RDMA/ucma: Fix AB-BA deadlock
[linux-2.6.git] / drivers / infiniband / core / ucma.c
blob5861cdb22b7c51c875720bd84cd1083d9133e2d0
1 /*
2 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
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.
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.
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>
46 #include <rdma/rdma_user_cm.h>
47 #include <rdma/ib_marshall.h>
48 #include <rdma/rdma_cm.h>
49 #include <rdma/rdma_cm_ib.h>
51 MODULE_AUTHOR("Sean Hefty");
52 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
53 MODULE_LICENSE("Dual BSD/GPL");
55 static unsigned int max_backlog = 1024;
57 static struct ctl_table_header *ucma_ctl_table_hdr;
58 static ctl_table ucma_ctl_table[] = {
60 .procname = "max_backlog",
61 .data = &max_backlog,
62 .maxlen = sizeof max_backlog,
63 .mode = 0644,
64 .proc_handler = proc_dointvec,
66 { }
69 static struct ctl_path ucma_ctl_path[] = {
70 { .procname = "net" },
71 { .procname = "rdma_ucm" },
72 { }
75 struct ucma_file {
76 struct mutex mut;
77 struct file *filp;
78 struct list_head ctx_list;
79 struct list_head event_list;
80 wait_queue_head_t poll_wait;
83 struct ucma_context {
84 int id;
85 struct completion comp;
86 atomic_t ref;
87 int events_reported;
88 int backlog;
90 struct ucma_file *file;
91 struct rdma_cm_id *cm_id;
92 u64 uid;
94 struct list_head list;
95 struct list_head mc_list;
98 struct ucma_multicast {
99 struct ucma_context *ctx;
100 int id;
101 int events_reported;
103 u64 uid;
104 struct list_head list;
105 struct sockaddr_storage addr;
108 struct ucma_event {
109 struct ucma_context *ctx;
110 struct ucma_multicast *mc;
111 struct list_head list;
112 struct rdma_cm_id *cm_id;
113 struct rdma_ucm_event_resp resp;
116 static DEFINE_MUTEX(mut);
117 static DEFINE_IDR(ctx_idr);
118 static DEFINE_IDR(multicast_idr);
120 static inline struct ucma_context *_ucma_find_context(int id,
121 struct ucma_file *file)
123 struct ucma_context *ctx;
125 ctx = idr_find(&ctx_idr, id);
126 if (!ctx)
127 ctx = ERR_PTR(-ENOENT);
128 else if (ctx->file != file)
129 ctx = ERR_PTR(-EINVAL);
130 return ctx;
133 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
135 struct ucma_context *ctx;
137 mutex_lock(&mut);
138 ctx = _ucma_find_context(id, file);
139 if (!IS_ERR(ctx))
140 atomic_inc(&ctx->ref);
141 mutex_unlock(&mut);
142 return ctx;
145 static void ucma_put_ctx(struct ucma_context *ctx)
147 if (atomic_dec_and_test(&ctx->ref))
148 complete(&ctx->comp);
151 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
153 struct ucma_context *ctx;
154 int ret;
156 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
157 if (!ctx)
158 return NULL;
160 atomic_set(&ctx->ref, 1);
161 init_completion(&ctx->comp);
162 INIT_LIST_HEAD(&ctx->mc_list);
163 ctx->file = file;
165 do {
166 ret = idr_pre_get(&ctx_idr, GFP_KERNEL);
167 if (!ret)
168 goto error;
170 mutex_lock(&mut);
171 ret = idr_get_new(&ctx_idr, ctx, &ctx->id);
172 mutex_unlock(&mut);
173 } while (ret == -EAGAIN);
175 if (ret)
176 goto error;
178 list_add_tail(&ctx->list, &file->ctx_list);
179 return ctx;
181 error:
182 kfree(ctx);
183 return NULL;
186 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
188 struct ucma_multicast *mc;
189 int ret;
191 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
192 if (!mc)
193 return NULL;
195 do {
196 ret = idr_pre_get(&multicast_idr, GFP_KERNEL);
197 if (!ret)
198 goto error;
200 mutex_lock(&mut);
201 ret = idr_get_new(&multicast_idr, mc, &mc->id);
202 mutex_unlock(&mut);
203 } while (ret == -EAGAIN);
205 if (ret)
206 goto error;
208 mc->ctx = ctx;
209 list_add_tail(&mc->list, &ctx->mc_list);
210 return mc;
212 error:
213 kfree(mc);
214 return NULL;
217 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
218 struct rdma_conn_param *src)
220 if (src->private_data_len)
221 memcpy(dst->private_data, src->private_data,
222 src->private_data_len);
223 dst->private_data_len = src->private_data_len;
