?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
locking/refcounts: Include fewer headers in <linux/refcount.h>
[linux/fpc-iii.git] / drivers / infiniband / core / ucma.c
blobec8fb289621fb7590dd3e3f4000967fa2b6c9aae
1 /*
2 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33 #include <linux/completion.h>
34 #include <linux/file.h>
35 #include <linux/mutex.h>
36 #include <linux/poll.h>
37 #include <linux/sched.h>
38 #include <linux/idr.h>
39 #include <linux/in.h>
40 #include <linux/in6.h>
41 #include <linux/miscdevice.h>
42 #include <linux/slab.h>
43 #include <linux/sysctl.h>
44 #include <linux/module.h>
45 #include <linux/nsproxy.h>
46
47 #include <rdma/rdma_user_cm.h>
48 #include <rdma/ib_marshall.h>
49 #include <rdma/rdma_cm.h>
50 #include <rdma/rdma_cm_ib.h>
51 #include <rdma/ib_addr.h>
52 #include <rdma/ib.h>
53
54 MODULE_AUTHOR("Sean Hefty");
55 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
56 MODULE_LICENSE("Dual BSD/GPL");
57
58 static unsigned int max_backlog = 1024;
59
60 static struct ctl_table_header *ucma_ctl_table_hdr;
61 static struct ctl_table ucma_ctl_table[] = {
62 {
63 .procname = "max_backlog",
64 .data = &max_backlog,
65 .maxlen = sizeof max_backlog,
66 .mode = 0644,
67 .proc_handler = proc_dointvec,
68 },
69 { }
70 };
71
72 struct ucma_file {
73 struct mutex mut;
74 struct file *filp;
75 struct list_head ctx_list;
76 struct list_head event_list;
77 wait_queue_head_t poll_wait;
78 struct workqueue_struct *close_wq;
79 };
80
81 struct ucma_context {
82 int id;
83 struct completion comp;
84 atomic_t ref;
85 int events_reported;
86 int backlog;
87
88 struct ucma_file *file;
89 struct rdma_cm_id *cm_id;
90 u64 uid;
91
92 struct list_head list;
93 struct list_head mc_list;
94 /* mark that device is in process of destroying the internal HW
95 * resources, protected by the global mut
96 */
97 int closing;
98 /* sync between removal event and id destroy, protected by file mut */
99 int destroying;
100 struct work_struct close_work;
101 };
102
103 struct ucma_multicast {
104 struct ucma_context *ctx;
105 int id;
106 int events_reported;
107
108 u64 uid;
109 u8 join_state;
110 struct list_head list;
111 struct sockaddr_storage addr;
112 };
113
114 struct ucma_event {
115 struct ucma_context *ctx;
116 struct ucma_multicast *mc;
117 struct list_head list;
118 struct rdma_cm_id *cm_id;
119 struct rdma_ucm_event_resp resp;
120 struct work_struct close_work;
121 };
122
123 static DEFINE_MUTEX(mut);
124 static DEFINE_IDR(ctx_idr);
125 static DEFINE_IDR(multicast_idr);
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 = idr_find(&ctx_idr, id);
133 if (!ctx)
134 ctx = ERR_PTR(-ENOENT);
135 else if (ctx->file != file || !ctx->cm_id)
136 ctx = ERR_PTR(-EINVAL);
137 return ctx;
138 }
139
140 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
141 {
142 struct ucma_context *ctx;
143
144 mutex_lock(&mut);
145 ctx = _ucma_find_context(id, file);
146 if (!IS_ERR(ctx)) {
147 if (ctx->closing)
148 ctx = ERR_PTR(-EIO);
149 else
150 atomic_inc(&ctx->ref);
151 }
152 mutex_unlock(&mut);
153 return ctx;
154 }
155
156 static void ucma_put_ctx(struct ucma_context *ctx)
157 {
158 if (atomic_dec_and_test(&ctx->ref))
159 complete(&ctx->comp);
160 }
161
162 /*
163 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
164 * CM_ID is bound.
165 */
166 static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
167 {
168 struct ucma_context *ctx = ucma_get_ctx(file, id);
169
170 if (IS_ERR(ctx))
171 return ctx;
172 if (!ctx->cm_id->device) {
173 ucma_put_ctx(ctx);
174 return ERR_PTR(-EINVAL);
175 }
176 return ctx;
177 }
178
179 static void ucma_close_event_id(struct work_struct *work)
180 {
181 struct ucma_event *uevent_close = container_of(work, struct ucma_event, close_work);
182
183 rdma_destroy_id(uevent_close->cm_id);
184 kfree(uevent_close);
185 }
186
187 static void ucma_close_id(struct work_struct *work)
188 {
189 struct ucma_context *ctx = container_of(work, struct ucma_context, close_work);
190
191 /* once all inflight tasks are finished, we close all underlying
192 * resources. The context is still alive till its explicit destryoing
193 * by its creator.
