819 .name = "rdma_cm",
1820 .get_global_nl_info = ucma_get_global_nl_info,
1821 };
1822 MODULE_ALIAS_RDMA_CLIENT("rdma_cm");
1823
1824 static ssize_t show_abi_version(struct device *dev,
1825 struct device_attribute *attr,
1826 char *buf)
1827 {
1828 return sysfs_emit(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1829 }
1830 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1831
1832 static int __init ucma_init(void)
1833 {
1834 int ret;
1835
1836 ret = misc_register(&ucma_misc);
1837 if (ret)
1838 return ret;
1839
1840 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1841 if (ret) {
1842 pr_err("rdma_ucm: couldn't create abi_version attr\n");
1843 goto err1;
1844 }
1845
1846 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1847 if (!ucma_ctl_table_hdr) {
1848 pr_err("rdma_ucm: couldn't register sysctl paths\n");
1849 ret = -ENOMEM;
1850 goto err2;
1851 }
1852
1853 ret = ib_register_client(&rdma_cma_client);
1854 if (ret)
1855 goto err3;
1856
1857 return 0;
1858 err3:
1859 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1860 err2:
1861 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1862 err1:
1863 misc_deregister(&ucma_misc);
1864 return ret;
1865 }
1866
1867 static void __exit ucma_cleanup(void)
1868 {
1869 ib_unregister_client(&rdma_cma_client);
1870 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1871 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1872 misc_deregister(&ucma_misc);
1873 }
1874
1875 module_init(ucma_init);
1876 module_exit(ucma_cleanup);
Linux with added FPC-III support