224 dst->responder_resources =src->responder_resources;
225 dst->initiator_depth = src->initiator_depth;
226 dst->flow_control = src->flow_control;
227 dst->retry_count = src->retry_count;
228 dst->rnr_retry_count = src->rnr_retry_count;
229 dst->srq = src->srq;
230 dst->qp_num = src->qp_num;
233 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
234 struct rdma_ud_param *src)
236 if (src->private_data_len)
237 memcpy(dst->private_data, src->private_data,
238 src->private_data_len);
239 dst->private_data_len = src->private_data_len;
240 ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
241 dst->qp_num = src->qp_num;
242 dst->qkey = src->qkey;
245 static void ucma_set_event_context(struct ucma_context *ctx,
246 struct rdma_cm_event *event,
247 struct ucma_event *uevent)
249 uevent->ctx = ctx;
250 switch (event->event) {
251 case RDMA_CM_EVENT_MULTICAST_JOIN:
252 case RDMA_CM_EVENT_MULTICAST_ERROR:
253 uevent->mc = (struct ucma_multicast *)
254 event->param.ud.private_data;
255 uevent->resp.uid = uevent->mc->uid;
256 uevent->resp.id = uevent->mc->id;
257 break;
258 default:
259 uevent->resp.uid = ctx->uid;
260 uevent->resp.id = ctx->id;
261 break;
265 static int ucma_event_handler(struct rdma_cm_id *cm_id,
266 struct rdma_cm_event *event)
268 struct ucma_event *uevent;
269 struct ucma_context *ctx = cm_id->context;
270 int ret = 0;
272 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
273 if (!uevent)
274 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
276 uevent->cm_id = cm_id;
277 ucma_set_event_context(ctx, event, uevent);
278 uevent->resp.event = event->event;
279 uevent->resp.status = event->status;
280 if (cm_id->qp_type == IB_QPT_UD)
281 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
282 else
283 ucma_copy_conn_event(&uevent->resp.param.conn,
284 &event->param.conn);
286 mutex_lock(&ctx->file->mut);
287 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
288 if (!ctx->backlog) {
289 ret = -ENOMEM;
290 kfree(uevent);
291 goto out;
293 ctx->backlog--;
294 } else if (!ctx->uid) {
296 * We ignore events for new connections until userspace has set
297 * their context. This can only happen if an error occurs on a
298 * new connection before the user accepts it. This is okay,
299 * since the accept will just fail later.
301 kfree(uevent);
302 goto out;
305 list_add_tail(&uevent->list, &ctx->file->event_list);
306 wake_up_interruptible(&ctx->file->poll_wait);
307 out:
308 mutex_unlock(&ctx->file->mut);
309 return ret;
312 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
313 int in_len, int out_len)
315 struct ucma_context *ctx;
316 struct rdma_ucm_get_event cmd;
317 struct ucma_event *uevent;
318 int ret = 0;
319 DEFINE_WAIT(wait);
321 if (out_len < sizeof uevent->resp)
322 return -ENOSPC;
324 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
325 return -EFAULT;
327 mutex_lock(&file->mut);
328 while (list_empty(&file->event_list)) {
329 mutex_unlock(&file->mut);
331 if (file->filp->f_flags & O_NONBLOCK)
332 return -EAGAIN;
334 if (wait_event_interruptible(file->poll_wait,
335 !list_empty(&file->event_list)))
336 return -ERESTARTSYS;
338 mutex_lock(&file->mut);
341 uevent = list_entry(file->event_list.next, struct ucma_event, list);
343 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
344 ctx = ucma_alloc_ctx(file);
345 if (!ctx) {
346 ret = -ENOMEM;
347 goto done;
349 uevent->ctx->backlog++;
350 ctx->cm_id = uevent->cm_id;
351 ctx->cm_id->context = ctx;
352 uevent->resp.id = ctx->id;
355 if (copy_to_user((void __user *)(unsigned long)cmd.response,
356 &uevent->resp, sizeof uevent->resp)) {
357 ret = -EFAULT;
358 goto done;
361 list_del(&uevent->list);
362 uevent->ctx->events_reported++;
363 if (uevent->mc)
364 uevent->mc->events_reported++;
365 kfree(uevent);
366 done:
367 mutex_unlock(&file->mut);
368 return ret;
371 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
373 switch (cmd->ps) {
374 case RDMA_PS_TCP:
375 *qp_type = IB_QPT_RC;
376 return 0;