194 */
195 ucma_put_ctx(ctx);
196 wait_for_completion(&ctx->comp);
197 /* No new events will be generated after destroying the id. */
198 rdma_destroy_id(ctx->cm_id);
199 }
200
201 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
202 {
203 struct ucma_context *ctx;
204
205 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
206 if (!ctx)
207 return NULL;
208
209 INIT_WORK(&ctx->close_work, ucma_close_id);
210 atomic_set(&ctx->ref, 1);
211 init_completion(&ctx->comp);
212 INIT_LIST_HEAD(&ctx->mc_list);
213 ctx->file = file;
214
215 mutex_lock(&mut);
216 ctx->id = idr_alloc(&ctx_idr, ctx, 0, 0, GFP_KERNEL);
217 mutex_unlock(&mut);
218 if (ctx->id < 0)
219 goto error;
220
221 list_add_tail(&ctx->list, &file->ctx_list);
222 return ctx;
223
224 error:
225 kfree(ctx);
226 return NULL;
227 }
228
229 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
230 {
231 struct ucma_multicast *mc;
232
233 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
234 if (!mc)
235 return NULL;
236
237 mutex_lock(&mut);
238 mc->id = idr_alloc(&multicast_idr, NULL, 0, 0, GFP_KERNEL);
239 mutex_unlock(&mut);
240 if (mc->id < 0)
241 goto error;
242
243 mc->ctx = ctx;
244 list_add_tail(&mc->list, &ctx->mc_list);
245 return mc;
246
247 error:
248 kfree(mc);
249 return NULL;
250 }
251
252 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
253 struct rdma_conn_param *src)
254 {
255 if (src->private_data_len)
256 memcpy(dst->private_data, src->private_data,
257 src->private_data_len);
258 dst->private_data_len = src->private_data_len;
259 dst->responder_resources =src->responder_resources;
260 dst->initiator_depth = src->initiator_depth;
261 dst->flow_control = src->flow_control;
262 dst->retry_count = src->retry_count;
263 dst->rnr_retry_count = src->rnr_retry_count;
264 dst->srq = src->srq;
265 dst->qp_num = src->qp_num;
266 }
267
268 static void ucma_copy_ud_event(struct ib_device *device,
269 struct rdma_ucm_ud_param *dst,
270 struct rdma_ud_param *src)
271 {
272 if (src->private_data_len)
273 memcpy(dst->private_data, src->private_data,
274 src->private_data_len);
275 dst->private_data_len = src->private_data_len;
276 ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
277 dst->qp_num = src->qp_num;
278 dst->qkey = src->qkey;
279 }
280
281 static void ucma_set_event_context(struct ucma_context *ctx,
282 struct rdma_cm_event *event,
283 struct ucma_event *uevent)
284 {
285 uevent->ctx = ctx;
286 switch (event->event) {
287 case RDMA_CM_EVENT_MULTICAST_JOIN:
288 case RDMA_CM_EVENT_MULTICAST_ERROR:
289 uevent->mc = (struct ucma_multicast *)
290 event->param.ud.private_data;
291 uevent->resp.uid = uevent->mc->uid;
292 uevent->resp.id = uevent->mc->id;
293 break;
294 default:
295 uevent->resp.uid = ctx->uid;
296 uevent->resp.id = ctx->id;
297 break;
298 }
299 }
300
301 /* Called with file->mut locked for the relevant context. */
302 static void ucma_removal_event_handler(struct rdma_cm_id *cm_id)
303 {
304 struct ucma_context *ctx = cm_id->context;
305 struct ucma_event *con_req_eve;
306 int event_found = 0;
307
308 if (ctx->destroying)
309 return;
310
311 /* only if context is pointing to cm_id that it owns it and can be
312 * queued to be closed, otherwise that cm_id is an inflight one that
313 * is part of that context event list pending to be detached and
314 * reattached to its new context as part of ucma_get_event,
315 * handled separately below.
316 */
317 if (ctx->cm_id == cm_id) {
318 mutex_lock(&mut);
319 ctx->closing = 1;
320 mutex_unlock(&mut);
321 queue_work(ctx->file->close_wq, &ctx->close_work);
322 return;
323 }
324
325 list_for_each_entry(con_req_eve, &ctx->file->event_list, list) {
326 if (con_req_eve->cm_id == cm_id &&
327 con_req_eve->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
328 list_del(&con_req_eve->list);
329 INIT_WORK(&con_req_eve->close_work, ucma_close_event_id);
330 queue_work(ctx->file->close_wq, &con_req_eve->close_work);
331 event_found = 1;
332 break;
333 }
334 }
335 if (!event_found)
336 pr_err("ucma_removal_event_handler: warning: connect request event wasn't found\n");
337 }
338
339 static int ucma_event_handler(struct rdma_cm_id *cm_id,
340 struct rdma_cm_event *event)
341 {
342 struct ucma_event *uevent;
343 struct ucma_context *ctx = cm_id->context;
344 int ret = 0;
345
346 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
347 if (!uevent)
348 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
349
350 mutex_lock(&ctx->file->mut);
351 uevent->cm_id = cm_id;
352 ucma_set_event_context(ctx, event, uevent);
353 uevent->resp.event = event->event;
354 uevent->resp.status = event->status;
355 if (cm_id->qp_type == IB_QPT_UD)
356 ucma_copy_ud_event(cm_id->device, &uevent->resp.param.ud,
357 &event->param.ud);
358 else
359 ucma_copy_conn_event(&uevent->resp.param.conn,
360 &event->param.conn);
361
362 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
363 if (!ctx->backlog) {
364 ret = -ENOMEM;
365 kfree(uevent);
366 goto out;
367 }
368 ctx->backlog--;
369 } else if (!ctx->uid || ctx->cm_id != cm_id) {
370 /*
371 * We ignore events for new connections until userspace has set
372 * their context. This can only happen if an error occurs on a
373 * new connection before the user accepts it. This is okay,
374 * since the accept will just fail later. However, we do need
375 * to release the underlying HW resources in case of a device
376 * removal event.
377 */
378 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
379 ucma_removal_event_handler(cm_id);
380
381 kfree(uevent);
382 goto out;
383 }
384
385 list_add_tail(&uevent->list, &ctx->file->event_list);
386 wake_up_interruptible(&ctx->file->poll_wait);
387 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
388 ucma_removal_event_handler(cm_id);
389 out:
390 mutex_unlock(&ctx->file->mut);
391 return ret;
392 }
393
394 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
395 int in_len, int out_len)
396 {
397 struct ucma_context *ctx;
398 struct rdma_ucm_get_event cmd;
399 struct ucma_event *uevent;
400 int ret = 0;
401
402 /*
403 * Old 32 bit user space does not send the 4 byte padding in the
404 * reserved field. We don't care, allow it to keep working.
