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
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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
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>
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>
50 #include <rdma/ib_addr.h>
53 MODULE_AUTHOR("Sean Hefty");
54 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
55 MODULE_LICENSE("Dual BSD/GPL");
57 static unsigned int max_backlog = 1024;
59 static struct ctl_table_header *ucma_ctl_table_hdr;
60 static struct ctl_table ucma_ctl_table[] = {
62 .procname = "max_backlog",
64 .maxlen = sizeof max_backlog,
66 .proc_handler = proc_dointvec,
74 struct list_head ctx_list;
75 struct list_head event_list;
76 wait_queue_head_t poll_wait;
81 struct completion comp;
86 struct ucma_file *file;
87 struct rdma_cm_id *cm_id;
90 struct list_head list;
91 struct list_head mc_list;
94 struct ucma_multicast {
95 struct ucma_context *ctx;
100 struct list_head list;
101 struct sockaddr_storage addr;
105 struct ucma_context *ctx;
106 struct ucma_multicast *mc;
107 struct list_head list;
108 struct rdma_cm_id *cm_id;
109 struct rdma_ucm_event_resp resp;
112 static DEFINE_MUTEX(mut);
113 static DEFINE_IDR(ctx_idr);
114 static DEFINE_IDR(multicast_idr);
116 static inline struct ucma_context *_ucma_find_context(int id,
117 struct ucma_file *file)
119 struct ucma_context *ctx;
121 ctx = idr_find(&ctx_idr, id);
123 ctx = ERR_PTR(-ENOENT);
124 else if (ctx->file != file)
125 ctx = ERR_PTR(-EINVAL);
129 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
131 struct ucma_context *ctx;
134 ctx = _ucma_find_context(id, file);
136 atomic_inc(&ctx->ref);
141 static void ucma_put_ctx(struct ucma_context *ctx)
143 if (atomic_dec_and_test(&ctx->ref))
144 complete(&ctx->comp);
147 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
149 struct ucma_context *ctx;
151 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
155 atomic_set(&ctx->ref, 1);
156 init_completion(&ctx->comp);
157 INIT_LIST_HEAD(&ctx->mc_list);
161 ctx->id = idr_alloc(&ctx_idr, ctx, 0, 0, GFP_KERNEL);
166 list_add_tail(&ctx->list, &file->ctx_list);
174 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
176 struct ucma_multicast *mc;
178 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
183 mc->id = idr_alloc(&multicast_idr, mc, 0, 0, GFP_KERNEL);
189 list_add_tail(&mc->list, &ctx->mc_list);
197 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
198 struct rdma_conn_param *src)
200 if (src->private_data_len)
201 memcpy(dst->private_data, src->private_data,
202 src->private_data_len);
203 dst->private_data_len = src->private_data_len;
204 dst->responder_resources =src->responder_resources;
205 dst->initiator_depth = src->initiator_depth;
206 dst->flow_control = src->flow_control;
207 dst->retry_count = src->retry_count;
208 dst->rnr_retry_count = src->rnr_retry_count;
210 dst->qp_num = src->qp_num;
213 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
214 struct rdma_ud_param *src)
216 if (src->private_data_len)
217 memcpy(dst->private_data, src->private_data,
218 src->private_data_len);
219 dst->private_data_len = src->private_data_len;
220 ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
221 dst->qp_num = src->qp_num;
222 dst->qkey = src->qkey;
225 static void ucma_set_event_context(struct ucma_context *ctx,
226 struct rdma_cm_event *event,
227 struct ucma_event *uevent)
230 switch (event->event) {
231 case RDMA_CM_EVENT_MULTICAST_JOIN:
232 case RDMA_CM_EVENT_MULTICAST_ERROR:
233 uevent->mc = (struct ucma_multicast *)
234 event->param.ud.private_data;
235 uevent->resp.uid = uevent->mc->uid;
236 uevent->resp.id = uevent->mc->id;
239 uevent->resp.uid = ctx->uid;
240 uevent->resp.id = ctx->id;
245 static int ucma_event_handler(struct rdma_cm_id *cm_id,
246 struct rdma_cm_event *event)
248 struct ucma_event *uevent;
249 struct ucma_context *ctx = cm_id->context;
252 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
254 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
256 mutex_lock(&ctx->file->mut);
257 uevent->cm_id = cm_id;
258 ucma_set_event_context(ctx, event, uevent);
259 uevent->resp.event = event->event;
260 uevent->resp.status = event->status;
261 if (cm_id->qp_type == IB_QPT_UD)
262 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
264 ucma_copy_conn_event(&uevent->resp.param.conn,
267 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
274 } else if (!ctx->uid) {
276 * We ignore events for new connections until userspace has set
277 * their context. This can only happen if an error occurs on a
278 * new connection before the user accepts it. This is okay,
279 * since the accept will just fail later.