377 case RDMA_PS_UDP:
378 case RDMA_PS_IPOIB:
379 *qp_type = IB_QPT_UD;
380 return 0;
381 case RDMA_PS_IB:
382 *qp_type = cmd->qp_type;
383 return 0;
384 default:
385 return -EINVAL;
389 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
390 int in_len, int out_len)
392 struct rdma_ucm_create_id cmd;
393 struct rdma_ucm_create_id_resp resp;
394 struct ucma_context *ctx;
395 enum ib_qp_type qp_type;
396 int ret;
398 if (out_len < sizeof(resp))
399 return -ENOSPC;
401 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
402 return -EFAULT;
404 ret = ucma_get_qp_type(&cmd, &qp_type);
405 if (ret)
406 return ret;
408 mutex_lock(&file->mut);
409 ctx = ucma_alloc_ctx(file);
410 mutex_unlock(&file->mut);
411 if (!ctx)
412 return -ENOMEM;
414 ctx->uid = cmd.uid;
415 ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps, qp_type);
416 if (IS_ERR(ctx->cm_id)) {
417 ret = PTR_ERR(ctx->cm_id);
418 goto err1;
421 resp.id = ctx->id;
422 if (copy_to_user((void __user *)(unsigned long)cmd.response,
423 &resp, sizeof(resp))) {
424 ret = -EFAULT;
425 goto err2;
427 return 0;
429 err2:
430 rdma_destroy_id(ctx->cm_id);
431 err1:
432 mutex_lock(&mut);
433 idr_remove(&ctx_idr, ctx->id);
434 mutex_unlock(&mut);
435 kfree(ctx);
436 return ret;
439 static void ucma_cleanup_multicast(struct ucma_context *ctx)
441 struct ucma_multicast *mc, *tmp;
443 mutex_lock(&mut);
444 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
445 list_del(&mc->list);
446 idr_remove(&multicast_idr, mc->id);
447 kfree(mc);
449 mutex_unlock(&mut);
452 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
454 struct ucma_event *uevent, *tmp;
456 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
457 if (uevent->mc != mc)
458 continue;
460 list_del(&uevent->list);
461 kfree(uevent);
466 * We cannot hold file->mut when calling rdma_destroy_id() or we can
467 * deadlock. We also acquire file->mut in ucma_event_handler(), and
468 * rdma_destroy_id() will wait until all callbacks have completed.
470 static int ucma_free_ctx(struct ucma_context *ctx)
472 int events_reported;
473 struct ucma_event *uevent, *tmp;
474 LIST_HEAD(list);
476 /* No new events will be generated after destroying the id. */
477 rdma_destroy_id(ctx->cm_id);
479 ucma_cleanup_multicast(ctx);
481 /* Cleanup events not yet reported to the user. */
482 mutex_lock(&ctx->file->mut);
483 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
484 if (uevent->ctx == ctx)
485 list_move_tail(&uevent->list, &list);
487 list_del(&ctx->list);
488 mutex_unlock(&ctx->file->mut);
490 list_for_each_entry_safe(uevent, tmp, &list, list) {
491 list_del(&uevent->list);
492 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
493 rdma_destroy_id(uevent->cm_id);
494 kfree(uevent);
497 events_reported = ctx->events_reported;
498 kfree(ctx);
499 return events_reported;
502 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
503 int in_len, int out_len)
505 struct rdma_ucm_destroy_id cmd;
506 struct rdma_ucm_destroy_id_resp resp;
507 struct ucma_context *ctx;
508 int ret = 0;
510 if (out_len < sizeof(resp))
511 return -ENOSPC;
513 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
514 return -EFAULT;
516 mutex_lock(&mut);
517 ctx = _ucma_find_context(cmd.id, file);
518 if (!IS_ERR(ctx))
519 idr_remove(&ctx_idr, ctx->id);
520 mutex_unlock(&mut);
522 if (IS_ERR(ctx))
523 return PTR_ERR(ctx);
525 ucma_put_ctx(ctx);
526 wait_for_completion(&ctx->comp);
527 resp.events_reported = ucma_free_ctx(ctx);
529 if (copy_to_user((void __user *)(unsigned long)cmd.response,
530 &resp, sizeof(resp)))
531 ret = -EFAULT;
533 return ret;
536 static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf,
537 int in_len, int out_len)
539 struct rdma_ucm_bind_addr cmd;
540 struct ucma_context *ctx;
541 int ret;
543 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
544 return -EFAULT;
546 ctx = ucma_get_ctx(file, cmd.id);
547 if (IS_ERR(ctx))
548 return PTR_ERR(ctx);