405 */
406 if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved))
407 return -ENOSPC;
408
409 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
410 return -EFAULT;
411
412 mutex_lock(&file->mut);
413 while (list_empty(&file->event_list)) {
414 mutex_unlock(&file->mut);
415
416 if (file->filp->f_flags & O_NONBLOCK)
417 return -EAGAIN;
418
419 if (wait_event_interruptible(file->poll_wait,
420 !list_empty(&file->event_list)))
421 return -ERESTARTSYS;
422
423 mutex_lock(&file->mut);
424 }
425
426 uevent = list_entry(file->event_list.next, struct ucma_event, list);
427
428 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
429 ctx = ucma_alloc_ctx(file);
430 if (!ctx) {
431 ret = -ENOMEM;
432 goto done;
433 }
434 uevent->ctx->backlog++;
435 ctx->cm_id = uevent->cm_id;
436 ctx->cm_id->context = ctx;
437 uevent->resp.id = ctx->id;
438 }
439
440 if (copy_to_user(u64_to_user_ptr(cmd.response),
441 &uevent->resp,
442 min_t(size_t, out_len, sizeof(uevent->resp)))) {
443 ret = -EFAULT;
444 goto done;
445 }
446
447 list_del(&uevent->list);
448 uevent->ctx->events_reported++;
449 if (uevent->mc)
450 uevent->mc->events_reported++;
451 kfree(uevent);
452 done:
453 mutex_unlock(&file->mut);
454 return ret;
455 }
456
457 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
458 {
459 switch (cmd->ps) {
460 case RDMA_PS_TCP:
461 *qp_type = IB_QPT_RC;
462 return 0;
463 case RDMA_PS_UDP:
464 case RDMA_PS_IPOIB:
465 *qp_type = IB_QPT_UD;
466 return 0;
467 case RDMA_PS_IB:
468 *qp_type = cmd->qp_type;
469 return 0;
470 default:
471 return -EINVAL;
472 }
473 }
474
475 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
476 int in_len, int out_len)
477 {
478 struct rdma_ucm_create_id cmd;
479 struct rdma_ucm_create_id_resp resp;
480 struct ucma_context *ctx;
481 struct rdma_cm_id *cm_id;
482 enum ib_qp_type qp_type;
483 int ret;
484
485 if (out_len < sizeof(resp))
486 return -ENOSPC;
487
488 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
489 return -EFAULT;
490
491 ret = ucma_get_qp_type(&cmd, &qp_type);
492 if (ret)
493 return ret;
494
495 mutex_lock(&file->mut);
496 ctx = ucma_alloc_ctx(file);
497 mutex_unlock(&file->mut);
498 if (!ctx)
499 return -ENOMEM;
500
501 ctx->uid = cmd.uid;
502 cm_id = __rdma_create_id(current->nsproxy->net_ns,
503 ucma_event_handler, ctx, cmd.ps, qp_type, NULL);
504 if (IS_ERR(cm_id)) {
505 ret = PTR_ERR(cm_id);
506 goto err1;
507 }
508
509 resp.id = ctx->id;
510 if (copy_to_user(u64_to_user_ptr(cmd.response),
511 &resp, sizeof(resp))) {
512 ret = -EFAULT;
513 goto err2;
514 }
515
516 ctx->cm_id = cm_id;
517 return 0;
518
519 err2:
520 rdma_destroy_id(cm_id);
521 err1:
522 mutex_lock(&mut);
523 idr_remove(&ctx_idr, ctx->id);
524 mutex_unlock(&mut);
525 mutex_lock(&file->mut);
526 list_del(&ctx->list);
527 mutex_unlock(&file->mut);
528 kfree(ctx);
529 return ret;
530 }
531
532 static void ucma_cleanup_multicast(struct ucma_context *ctx)
533 {
534 struct ucma_multicast *mc, *tmp;
535
536 mutex_lock(&mut);
537 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
538 list_del(&mc->list);
539 idr_remove(&multicast_idr, mc->id);
540 kfree(mc);
541 }
542 mutex_unlock(&mut);
543 }
544
545 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
546 {
547 struct ucma_event *uevent, *tmp;
548
549 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
550 if (uevent->mc != mc)
551 continue;
552
553 list_del(&uevent->list);
554 kfree(uevent);
555 }
556 }
557
558 /*
559 * ucma_free_ctx is called after the underlying rdma CM-ID is destroyed. At
560 * this point, no new events will be reported from the hardware. However, we
561 * still need to cleanup the UCMA context for this ID. Specifically, there
562 * might be events that have not yet been consumed by the user space software.
563 * These might include pending connect requests which we have not completed
564 * processing. We cannot call rdma_destroy_id while holding the lock of the
565 * context (file->mut), as it might cause a deadlock. We therefore extract all
566 * relevant events from the context pending events list while holding the
567 * mutex. After that we release them as needed.
568 */
569 static int ucma_free_ctx(struct ucma_context *ctx)
570 {
571 int events_reported;
572 struct ucma_event *uevent, *tmp;
573 LIST_HEAD(list);
574
575
576 ucma_cleanup_multicast(ctx);
577
578 /* Cleanup events not yet reported to the user. */
579 mutex_lock(&ctx->file->mut);
580 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
581 if (uevent->ctx == ctx)
582 list_move_tail(&uevent->list, &list);
583 }
584 list_del(&ctx->list);
585 mutex_unlock(&ctx->file->mut);
586
587 list_for_each_entry_safe(uevent, tmp, &list, list) {
588 list_del(&uevent->list);
589 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
590 rdma_destroy_id(uevent->cm_id);
591 kfree(uevent);
592 }
593
594 events_reported = ctx->events_reported;
595 kfree(ctx);
596 return events_reported;
597 }
598
599 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
600 int in_len, int out_len)
601 {
602 struct rdma_ucm_destroy_id cmd;
603 struct rdma_ucm_destroy_id_resp resp;
604 struct ucma_context *ctx;
605 int ret = 0;
606
607 if (out_len < sizeof(resp))
608 return -ENOSPC;
609
610 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
611 return -EFAULT;
612
613 mutex_lock(&mut);
614 ctx = _ucma_find_context(cmd.id, file);
615 if (!IS_ERR(ctx))
616 idr_remove(&ctx_idr, ctx->id);
617 mutex_unlock(&mut);
618
619 if (IS_ERR(ctx))
620 return PTR_ERR(ctx);
621
622 mutex_lock(&ctx->file->mut);
623 ctx->destroying = 1;
624 mutex_unlock(&ctx->file->mut);
625
626 flush_workqueue(ctx->file->close_wq);
627 /* At this point it's guaranteed that there is no inflight
628 * closing task */
629 mutex_lock(&mut);
630 if (!ctx->closing) {
631 mutex_unlock(&mut);
632 ucma_put_ctx(ctx);
633 wait_for_completion(&ctx->comp);
634 rdma_destroy_id(ctx->cm_id);
635 } else {
636 mutex_unlock(&mut);
637 }
638
639 resp.events_reported = ucma_free_ctx(ctx);
640 if (copy_to_user(u64_to_user_ptr(cmd.response),
641 &resp, sizeof(resp)))
642 ret = -EFAULT;
643
644 return ret;
645 }
646
647 static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
648 int in_len, int out_len)
649 {
650 struct rdma_ucm_bind_ip cmd;
651 struct ucma_context *ctx;
652 int ret;
653