285 list_add_tail(&uevent->list, &ctx->file->event_list);
286 wake_up_interruptible(&ctx->file->poll_wait);
288 mutex_unlock(&ctx->file->mut);
292 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
293 int in_len, int out_len)
295 struct ucma_context *ctx;
296 struct rdma_ucm_get_event cmd;
297 struct ucma_event *uevent;
300 if (out_len < sizeof uevent->resp)
303 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
306 mutex_lock(&file->mut);
307 while (list_empty(&file->event_list)) {
308 mutex_unlock(&file->mut);
310 if (file->filp->f_flags & O_NONBLOCK)
313 if (wait_event_interruptible(file->poll_wait,
314 !list_empty(&file->event_list)))
317 mutex_lock(&file->mut);
320 uevent = list_entry(file->event_list.next, struct ucma_event, list);
322 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
323 ctx = ucma_alloc_ctx(file);
328 uevent->ctx->backlog++;
329 ctx->cm_id = uevent->cm_id;
330 ctx->cm_id->context = ctx;
331 uevent->resp.id = ctx->id;
334 if (copy_to_user((void __user *)(unsigned long)cmd.response,
335 &uevent->resp, sizeof uevent->resp)) {
340 list_del(&uevent->list);
341 uevent->ctx->events_reported++;
343 uevent->mc->events_reported++;
346 mutex_unlock(&file->mut);
350 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
354 *qp_type = IB_QPT_RC;
358 *qp_type = IB_QPT_UD;
361 *qp_type = cmd->qp_type;
368 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
369 int in_len, int out_len)
371 struct rdma_ucm_create_id cmd;
372 struct rdma_ucm_create_id_resp resp;
373 struct ucma_context *ctx;
374 enum ib_qp_type qp_type;
377 if (out_len < sizeof(resp))
380 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
383 ret = ucma_get_qp_type(&cmd, &qp_type);
387 mutex_lock(&file->mut);
388 ctx = ucma_alloc_ctx(file);
389 mutex_unlock(&file->mut);
394 ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps, qp_type);
395 if (IS_ERR(ctx->cm_id)) {
396 ret = PTR_ERR(ctx->cm_id);
401 if (copy_to_user((void __user *)(unsigned long)cmd.response,
402 &resp, sizeof(resp))) {
409 rdma_destroy_id(ctx->cm_id);
412 idr_remove(&ctx_idr, ctx->id);
418 static void ucma_cleanup_multicast(struct ucma_context *ctx)
420 struct ucma_multicast *mc, *tmp;
423 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
425 idr_remove(&multicast_idr, mc->id);
431 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
433 struct ucma_event *uevent, *tmp;
435 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
436 if (uevent->mc != mc)
439 list_del(&uevent->list);
445 * We cannot hold file->mut when calling rdma_destroy_id() or we can
446 * deadlock. We also acquire file->mut in ucma_event_handler(), and
447 * rdma_destroy_id() will wait until all callbacks have completed.
449 static int ucma_free_ctx(struct ucma_context *ctx)
452 struct ucma_event *uevent, *tmp;
455 /* No new events will be generated after destroying the id. */
456 rdma_destroy_id(ctx->cm_id);
458 ucma_cleanup_multicast(ctx);
460 /* Cleanup events not yet reported to the user. */
461 mutex_lock(&ctx->file->mut);
462 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
463 if (uevent->ctx == ctx)
464 list_move_tail(&uevent->list, &list);
466 list_del(&ctx->list);
467 mutex_unlock(&ctx->file->mut);
469 list_for_each_entry_safe(uevent, tmp, &list, list) {
470 list_del(&uevent->list);
471 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
472 rdma_destroy_id(uevent->cm_id);
476 events_reported = ctx->events_reported;
478 return events_reported;
481 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
482 int in_len, int out_len)
484 struct rdma_ucm_destroy_id cmd;
485 struct rdma_ucm_destroy_id_resp resp;
486 struct ucma_context *ctx;
489 if (out_len < sizeof(resp))
492 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
496 ctx = _ucma_find_context(cmd.id, file);
498 idr_remove(&ctx_idr, ctx->id);
505 wait_for_completion(&ctx->comp);
506 resp.events_reported = ucma_free_ctx(ctx);
508 if (copy_to_user((void __user *)(unsigned long)cmd.response,
509 &resp, sizeof(resp)))
515 static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
516 int in_len, int out_len)
518 struct rdma_ucm_bind_ip cmd;
519 struct ucma_context *ctx;
522 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
525 ctx = ucma_get_ctx(file, cmd.id);
529 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