550 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
551 ucma_put_ctx(ctx);
552 return ret;
555 static ssize_t ucma_resolve_addr(struct ucma_file *file,
556 const char __user *inbuf,
557 int in_len, int out_len)
559 struct rdma_ucm_resolve_addr cmd;
560 struct ucma_context *ctx;
561 int ret;
563 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
564 return -EFAULT;
566 ctx = ucma_get_ctx(file, cmd.id);
567 if (IS_ERR(ctx))
568 return PTR_ERR(ctx);
570 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
571 (struct sockaddr *) &cmd.dst_addr,
572 cmd.timeout_ms);
573 ucma_put_ctx(ctx);
574 return ret;
577 static ssize_t ucma_resolve_route(struct ucma_file *file,
578 const char __user *inbuf,
579 int in_len, int out_len)
581 struct rdma_ucm_resolve_route cmd;
582 struct ucma_context *ctx;
583 int ret;
585 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
586 return -EFAULT;
588 ctx = ucma_get_ctx(file, cmd.id);
589 if (IS_ERR(ctx))
590 return PTR_ERR(ctx);
592 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
593 ucma_put_ctx(ctx);
594 return ret;
597 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
598 struct rdma_route *route)
600 struct rdma_dev_addr *dev_addr;
602 resp->num_paths = route->num_paths;
603 switch (route->num_paths) {
604 case 0:
605 dev_addr = &route->addr.dev_addr;
606 rdma_addr_get_dgid(dev_addr,
607 (union ib_gid *) &resp->ib_route[0].dgid);
608 rdma_addr_get_sgid(dev_addr,
609 (union ib_gid *) &resp->ib_route[0].sgid);
610 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
611 break;
612 case 2:
613 ib_copy_path_rec_to_user(&resp->ib_route[1],
614 &route->path_rec[1]);
615 /* fall through */
616 case 1:
617 ib_copy_path_rec_to_user(&resp->ib_route[0],
618 &route->path_rec[0]);
619 break;
620 default:
621 break;
625 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
626 struct rdma_route *route)
628 struct rdma_dev_addr *dev_addr;
629 struct net_device *dev;
630 u16 vid = 0;
632 resp->num_paths = route->num_paths;
633 switch (route->num_paths) {
634 case 0:
635 dev_addr = &route->addr.dev_addr;
636 dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
637 if (dev) {
638 vid = rdma_vlan_dev_vlan_id(dev);
639 dev_put(dev);
642 iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid,
643 dev_addr->dst_dev_addr, vid);
644 iboe_addr_get_sgid(dev_addr,
645 (union ib_gid *) &resp->ib_route[0].sgid);
646 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
647 break;
648 case 2:
649 ib_copy_path_rec_to_user(&resp->ib_route[1],
650 &route->path_rec[1]);
651 /* fall through */
652 case 1:
653 ib_copy_path_rec_to_user(&resp->ib_route[0],
654 &route->path_rec[0]);
655 break;
656 default:
657 break;
661 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
662 struct rdma_route *route)
664 struct rdma_dev_addr *dev_addr;
666 dev_addr = &route->addr.dev_addr;
667 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
668 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
671 static ssize_t ucma_query_route(struct ucma_file *file,
672 const char __user *inbuf,
673 int in_len, int out_len)
675 struct rdma_ucm_query_route cmd;
676 struct rdma_ucm_query_route_resp resp;
677 struct ucma_context *ctx;
678 struct sockaddr *addr;
679 int ret = 0;
681 if (out_len < sizeof(resp))
682 return -ENOSPC;
684 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
685 return -EFAULT;
687 ctx = ucma_get_ctx(file, cmd.id);
688 if (IS_ERR(ctx))
689 return PTR_ERR(ctx);
691 memset(&resp, 0, sizeof resp);
692 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
693 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
694 sizeof(struct sockaddr_in) :
695 sizeof(struct sockaddr_in6));
696 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
697 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
698 sizeof(struct sockaddr_in) :
699 sizeof(struct sockaddr_in6));
700 if (!ctx->cm_id->device)
701 goto out;
703 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
704 resp.port_num = ctx->cm_id->port_num;