654 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
655 return -EFAULT;
656
657 if (!rdma_addr_size_in6(&cmd.addr))
658 return -EINVAL;
659
660 ctx = ucma_get_ctx(file, cmd.id);
661 if (IS_ERR(ctx))
662 return PTR_ERR(ctx);
663
664 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
665 ucma_put_ctx(ctx);
666 return ret;
667 }
668
669 static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
670 int in_len, int out_len)
671 {
672 struct rdma_ucm_bind cmd;
673 struct ucma_context *ctx;
674 int ret;
675
676 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
677 return -EFAULT;
678
679 if (cmd.reserved || !cmd.addr_size ||
680 cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
681 return -EINVAL;
682
683 ctx = ucma_get_ctx(file, cmd.id);
684 if (IS_ERR(ctx))
685 return PTR_ERR(ctx);
686
687 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
688 ucma_put_ctx(ctx);
689 return ret;
690 }
691
692 static ssize_t ucma_resolve_ip(struct ucma_file *file,
693 const char __user *inbuf,
694 int in_len, int out_len)
695 {
696 struct rdma_ucm_resolve_ip cmd;
697 struct ucma_context *ctx;
698 int ret;
699
700 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
701 return -EFAULT;
702
703 if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
704 !rdma_addr_size_in6(&cmd.dst_addr))
705 return -EINVAL;
706
707 ctx = ucma_get_ctx(file, cmd.id);
708 if (IS_ERR(ctx))
709 return PTR_ERR(ctx);
710
711 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
712 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
713 ucma_put_ctx(ctx);
714 return ret;
715 }
716
717 static ssize_t ucma_resolve_addr(struct ucma_file *file,
718 const char __user *inbuf,
719 int in_len, int out_len)
720 {
721 struct rdma_ucm_resolve_addr cmd;
722 struct ucma_context *ctx;
723 int ret;
724
725 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
726 return -EFAULT;
727
728 if (cmd.reserved ||
729 (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
730 !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
731 return -EINVAL;
732
733 ctx = ucma_get_ctx(file, cmd.id);
734 if (IS_ERR(ctx))
735 return PTR_ERR(ctx);
736
737 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
738 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
739 ucma_put_ctx(ctx);
740 return ret;
741 }
742
743 static ssize_t ucma_resolve_route(struct ucma_file *file,
744 const char __user *inbuf,
745 int in_len, int out_len)
746 {
747 struct rdma_ucm_resolve_route cmd;
748 struct ucma_context *ctx;
749 int ret;
750
751 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
752 return -EFAULT;
753
754 ctx = ucma_get_ctx_dev(file, cmd.id);
755 if (IS_ERR(ctx))
756 return PTR_ERR(ctx);
757
758 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
759 ucma_put_ctx(ctx);
760 return ret;
761 }
762
763 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
764 struct rdma_route *route)
765 {
766 struct rdma_dev_addr *dev_addr;
767
768 resp->num_paths = route->num_paths;
769 switch (route->num_paths) {
770 case 0:
771 dev_addr = &route->addr.dev_addr;
772 rdma_addr_get_dgid(dev_addr,
773 (union ib_gid *) &resp->ib_route[0].dgid);
774 rdma_addr_get_sgid(dev_addr,
775 (union ib_gid *) &resp->ib_route[0].sgid);
776 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
777 break;
778 case 2:
779 ib_copy_path_rec_to_user(&resp->ib_route[1],
780 &route->path_rec[1]);
781 /* fall through */
782 case 1:
783 ib_copy_path_rec_to_user(&resp->ib_route[0],
784 &route->path_rec[0]);
785 break;
786 default:
787 break;
788 }
789 }
790
791 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
792 struct rdma_route *route)
793 {
794
795 resp->num_paths = route->num_paths;
796 switch (route->num_paths) {
797 case 0:
798 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
799 (union ib_gid *)&resp->ib_route[0].dgid);
800 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
801 (union ib_gid *)&resp->ib_route[0].sgid);
802 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
803 break;
804 case 2:
805 ib_copy_path_rec_to_user(&resp->ib_route[1],
806 &route->path_rec[1]);
807 /* fall through */
808 case 1:
809 ib_copy_path_rec_to_user(&resp->ib_route[0],
810 &route->path_rec[0]);
811 break;
812 default:
813 break;
814 }
815 }
816
817 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
818 struct rdma_route *route)
819 {
820 struct rdma_dev_addr *dev_addr;
821
822 dev_addr = &route->addr.dev_addr;
823 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
824 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
825 }
826
827 static ssize_t ucma_query_route(struct ucma_file *file,
828 const char __user *inbuf,
829 int in_len, int out_len)
830 {
831 struct rdma_ucm_query cmd;
832 struct rdma_ucm_query_route_resp resp;
833 struct ucma_context *ctx;
834 struct sockaddr *addr;
835 int ret = 0;
836
837 if (out_len < sizeof(resp))
838 return -ENOSPC;
839
840 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
841 return -EFAULT;
842
843 ctx = ucma_get_ctx(file, cmd.id);
844 if (IS_ERR(ctx))
845 return PTR_ERR(ctx);
846
847 memset(&resp, 0, sizeof resp);
848 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
849 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
850 sizeof(struct sockaddr_in) :
851 sizeof(struct sockaddr_in6));
852 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
853 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
854 sizeof(struct sockaddr_in) :
855 sizeof(struct sockaddr_in6));
856 if (!ctx->cm_id->device)
857 goto out;
858
859 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
860 resp.port_num = ctx->cm_id->port_num;
861
862 if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
863 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
864 else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
865 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
866 else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
867 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
868
869 out:
870 if (copy_to_user(u64_to_user_ptr(cmd.response),
871 &resp, sizeof(resp)))
872 ret = -EFAULT;
873
874 ucma_put_ctx(ctx);
875 return ret;
876 }
877
878 static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
879 struct rdma_ucm_query_addr_resp *resp)
880 {
881 if (!cm_id->device)
882 return;
883