534 static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
535 int in_len, int out_len)
537 struct rdma_ucm_bind cmd;
538 struct sockaddr *addr;
539 struct ucma_context *ctx;
542 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
545 addr = (struct sockaddr *) &cmd.addr;
546 if (cmd.reserved || !cmd.addr_size || (cmd.addr_size != rdma_addr_size(addr)))
549 ctx = ucma_get_ctx(file, cmd.id);
553 ret = rdma_bind_addr(ctx->cm_id, addr);
558 static ssize_t ucma_resolve_ip(struct ucma_file *file,
559 const char __user *inbuf,
560 int in_len, int out_len)
562 struct rdma_ucm_resolve_ip cmd;
563 struct ucma_context *ctx;
566 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
569 ctx = ucma_get_ctx(file, cmd.id);
573 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
574 (struct sockaddr *) &cmd.dst_addr,
580 static ssize_t ucma_resolve_addr(struct ucma_file *file,
581 const char __user *inbuf,
582 int in_len, int out_len)
584 struct rdma_ucm_resolve_addr cmd;
585 struct sockaddr *src, *dst;
586 struct ucma_context *ctx;
589 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
592 src = (struct sockaddr *) &cmd.src_addr;
593 dst = (struct sockaddr *) &cmd.dst_addr;
594 if (cmd.reserved || (cmd.src_size && (cmd.src_size != rdma_addr_size(src))) ||
595 !cmd.dst_size || (cmd.dst_size != rdma_addr_size(dst)))
598 ctx = ucma_get_ctx(file, cmd.id);
602 ret = rdma_resolve_addr(ctx->cm_id, src, dst, cmd.timeout_ms);
607 static ssize_t ucma_resolve_route(struct ucma_file *file,
608 const char __user *inbuf,
609 int in_len, int out_len)
611 struct rdma_ucm_resolve_route cmd;
612 struct ucma_context *ctx;
615 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
618 ctx = ucma_get_ctx(file, cmd.id);
622 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
627 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
628 struct rdma_route *route)
630 struct rdma_dev_addr *dev_addr;
632 resp->num_paths = route->num_paths;
633 switch (route->num_paths) {
635 dev_addr = &route->addr.dev_addr;
636 rdma_addr_get_dgid(dev_addr,
637 (union ib_gid *) &resp->ib_route[0].dgid);
638 rdma_addr_get_sgid(dev_addr,
639 (union ib_gid *) &resp->ib_route[0].sgid);
640 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
643 ib_copy_path_rec_to_user(&resp->ib_route[1],
644 &route->path_rec[1]);
647 ib_copy_path_rec_to_user(&resp->ib_route[0],
648 &route->path_rec[0]);
655 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
656 struct rdma_route *route)
658 struct rdma_dev_addr *dev_addr;
659 struct net_device *dev;
662 resp->num_paths = route->num_paths;
663 switch (route->num_paths) {
665 dev_addr = &route->addr.dev_addr;
666 dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
668 vid = rdma_vlan_dev_vlan_id(dev);
672 iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid,
673 dev_addr->dst_dev_addr, vid);
674 iboe_addr_get_sgid(dev_addr,
675 (union ib_gid *) &resp->ib_route[0].sgid);
676 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
679 ib_copy_path_rec_to_user(&resp->ib_route[1],
680 &route->path_rec[1]);
683 ib_copy_path_rec_to_user(&resp->ib_route[0],
684 &route->path_rec[0]);
691 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
692 struct rdma_route *route)
694 struct rdma_dev_addr *dev_addr;
696 dev_addr = &route->addr.dev_addr;
697 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
698 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
701 static ssize_t ucma_query_route(struct ucma_file *file,
702 const char __user *inbuf,
703 int in_len, int out_len)
705 struct rdma_ucm_query cmd;
706 struct rdma_ucm_query_route_resp resp;
707 struct ucma_context *ctx;
708 struct sockaddr *addr;
711 if (out_len < sizeof(resp))
714 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
717 ctx = ucma_get_ctx(file, cmd.id);
721 memset(&resp, 0, sizeof resp);
722 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
723 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
724 sizeof(struct sockaddr_in) :
725 sizeof(struct sockaddr_in6));
726 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
727 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
728 sizeof(struct sockaddr_in) :
729 sizeof(struct sockaddr_in6));
730 if (!ctx->cm_id->device)
733 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
734 resp.port_num = ctx->cm_id->port_num;
735 switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