705 switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
706 case RDMA_TRANSPORT_IB:
707 switch (rdma_port_get_link_layer(ctx->cm_id->device,
708 ctx->cm_id->port_num)) {
709 case IB_LINK_LAYER_INFINIBAND:
710 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
711 break;
712 case IB_LINK_LAYER_ETHERNET:
713 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
714 break;
715 default:
716 break;
718 break;
719 case RDMA_TRANSPORT_IWARP:
720 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
721 break;
722 default:
723 break;
726 out:
727 if (copy_to_user((void __user *)(unsigned long)cmd.response,
728 &resp, sizeof(resp)))
729 ret = -EFAULT;
731 ucma_put_ctx(ctx);
732 return ret;
735 static void ucma_copy_conn_param(struct rdma_conn_param *dst,
736 struct rdma_ucm_conn_param *src)
738 dst->private_data = src->private_data;
739 dst->private_data_len = src->private_data_len;
740 dst->responder_resources =src->responder_resources;
741 dst->initiator_depth = src->initiator_depth;
742 dst->flow_control = src->flow_control;
743 dst->retry_count = src->retry_count;
744 dst->rnr_retry_count = src->rnr_retry_count;
745 dst->srq = src->srq;
746 dst->qp_num = src->qp_num;
749 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
750 int in_len, int out_len)
752 struct rdma_ucm_connect cmd;
753 struct rdma_conn_param conn_param;
754 struct ucma_context *ctx;
755 int ret;
757 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
758 return -EFAULT;
760 if (!cmd.conn_param.valid)
761 return -EINVAL;
763 ctx = ucma_get_ctx(file, cmd.id);
764 if (IS_ERR(ctx))
765 return PTR_ERR(ctx);
767 ucma_copy_conn_param(&conn_param, &cmd.conn_param);
768 ret = rdma_connect(ctx->cm_id, &conn_param);
769 ucma_put_ctx(ctx);
770 return ret;
773 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
774 int in_len, int out_len)
776 struct rdma_ucm_listen cmd;
777 struct ucma_context *ctx;
778 int ret;
780 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
781 return -EFAULT;
783 ctx = ucma_get_ctx(file, cmd.id);
784 if (IS_ERR(ctx))
785 return PTR_ERR(ctx);
787 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
788 cmd.backlog : max_backlog;
789 ret = rdma_listen(ctx->cm_id, ctx->backlog);
790 ucma_put_ctx(ctx);
791 return ret;
794 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
795 int in_len, int out_len)
797 struct rdma_ucm_accept cmd;
798 struct rdma_conn_param conn_param;
799 struct ucma_context *ctx;
800 int ret;
802 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
803 return -EFAULT;
805 ctx = ucma_get_ctx(file, cmd.id);
806 if (IS_ERR(ctx))
807 return PTR_ERR(ctx);
809 if (cmd.conn_param.valid) {
810 ucma_copy_conn_param(&conn_param, &cmd.conn_param);
811 mutex_lock(&file->mut);
812 ret = rdma_accept(ctx->cm_id, &conn_param);
813 if (!ret)
814 ctx->uid = cmd.uid;
815 mutex_unlock(&file->mut);
816 } else
817 ret = rdma_accept(ctx->cm_id, NULL);
819 ucma_put_ctx(ctx);
820 return ret;
823 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
824 int in_len, int out_len)
826 struct rdma_ucm_reject cmd;
827 struct ucma_context *ctx;
828 int ret;
830 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
831 return -EFAULT;
833 ctx = ucma_get_ctx(file, cmd.id);
834 if (IS_ERR(ctx))
835 return PTR_ERR(ctx);
837 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
838 ucma_put_ctx(ctx);
839 return ret;
842 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
843 int in_len, int out_len)
845 struct rdma_ucm_disconnect cmd;
846 struct ucma_context *ctx;
847 int ret;
849 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
850 return -EFAULT;
852 ctx = ucma_get_ctx(file, cmd.id);
853 if (IS_ERR(ctx))
854 return PTR_ERR(ctx);
856 ret = rdma_disconnect(ctx->cm_id);
857 ucma_put_ctx(ctx);
858 return ret;
861 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
862 const char __user *inbuf,
863 int in_len, int out_len)
865 struct rdma_ucm_init_qp_attr cmd;
866 struct ib_uverbs_qp_attr resp;