884 resp->node_guid = (__force __u64) cm_id->device->node_guid;
885 resp->port_num = cm_id->port_num;
886 resp->pkey = (__force __u16) cpu_to_be16(
887 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
888 }
889
890 static ssize_t ucma_query_addr(struct ucma_context *ctx,
891 void __user *response, int out_len)
892 {
893 struct rdma_ucm_query_addr_resp resp;
894 struct sockaddr *addr;
895 int ret = 0;
896
897 if (out_len < sizeof(resp))
898 return -ENOSPC;
899
900 memset(&resp, 0, sizeof resp);
901
902 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
903 resp.src_size = rdma_addr_size(addr);
904 memcpy(&resp.src_addr, addr, resp.src_size);
905
906 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
907 resp.dst_size = rdma_addr_size(addr);
908 memcpy(&resp.dst_addr, addr, resp.dst_size);
909
910 ucma_query_device_addr(ctx->cm_id, &resp);
911
912 if (copy_to_user(response, &resp, sizeof(resp)))
913 ret = -EFAULT;
914
915 return ret;
916 }
917
918 static ssize_t ucma_query_path(struct ucma_context *ctx,
919 void __user *response, int out_len)
920 {
921 struct rdma_ucm_query_path_resp *resp;
922 int i, ret = 0;
923
924 if (out_len < sizeof(*resp))
925 return -ENOSPC;
926
927 resp = kzalloc(out_len, GFP_KERNEL);
928 if (!resp)
929 return -ENOMEM;
930
931 resp->num_paths = ctx->cm_id->route.num_paths;
932 for (i = 0, out_len -= sizeof(*resp);
933 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
934 i++, out_len -= sizeof(struct ib_path_rec_data)) {
935 struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
936
937 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
938 IB_PATH_BIDIRECTIONAL;
939 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
940 struct sa_path_rec ib;
941
942 sa_convert_path_opa_to_ib(&ib, rec);
943 ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
944
945 } else {
946 ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
947 }
948 }
949
950 if (copy_to_user(response, resp,
951 sizeof(*resp) + (i * sizeof(struct ib_path_rec_data))))
952 ret = -EFAULT;
953
954 kfree(resp);
955 return ret;
956 }
957
958 static ssize_t ucma_query_gid(struct ucma_context *ctx,
959 void __user *response, int out_len)
960 {
961 struct rdma_ucm_query_addr_resp resp;
962 struct sockaddr_ib *addr;
963 int ret = 0;
964
965 if (out_len < sizeof(resp))
966 return -ENOSPC;
967
968 memset(&resp, 0, sizeof resp);
969
970 ucma_query_device_addr(ctx->cm_id, &resp);
971
972 addr = (struct sockaddr_ib *) &resp.src_addr;
973 resp.src_size = sizeof(*addr);
974 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
975 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
976 } else {
977 addr->sib_family = AF_IB;
978 addr->sib_pkey = (__force __be16) resp.pkey;
979 rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
980 NULL);
981 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
982 &ctx->cm_id->route.addr.src_addr);
983 }
984
985 addr = (struct sockaddr_ib *) &resp.dst_addr;
986 resp.dst_size = sizeof(*addr);
987 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
988 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
989 } else {
990 addr->sib_family = AF_IB;
991 addr->sib_pkey = (__force __be16) resp.pkey;
992 rdma_read_gids(ctx->cm_id, NULL,
993 (union ib_gid *)&addr->sib_addr);
994 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
995 &ctx->cm_id->route.addr.dst_addr);
996 }
997
998 if (copy_to_user(response, &resp, sizeof(resp)))
999 ret = -EFAULT;
1000
1001 return ret;
1002 }
1003
1004 static ssize_t ucma_query(struct ucma_file *file,
1005 const char __user *inbuf,
1006 int in_len, int out_len)
1007 {
1008 struct rdma_ucm_query cmd;
1009 struct ucma_context *ctx;
1010 void __user *response;
1011 int ret;
1012
1013 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1014 return -EFAULT;
1015
1016 response = u64_to_user_ptr(cmd.response);
1017 ctx = ucma_get_ctx(file, cmd.id);
1018 if (IS_ERR(ctx))
1019 return PTR_ERR(ctx);
1020
1021 switch (cmd.option) {
1022 case RDMA_USER_CM_QUERY_ADDR:
1023 ret = ucma_query_addr(ctx, response, out_len);
1024 break;
1025 case RDMA_USER_CM_QUERY_PATH:
1026 ret = ucma_query_path(ctx, response, out_len);
1027 break;
1028 case RDMA_USER_CM_QUERY_GID:
1029 ret = ucma_query_gid(ctx, response, out_len);
1030 break;
1031 default:
1032 ret = -ENOSYS;
1033 break;
1034 }
1035
1036 ucma_put_ctx(ctx);
1037 return ret;
1038 }
1039
1040 static void ucma_copy_conn_param(struct rdma_cm_id *id,
1041 struct rdma_conn_param *dst,
1042 struct rdma_ucm_conn_param *src)
1043 {
1044 dst->private_data = src->private_data;
1045 dst->private_data_len = src->private_data_len;
1046 dst->responder_resources =src->responder_resources;
1047 dst->initiator_depth = src->initiator_depth;
1048 dst->flow_control = src->flow_control;
1049 dst->retry_count = src->retry_count;
1050 dst->rnr_retry_count = src->rnr_retry_count;
1051 dst->srq = src->srq;
1052 dst->qp_num = src->qp_num;
1053 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1054 }
1055
1056 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1057 int in_len, int out_len)
1058 {
1059 struct rdma_ucm_connect cmd;
1060 struct rdma_conn_param conn_param;
1061 struct ucma_context *ctx;
1062 int ret;
1063
1064 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1065 return -EFAULT;
1066
1067 if (!cmd.conn_param.valid)
1068 return -EINVAL;
1069
1070 ctx = ucma_get_ctx_dev(file, cmd.id);
1071 if (IS_ERR(ctx))
1072 return PTR_ERR(ctx);
1073
1074 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1075 ret = rdma_connect(ctx->cm_id, &conn_param);
1076 ucma_put_ctx(ctx);
1077 return ret;
1078 }
1079
1080 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1081 int in_len, int out_len)
1082 {
1083 struct rdma_ucm_listen cmd;
1084 struct ucma_context *ctx;
1085 int ret;
1086
1087 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1088 return -EFAULT;
1089
1090 ctx = ucma_get_ctx(file, cmd.id);
1091 if (IS_ERR(ctx))
1092 return PTR_ERR(ctx);
1093
1094 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
1095 cmd.backlog : max_backlog;
1096 ret = rdma_listen(ctx->cm_id, ctx->backlog);
1097 ucma_put_ctx(ctx);
1098 return ret;
1099 }
1100
1101 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1102 int in_len, int out_len)
1103 {
1104 struct rdma_ucm_accept cmd;
1105 struct rdma_conn_param conn_param;