867 struct ucma_context *ctx;
868 struct ib_qp_attr qp_attr;
869 int ret;
871 if (out_len < sizeof(resp))
872 return -ENOSPC;
874 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
875 return -EFAULT;
877 ctx = ucma_get_ctx(file, cmd.id);
878 if (IS_ERR(ctx))
879 return PTR_ERR(ctx);
881 resp.qp_attr_mask = 0;
882 memset(&qp_attr, 0, sizeof qp_attr);
883 qp_attr.qp_state = cmd.qp_state;
884 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
885 if (ret)
886 goto out;
888 ib_copy_qp_attr_to_user(&resp, &qp_attr);
889 if (copy_to_user((void __user *)(unsigned long)cmd.response,
890 &resp, sizeof(resp)))
891 ret = -EFAULT;
893 out:
894 ucma_put_ctx(ctx);
895 return ret;
898 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
899 void *optval, size_t optlen)
901 int ret = 0;
903 switch (optname) {
904 case RDMA_OPTION_ID_TOS:
905 if (optlen != sizeof(u8)) {
906 ret = -EINVAL;
907 break;
909 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
910 break;
911 case RDMA_OPTION_ID_REUSEADDR:
912 if (optlen != sizeof(int)) {
913 ret = -EINVAL;
914 break;
916 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
917 break;
918 default:
919 ret = -ENOSYS;
922 return ret;
925 static int ucma_set_ib_path(struct ucma_context *ctx,
926 struct ib_path_rec_data *path_data, size_t optlen)
928 struct ib_sa_path_rec sa_path;
929 struct rdma_cm_event event;
930 int ret;
932 if (optlen % sizeof(*path_data))
933 return -EINVAL;
935 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
936 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
937 IB_PATH_BIDIRECTIONAL))
938 break;
941 if (!optlen)
942 return -EINVAL;
944 ib_sa_unpack_path(path_data->path_rec, &sa_path);
945 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
946 if (ret)
947 return ret;
949 memset(&event, 0, sizeof event);
950 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
951 return ucma_event_handler(ctx->cm_id, &event);
954 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
955 void *optval, size_t optlen)
957 int ret;
959 switch (optname) {
960 case RDMA_OPTION_IB_PATH:
961 ret = ucma_set_ib_path(ctx, optval, optlen);
962 break;
963 default:
964 ret = -ENOSYS;
967 return ret;
970 static int ucma_set_option_level(struct ucma_context *ctx, int level,
971 int optname, void *optval, size_t optlen)
973 int ret;
975 switch (level) {
976 case RDMA_OPTION_ID:
977 ret = ucma_set_option_id(ctx, optname, optval, optlen);
978 break;
979 case RDMA_OPTION_IB:
980 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
981 break;
982 default:
983 ret = -ENOSYS;
986 return ret;
989 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
990 int in_len, int out_len)
992 struct rdma_ucm_set_option cmd;
993 struct ucma_context *ctx;
994 void *optval;
995 int ret;
997 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
998 return -EFAULT;
1000 ctx = ucma_get_ctx(file, cmd.id);
1001 if (IS_ERR(ctx))
1002 return PTR_ERR(ctx);
1004 optval = kmalloc(cmd.optlen, GFP_KERNEL);
1005 if (!optval) {
1006 ret = -ENOMEM;
1007 goto out1;
1010 if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval,
1011 cmd.optlen)) {
1012 ret = -EFAULT;
1013 goto out2;
1016 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1017 cmd.optlen);
1018 out2:
1019 kfree(optval);
1020 out1:
1021 ucma_put_ctx(ctx);
1022 return ret;
1025 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1026 int in_len, int out_len)
1028 struct rdma_ucm_notify cmd;
1029 struct ucma_context *ctx;
1030 int ret;
1032 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1033 return -EFAULT;
1035 ctx = ucma_get_ctx(file, cmd.id);
1036 if (IS_ERR(ctx))
1037 return PTR_ERR(ctx);
1039 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1040 ucma_put_ctx(ctx);
1041 return ret;
1044 static ssize_t ucma_join_multicast(struct ucma_file *file,
1045 const char __user *inbuf,
1046 int in_len, int out_len)
1048 struct rdma_ucm_join_mcast cmd;
1049 struct rdma_ucm_create_id_resp resp;
1050 struct ucma_context *ctx;
1051 struct ucma_multicast *mc;
1052 int ret;
1054 if (out_len < sizeof(resp))
1055 return -ENOSPC;
1057 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1058 return -EFAULT;
1060 ctx = ucma_get_ctx(file, cmd.id);
1061 if (IS_ERR(ctx))
1062 return PTR_ERR(ctx);
1064 mutex_lock(&file->mut);
1065 mc = ucma_alloc_multicast(ctx);
1066 if (!mc) {
1067 ret = -ENOMEM;
1068 goto err1;
1071 mc->uid = cmd.uid;
1072 memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr);
1073 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1074 if (ret)
1075 goto err2;
1077 resp.id = mc->id;
1078 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1079 &resp, sizeof(resp))) {
1080 ret = -EFAULT;
1081 goto err3;
1084 mutex_unlock(&file->mut);
1085 ucma_put_ctx(ctx);
1086 return 0;
1088 err3:
1089 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1090 ucma_cleanup_mc_events(mc);
1091 err2:
1092 mutex_lock(&mut);
1093 idr_remove(&multicast_idr, mc->id);
1094 mutex_unlock(&mut);
1095 list_del(&mc->list);
1096 kfree(mc);
1097 err1:
1098 mutex_unlock(&file->mut);
1099 ucma_put_ctx(ctx);
1100 return ret;
1103 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1104 const char __user *inbuf,
1105 int in_len, int out_len)
1107 struct rdma_ucm_destroy_id cmd;
1108 struct rdma_ucm_destroy_id_resp resp;
1109 struct ucma_multicast *mc;
1110 int ret = 0;
1112 if (out_len < sizeof(resp))
1113 return -ENOSPC;
1115 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1116 return -EFAULT;
1118 mutex_lock(&mut);
1119 mc = idr_find(&multicast_idr, cmd.id);
1120 if (!mc)
1121 mc = ERR_PTR(-ENOENT);
1122 else if (mc->ctx->file != file)
1123 mc = ERR_PTR(-EINVAL);
1124 else {
1125 idr_remove(&multicast_idr, mc->id);
1126 atomic_inc(&mc->ctx->ref);
1128 mutex_unlock(&mut);
1130 if (IS_ERR(mc)) {
1131 ret = PTR_ERR(mc);
1132 goto out;
1135 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1136 mutex_lock(&mc->ctx->file->mut);
1137 ucma_cleanup_mc_events(mc);
1138 list_del(&mc->list);
1139 mutex_unlock(&mc->ctx->file->mut);
1141 ucma_put_ctx(mc->ctx);
1142 resp.events_reported = mc->events_reported;
1143 kfree(mc);
1145 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1146 &resp, sizeof(resp)))
1147 ret = -EFAULT;
1148 out:
1149 return ret;
1152 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1154 /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1155 if (file1 < file2) {
1156 mutex_lock(&file1->mut);
1157 mutex_lock(&file2->mut);
1158 } else {
1159 mutex_lock(&file2->mut);
1160 mutex_lock(&file1->mut);
1164 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1166 if (file1 < file2) {
1167 mutex_unlock(&file2->mut);
1168 mutex_unlock(&file1->mut);
1169 } else {
1170 mutex_unlock(&file1->mut);
1171 mutex_unlock(&file2->mut);
1175 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1177 struct ucma_event *uevent, *tmp;
1179 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1180 if (uevent->ctx == ctx)
1181 list_move_tail(&uevent->list, &file->event_list);
1184 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1185 const char __user *inbuf,
1186 int in_len, int out_len)
1188 struct rdma_ucm_migrate_id cmd;
1189 struct rdma_ucm_migrate_resp resp;
1190 struct ucma_context *ctx;
1191 struct file *filp;
1192 struct ucma_file *cur_file;
1193 int ret = 0;
1195 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1196 return -EFAULT;
1198 /* Get current fd to protect against it being closed */
1199 filp = fget(cmd.fd);
1200 if (!filp)
1201 return -ENOENT;
1203 /* Validate current fd and prevent destruction of id. */
1204 ctx = ucma_get_ctx(filp->private_data, cmd.id);
1205 if (IS_ERR(ctx)) {
1206 ret = PTR_ERR(ctx);
1207 goto file_put;
1210 cur_file = ctx->file;
1211 if (cur_file == new_file) {
1212 resp.events_reported = ctx->events_reported;
1213 goto response;
1217 * Migrate events between fd's, maintaining order, and avoiding new
1218 * events being added before existing events.