1106 struct ucma_context *ctx;
1107 int ret;
1108
1109 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1110 return -EFAULT;
1111
1112 ctx = ucma_get_ctx_dev(file, cmd.id);
1113 if (IS_ERR(ctx))
1114 return PTR_ERR(ctx);
1115
1116 if (cmd.conn_param.valid) {
1117 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1118 mutex_lock(&file->mut);
1119 ret = __rdma_accept(ctx->cm_id, &conn_param, NULL);
1120 if (!ret)
1121 ctx->uid = cmd.uid;
1122 mutex_unlock(&file->mut);
1123 } else
1124 ret = __rdma_accept(ctx->cm_id, NULL, NULL);
1125
1126 ucma_put_ctx(ctx);
1127 return ret;
1128 }
1129
1130 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1131 int in_len, int out_len)
1132 {
1133 struct rdma_ucm_reject cmd;
1134 struct ucma_context *ctx;
1135 int ret;
1136
1137 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1138 return -EFAULT;
1139
1140 ctx = ucma_get_ctx_dev(file, cmd.id);
1141 if (IS_ERR(ctx))
1142 return PTR_ERR(ctx);
1143
1144 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
1145 ucma_put_ctx(ctx);
1146 return ret;
1147 }
1148
1149 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1150 int in_len, int out_len)
1151 {
1152 struct rdma_ucm_disconnect cmd;
1153 struct ucma_context *ctx;
1154 int ret;
1155
1156 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1157 return -EFAULT;
1158
1159 ctx = ucma_get_ctx_dev(file, cmd.id);
1160 if (IS_ERR(ctx))
1161 return PTR_ERR(ctx);
1162
1163 ret = rdma_disconnect(ctx->cm_id);
1164 ucma_put_ctx(ctx);
1165 return ret;
1166 }
1167
1168 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1169 const char __user *inbuf,
1170 int in_len, int out_len)
1171 {
1172 struct rdma_ucm_init_qp_attr cmd;
1173 struct ib_uverbs_qp_attr resp;
1174 struct ucma_context *ctx;
1175 struct ib_qp_attr qp_attr;
1176 int ret;
1177
1178 if (out_len < sizeof(resp))
1179 return -ENOSPC;
1180
1181 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1182 return -EFAULT;
1183
1184 if (cmd.qp_state > IB_QPS_ERR)
1185 return -EINVAL;
1186
1187 ctx = ucma_get_ctx_dev(file, cmd.id);
1188 if (IS_ERR(ctx))
1189 return PTR_ERR(ctx);
1190
1191 resp.qp_attr_mask = 0;
1192 memset(&qp_attr, 0, sizeof qp_attr);
1193 qp_attr.qp_state = cmd.qp_state;
1194 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1195 if (ret)
1196 goto out;
1197
1198 ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1199 if (copy_to_user(u64_to_user_ptr(cmd.response),
1200 &resp, sizeof(resp)))
1201 ret = -EFAULT;
1202
1203 out:
1204 ucma_put_ctx(ctx);
1205 return ret;
1206 }
1207
1208 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1209 void *optval, size_t optlen)
1210 {
1211 int ret = 0;
1212
1213 switch (optname) {
1214 case RDMA_OPTION_ID_TOS:
1215 if (optlen != sizeof(u8)) {
1216 ret = -EINVAL;
1217 break;
1218 }
1219 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1220 break;
1221 case RDMA_OPTION_ID_REUSEADDR:
1222 if (optlen != sizeof(int)) {
1223 ret = -EINVAL;
1224 break;
1225 }
1226 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1227 break;
1228 case RDMA_OPTION_ID_AFONLY:
1229 if (optlen != sizeof(int)) {
1230 ret = -EINVAL;
1231 break;
1232 }
1233 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1234 break;
1235 default:
1236 ret = -ENOSYS;
1237 }
1238
1239 return ret;
1240 }
1241
1242 static int ucma_set_ib_path(struct ucma_context *ctx,
1243 struct ib_path_rec_data *path_data, size_t optlen)
1244 {
1245 struct sa_path_rec sa_path;
1246 struct rdma_cm_event event;
1247 int ret;
1248
1249 if (optlen % sizeof(*path_data))
1250 return -EINVAL;
1251
1252 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1253 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1254 IB_PATH_BIDIRECTIONAL))
1255 break;
1256 }
1257
1258 if (!optlen)
1259 return -EINVAL;
1260
1261 if (!ctx->cm_id->device)
1262 return -EINVAL;
1263
1264 memset(&sa_path, 0, sizeof(sa_path));
1265
1266 sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1267 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1268
1269 if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1270 struct sa_path_rec opa;
1271
1272 sa_convert_path_ib_to_opa(&opa, &sa_path);
1273 ret = rdma_set_ib_path(ctx->cm_id, &opa);
1274 } else {
1275 ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
1276 }
1277 if (ret)
1278 return ret;
1279
1280 memset(&event, 0, sizeof event);
1281 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1282 return ucma_event_handler(ctx->cm_id, &event);
1283 }
1284
1285 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1286 void *optval, size_t optlen)
1287 {
1288 int ret;
1289
1290 switch (optname) {
1291 case RDMA_OPTION_IB_PATH:
1292 ret = ucma_set_ib_path(ctx, optval, optlen);
1293 break;
1294 default:
1295 ret = -ENOSYS;
1296 }
1297
1298 return ret;
1299 }
1300
1301 static int ucma_set_option_level(struct ucma_context *ctx, int level,
1302 int optname, void *optval, size_t optlen)
1303 {
1304 int ret;
1305
1306 switch (level) {
1307 case RDMA_OPTION_ID:
1308 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1309 break;
1310 case RDMA_OPTION_IB:
1311 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1312 break;
1313 default:
1314 ret = -ENOSYS;
1315 }
1316
1317 return ret;
1318 }
1319
1320 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1321 int in_len, int out_len)
1322 {
1323 struct rdma_ucm_set_option cmd;
1324 struct ucma_context *ctx;
1325 void *optval;
1326 int ret;
1327
1328 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1329 return -EFAULT;
1330
1331 if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1332 return -EINVAL;
1333
1334 ctx = ucma_get_ctx(file, cmd.id);
1335 if (IS_ERR(ctx))
1336 return PTR_ERR(ctx);
1337
1338 optval = memdup_user(u64_to_user_ptr(cmd.optval),
1339 cmd.optlen);
1340 if (IS_ERR(optval)) {
1341 ret = PTR_ERR(optval);
1342 goto out;
1343 }
1344
1345 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1346 cmd.optlen);
1347 kfree(optval);
1348
1349 out:
1350 ucma_put_ctx(ctx);
1351 return ret;
1352 }
1353
1354 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1355 int in_len, int out_len)
1356 {
1357 struct rdma_ucm_notify cmd;
1358 struct ucma_context *ctx;
1359 int ret = -EINVAL;
1360
1361 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1362 return -EFAULT;
1363
1364 ctx = ucma_get_ctx(file, cmd.id);
1365 if (IS_ERR(ctx))
1366 return PTR_ERR(ctx);
1367
1368 if (ctx->cm_id->device)
1369 ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
1370
1371 ucma_put_ctx(ctx);
1372 return ret;
1373 }
1374
1375 static ssize_t ucma_process_join(struct ucma_file *file,
1376 struct rdma_ucm_join_mcast *cmd, int out_len)
1377 {
1378 struct rdma_ucm_create_id_resp resp;
1379 struct ucma_context *ctx;
1380 struct ucma_multicast *mc;
1381 struct sockaddr *addr;
1382 int ret;
1383 u8 join_state;
1384
1385 if (out_len < sizeof(resp))
1386 return -ENOSPC;
1387
1388 addr = (struct sockaddr *) &cmd->addr;
1389 if (cmd->addr_size != rdma_addr_size(addr))
1390 return -EINVAL;
1391
1392 if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1393 join_state = BIT(FULLMEMBER_JOIN);
1394 else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1395 join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1396 else
1397 return -EINVAL;
1398
1399 ctx = ucma_get_ctx_dev(file, cmd->id);
1400 if (IS_ERR(ctx))
1401 return PTR_ERR(ctx);
1402
1403 mutex_lock(&file->mut);
1404 mc = ucma_alloc_multicast(ctx);
1405 if (!mc) {
1406 ret = -ENOMEM;
1407 goto err1;
1408 }
1409 mc->join_state = join_state;
1410 mc->uid = cmd->uid;
1411 memcpy(&mc->addr, addr, cmd->addr_size);
1412 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1413 join_state, mc);
1414 if (ret)
1415 goto err2;
1416
1417 resp.id = mc->id;
1418 if (copy_to_user(u64_to_user_ptr(cmd->response),
1419 &resp, sizeof(resp))) {
1420 ret = -EFAULT;
1421 goto err3;
1422 }
1423
1424 mutex_lock(&mut);
1425 idr_replace(&multicast_idr, mc, mc->id);
1426 mutex_unlock(&mut);
1427
1428 mutex_unlock(&file->mut);
1429 ucma_put_ctx(ctx);
1430 return 0;
1431
1432 err3:
1433 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1434 ucma_cleanup_mc_events(mc);
1435 err2:
1436 mutex_lock(&mut);
1437 idr_remove(&multicast_idr, mc->id);
1438 mutex_unlock(&mut);
1439 list_del(&mc->list);
1440 kfree(mc);
1441 err1:
1442 mutex_unlock(&file->mut);
1443 ucma_put_ctx(ctx);
1444 return ret;
1445 }
1446
1447 static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1448 const char __user *inbuf,
1449 int in_len, int out_len)
1450 {
1451 struct rdma_ucm_join_ip_mcast cmd;
1452 struct rdma_ucm_join_mcast join_cmd;
1453
1454 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1455 return -EFAULT;
1456
1457 join_cmd.response = cmd.response;
1458 join_cmd.uid = cmd.uid;
1459 join_cmd.id = cmd.id;
1460 join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1461 if (!join_cmd.addr_size)
1462 return -EINVAL;
1463
1464 join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1465 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1466
1467 return ucma_process_join(file, &join_cmd, out_len);
1468 }
1469
1470 static ssize_t ucma_join_multicast(struct ucma_file *file,
1471 const char __user *inbuf,
1472 int in_len, int out_len)
1473 {
1474 struct rdma_ucm_join_mcast cmd;
1475
1476 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1477 return -EFAULT;
1478
1479 if (!rdma_addr_size_kss(&cmd.addr))
1480 return -EINVAL;
1481
1482 return ucma_process_join(file, &cmd, out_len);
1483 }
1484
1485 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1486 const char __user *inbuf,
1487 int in_len, int out_len)
1488 {
1489 struct rdma_ucm_destroy_id cmd;
1490 struct rdma_ucm_destroy_id_resp resp;
1491 struct ucma_multicast *mc;
1492 int ret = 0;
1493
1494 if (out_len < sizeof(resp))
1495 return -ENOSPC;
1496
1497 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1498 return -EFAULT;
1499
1500 mutex_lock(&mut);
1501 mc = idr_find(&multicast_idr, cmd.id);
1502 if (!mc)
1503 mc = ERR_PTR(-ENOENT);
1504 else if (mc->ctx->file != file)
1505 mc = ERR_PTR(-EINVAL);
1506 else if (!atomic_inc_not_zero(&mc->ctx->ref))
1507 mc = ERR_PTR(-ENXIO);
1508 else
1509 idr_remove(&multicast_idr, mc->id);
1510 mutex_unlock(&mut);
1511
1512 if (IS_ERR(mc)) {
1513 ret = PTR_ERR(mc);
1514 goto out;
1515 }
1516
1517 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1518 mutex_lock(&mc->ctx->file->mut);
1519 ucma_cleanup_mc_events(mc);
1520 list_del(&mc->list);
1521 mutex_unlock(&mc->ctx->file->mut);
1522
1523 ucma_put_ctx(mc->ctx);
1524 resp.events_reported = mc->events_reported;
1525 kfree(mc);
1526
1527 if (copy_to_user(u64_to_user_ptr(cmd.response),
1528 &resp, sizeof(resp)))
1529 ret = -EFAULT;
1530 out:
1531 return ret;
1532 }
1533
1534 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1535 {
1536 /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1537 if (file1 < file2) {
1538 mutex_lock(&file1->mut);
1539 mutex_lock_nested(&file2->mut, SINGLE_DEPTH_NESTING);
1540 } else {
1541 mutex_lock(&file2->mut);
1542 mutex_lock_nested(&file1->mut, SINGLE_DEPTH_NESTING);
1543 }
1544 }
1545
1546 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1547 {
1548 if (file1 < file2) {
1549 mutex_unlock(&file2->mut);
1550 mutex_unlock(&file1->mut);
1551 } else {
1552 mutex_unlock(&file1->mut);
1553 mutex_unlock(&file2->mut);
1554 }
1555 }
1556
1557 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1558 {
1559 struct ucma_event *uevent, *tmp;
1560
1561 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1562 if (uevent->ctx == ctx)
1563 list_move_tail(&uevent->list, &file->event_list);
1564 }
1565
1566 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1567 const char __user *inbuf,
1568 int in_len, int out_len)
1569 {
1570 struct rdma_ucm_migrate_id cmd;
1571 struct rdma_ucm_migrate_resp resp;
1572 struct ucma_context *ctx;
1573 struct fd f;
1574 struct ucma_file *cur_file;
1575 int ret = 0;
1576
1577 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1578 return -EFAULT;
1579
1580 /* Get current fd to protect against it being closed */
1581 f = fdget(cmd.fd);
1582 if (!f.file)
1583 return -ENOENT;
1584
1585 /* Validate current fd and prevent destruction of id. */
1586 ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1587 if (IS_ERR(ctx)) {
1588 ret = PTR_ERR(ctx);
1589 goto file_put;
1590 }
1591
1592 cur_file = ctx->file;
1593 if (cur_file == new_file) {
1594 resp.events_reported = ctx->events_reported;
1595 goto response;
1596 }
1597
1598 /*
1599 * Migrate events between fd's, maintaining order, and avoiding new
1600 * events being added before existing events.