1220 ucma_lock_files(cur_file, new_file);
1221 mutex_lock(&mut);
1223 list_move_tail(&ctx->list, &new_file->ctx_list);
1224 ucma_move_events(ctx, new_file);
1225 ctx->file = new_file;
1226 resp.events_reported = ctx->events_reported;
1228 mutex_unlock(&mut);
1229 ucma_unlock_files(cur_file, new_file);
1231 response:
1232 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1233 &resp, sizeof(resp)))
1234 ret = -EFAULT;
1236 ucma_put_ctx(ctx);
1237 file_put:
1238 fput(filp);
1239 return ret;
1242 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1243 const char __user *inbuf,
1244 int in_len, int out_len) = {
1245 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1246 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1247 [RDMA_USER_CM_CMD_BIND_ADDR] = ucma_bind_addr,
1248 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1249 [RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route,
1250 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1251 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1252 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1253 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1254 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1255 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1256 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1257 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1258 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1259 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1260 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1261 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast,
1262 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1263 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id
1266 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1267 size_t len, loff_t *pos)
1269 struct ucma_file *file = filp->private_data;
1270 struct rdma_ucm_cmd_hdr hdr;
1271 ssize_t ret;
1273 if (len < sizeof(hdr))
1274 return -EINVAL;
1276 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1277 return -EFAULT;
1279 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1280 return -EINVAL;
1282 if (hdr.in + sizeof(hdr) > len)
1283 return -EINVAL;
1285 if (!ucma_cmd_table[hdr.cmd])
1286 return -ENOSYS;
1288 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1289 if (!ret)
1290 ret = len;
1292 return ret;
1295 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1297 struct ucma_file *file = filp->private_data;
1298 unsigned int mask = 0;
1300 poll_wait(filp, &file->poll_wait, wait);
1302 if (!list_empty(&file->event_list))
1303 mask = POLLIN | POLLRDNORM;
1305 return mask;
1309 * ucma_open() does not need the BKL:
1311 * - no global state is referred to;
1312 * - there is no ioctl method to race against;
1313 * - no further module initialization is required for open to work
1314 * after the device is registered.
1316 static int ucma_open(struct inode *inode, struct file *filp)
1318 struct ucma_file *file;
1320 file = kmalloc(sizeof *file, GFP_KERNEL);
1321 if (!file)
1322 return -ENOMEM;
1324 INIT_LIST_HEAD(&file->event_list);
1325 INIT_LIST_HEAD(&file->ctx_list);
1326 init_waitqueue_head(&file->poll_wait);
1327 mutex_init(&file->mut);
1329 filp->private_data = file;
1330 file->filp = filp;
1332 return nonseekable_open(inode, filp);
1335 static int ucma_close(struct inode *inode, struct file *filp)
1337 struct ucma_file *file = filp->private_data;
1338 struct ucma_context *ctx, *tmp;
1340 mutex_lock(&file->mut);
1341 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1342 mutex_unlock(&file->mut);
1344 mutex_lock(&mut);
1345 idr_remove(&ctx_idr, ctx->id);
1346 mutex_unlock(&mut);
1348 ucma_free_ctx(ctx);
1349 mutex_lock(&file->mut);
1351 mutex_unlock(&file->mut);
1352 kfree(file);
1353 return 0;
1356 static const struct file_operations ucma_fops = {
1357 .owner = THIS_MODULE,
1358 .open = ucma_open,
1359 .release = ucma_close,
1360 .write = ucma_write,
1361 .poll = ucma_poll,
1362 .llseek = no_llseek,
1365 static struct miscdevice ucma_misc = {
1366 .minor = MISC_DYNAMIC_MINOR,
1367 .name = "rdma_cm",
1368 .nodename = "infiniband/rdma_cm",
1369 .mode = 0666,
1370 .fops = &ucma_fops,
1373 static ssize_t show_abi_version(struct device *dev,
1374 struct device_attribute *attr,
1375 char *buf)
1377 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1379 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1381 static int __init ucma_init(void)
1383 int ret;
1385 ret = misc_register(&ucma_misc);
1386 if (ret)
1387 return ret;
1389 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1390 if (ret) {
1391 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1392 goto err1;
1395 ucma_ctl_table_hdr = register_sysctl_paths(ucma_ctl_path, ucma_ctl_table);
1396 if (!ucma_ctl_table_hdr) {
1397 printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n");
1398 ret = -ENOMEM;
1399 goto err2;
1401 return 0;
1402 err2:
1403 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1404 err1:
1405 misc_deregister(&ucma_misc);
1406 return ret;
1409 static void __exit ucma_cleanup(void)
1411 unregister_sysctl_table(ucma_ctl_table_hdr);
1412 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1413 misc_deregister(&ucma_misc);
1414 idr_destroy(&ctx_idr);
1417 module_init(ucma_init);
1418 module_exit(ucma_cleanup);