1601 */
1602 ucma_lock_files(cur_file, new_file);
1603 mutex_lock(&mut);
1604
1605 list_move_tail(&ctx->list, &new_file->ctx_list);
1606 ucma_move_events(ctx, new_file);
1607 ctx->file = new_file;
1608 resp.events_reported = ctx->events_reported;
1609
1610 mutex_unlock(&mut);
1611 ucma_unlock_files(cur_file, new_file);
1612
1613 response:
1614 if (copy_to_user(u64_to_user_ptr(cmd.response),
1615 &resp, sizeof(resp)))
1616 ret = -EFAULT;
1617
1618 ucma_put_ctx(ctx);
1619 file_put:
1620 fdput(f);
1621 return ret;
1622 }
1623
1624 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1625 const char __user *inbuf,
1626 int in_len, int out_len) = {
1627 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1628 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1629 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1630 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1631 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1632 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1633 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1634 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1635 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1636 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1637 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1638 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1639 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1640 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1641 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1642 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1643 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1644 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1645 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1646 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1647 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1648 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1649 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1650 };
1651
1652 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1653 size_t len, loff_t *pos)
1654 {
1655 struct ucma_file *file = filp->private_data;
1656 struct rdma_ucm_cmd_hdr hdr;
1657 ssize_t ret;
1658
1659 if (!ib_safe_file_access(filp)) {
1660 pr_err_once("ucma_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1661 task_tgid_vnr(current), current->comm);
1662 return -EACCES;
1663 }
1664
1665 if (len < sizeof(hdr))
1666 return -EINVAL;
1667
1668 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1669 return -EFAULT;
1670
1671 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1672 return -EINVAL;
1673
1674 if (hdr.in + sizeof(hdr) > len)
1675 return -EINVAL;
1676
1677 if (!ucma_cmd_table[hdr.cmd])
1678 return -ENOSYS;
1679
1680 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1681 if (!ret)
1682 ret = len;
1683
1684 return ret;
1685 }
1686
1687 static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1688 {
1689 struct ucma_file *file = filp->private_data;
1690 __poll_t mask = 0;
1691
1692 poll_wait(filp, &file->poll_wait, wait);
1693
1694 if (!list_empty(&file->event_list))
1695 mask = EPOLLIN | EPOLLRDNORM;
1696
1697 return mask;
1698 }
1699
1700 /*
1701 * ucma_open() does not need the BKL:
1702 *
1703 * - no global state is referred to;
1704 * - there is no ioctl method to race against;
1705 * - no further module initialization is required for open to work
1706 * after the device is registered.
1707 */
1708 static int ucma_open(struct inode *inode, struct file *filp)
1709 {
1710 struct ucma_file *file;
1711
1712 file = kmalloc(sizeof *file, GFP_KERNEL);
1713 if (!file)
1714 return -ENOMEM;
1715
1716 file->close_wq = alloc_ordered_workqueue("ucma_close_id",
1717 WQ_MEM_RECLAIM);
1718 if (!file->close_wq) {
1719 kfree(file);
1720 return -ENOMEM;
1721 }
1722
1723 INIT_LIST_HEAD(&file->event_list);
1724 INIT_LIST_HEAD(&file->ctx_list);
1725 init_waitqueue_head(&file->poll_wait);
1726 mutex_init(&file->mut);
1727
1728 filp->private_data = file;
1729 file->filp = filp;
1730
1731 return nonseekable_open(inode, filp);
1732 }
1733
1734 static int ucma_close(struct inode *inode, struct file *filp)
1735 {
1736 struct ucma_file *file = filp->private_data;
1737 struct ucma_context *ctx, *tmp;
1738
1739 mutex_lock(&file->mut);
1740 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1741 ctx->destroying = 1;
1742 mutex_unlock(&file->mut);
1743
1744 mutex_lock(&mut);
1745 idr_remove(&ctx_idr, ctx->id);
1746 mutex_unlock(&mut);
1747
1748 flush_workqueue(file->close_wq);
1749 /* At that step once ctx was marked as destroying and workqueue
1750 * was flushed we are safe from any inflights handlers that
1751 * might put other closing task.
1752 */
1753 mutex_lock(&mut);
1754 if (!ctx->closing) {
1755 mutex_unlock(&mut);
1756 /* rdma_destroy_id ensures that no event handlers are
1757 * inflight for that id before releasing it.
1758 */
1759 rdma_destroy_id(ctx->cm_id);
1760 } else {
1761 mutex_unlock(&mut);
1762 }
1763
1764 ucma_free_ctx(ctx);
1765 mutex_lock(&file->mut);
1766 }
1767 mutex_unlock(&file->mut);
1768 destroy_workqueue(file->close_wq);
1769 kfree(file);
1770 return 0;
1771 }
1772
1773 static const struct file_operations ucma_fops = {
1774 .owner = THIS_MODULE,
1775 .open = ucma_open,
1776 .release = ucma_close,
1777 .write = ucma_write,
1778 .poll = ucma_poll,
1779 .llseek = no_llseek,
1780 };
1781
1782 static struct miscdevice ucma_misc = {
1783 .minor = MISC_DYNAMIC_MINOR,
1784 .name = "rdma_cm",
1785 .nodename = "infiniband/rdma_cm",
1786 .mode = 0666,
1787 .fops = &ucma_fops,
1788 };
1789
1790 static ssize_t show_abi_version(struct device *dev,
1791 struct device_attribute *attr,
1792 char *buf)
1793 {
1794 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1795 }
1796 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1797
1798 static int __init ucma_init(void)
1799 {
1800 int ret;
1801
1802 ret = misc_register(&ucma_misc);
1803 if (ret)
1804 return ret;
1805
1806 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1807 if (ret) {
1808 pr_err("rdma_ucm: couldn't create abi_version attr\n");
1809 goto err1;
1810 }
1811
1812 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1813 if (!ucma_ctl_table_hdr) {
1814 pr_err("rdma_ucm: couldn't register sysctl paths\n");
1815 ret = -ENOMEM;
1816 goto err2;
1817 }
1818 return 0;
1819 err2:
1820 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1821 err1:
1822 misc_deregister(&ucma_misc);
1823 return ret;
1824 }
1825
1826 static void __exit ucma_cleanup(void)
1827 {
1828 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1829 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1830 misc_deregister(&ucma_misc);
1831 idr_destroy(&ctx_idr);
1832 idr_destroy(&multicast_idr);
1833 }
1834
1835 module_init(ucma_init);
1836 module_exit(ucma_cleanup);
Linux with added FPC-III support