736 case RDMA_TRANSPORT_IB:
737 switch (rdma_port_get_link_layer(ctx->cm_id->device,
738 ctx->cm_id->port_num)) {
739 case IB_LINK_LAYER_INFINIBAND:
740 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
742 case IB_LINK_LAYER_ETHERNET:
743 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
749 case RDMA_TRANSPORT_IWARP:
750 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
757 if (copy_to_user((void __user *)(unsigned long)cmd.response,
758 &resp, sizeof(resp)))
765 static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
766 struct rdma_ucm_query_addr_resp *resp)
771 resp->node_guid = (__force __u64) cm_id->device->node_guid;
772 resp->port_num = cm_id->port_num;
773 resp->pkey = (__force __u16) cpu_to_be16(
774 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
777 static ssize_t ucma_query_addr(struct ucma_context *ctx,
778 void __user *response, int out_len)
780 struct rdma_ucm_query_addr_resp resp;
781 struct sockaddr *addr;
784 if (out_len < sizeof(resp))
787 memset(&resp, 0, sizeof resp);
789 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
790 resp.src_size = rdma_addr_size(addr);
791 memcpy(&resp.src_addr, addr, resp.src_size);
793 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
794 resp.dst_size = rdma_addr_size(addr);
795 memcpy(&resp.dst_addr, addr, resp.dst_size);
797 ucma_query_device_addr(ctx->cm_id, &resp);
799 if (copy_to_user(response, &resp, sizeof(resp)))
805 static ssize_t ucma_query_path(struct ucma_context *ctx,
806 void __user *response, int out_len)
808 struct rdma_ucm_query_path_resp *resp;
811 if (out_len < sizeof(*resp))
814 resp = kzalloc(out_len, GFP_KERNEL);
818 resp->num_paths = ctx->cm_id->route.num_paths;
819 for (i = 0, out_len -= sizeof(*resp);
820 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
821 i++, out_len -= sizeof(struct ib_path_rec_data)) {
823 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
824 IB_PATH_BIDIRECTIONAL;
825 ib_sa_pack_path(&ctx->cm_id->route.path_rec[i],
826 &resp->path_data[i].path_rec);
829 if (copy_to_user(response, resp,
830 sizeof(*resp) + (i * sizeof(struct ib_path_rec_data))))
837 static ssize_t ucma_query_gid(struct ucma_context *ctx,
838 void __user *response, int out_len)
840 struct rdma_ucm_query_addr_resp resp;
841 struct sockaddr_ib *addr;
844 if (out_len < sizeof(resp))
847 memset(&resp, 0, sizeof resp);
849 ucma_query_device_addr(ctx->cm_id, &resp);
851 addr = (struct sockaddr_ib *) &resp.src_addr;
852 resp.src_size = sizeof(*addr);
853 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
854 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
856 addr->sib_family = AF_IB;
857 addr->sib_pkey = (__force __be16) resp.pkey;
858 rdma_addr_get_sgid(&ctx->cm_id->route.addr.dev_addr,
859 (union ib_gid *) &addr->sib_addr);
860 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
861 &ctx->cm_id->route.addr.src_addr);
864 addr = (struct sockaddr_ib *) &resp.dst_addr;
865 resp.dst_size = sizeof(*addr);
866 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
867 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
869 addr->sib_family = AF_IB;
870 addr->sib_pkey = (__force __be16) resp.pkey;
871 rdma_addr_get_dgid(&ctx->cm_id->route.addr.dev_addr,
872 (union ib_gid *) &addr->sib_addr);
873 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
874 &ctx->cm_id->route.addr.dst_addr);
877 if (copy_to_user(response, &resp, sizeof(resp)))
883 static ssize_t ucma_query(struct ucma_file *file,
884 const char __user *inbuf,
885 int in_len, int out_len)
887 struct rdma_ucm_query cmd;
888 struct ucma_context *ctx;
889 void __user *response;
892 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
895 response = (void __user *)(unsigned long) cmd.response;
896 ctx = ucma_get_ctx(file, cmd.id);
900 switch (cmd.option) {
901 case RDMA_USER_CM_QUERY_ADDR:
902 ret = ucma_query_addr(ctx, response, out_len);
904 case RDMA_USER_CM_QUERY_PATH:
905 ret = ucma_query_path(ctx, response, out_len);
907 case RDMA_USER_CM_QUERY_GID:
908 ret = ucma_query_gid(ctx, response, out_len);
919 static void ucma_copy_conn_param(struct rdma_cm_id *id,
920 struct rdma_conn_param *dst,
921 struct rdma_ucm_conn_param *src)
923 dst->private_data = src->private_data;
924 dst->private_data_len = src->private_data_len;
925 dst->responder_resources =src->responder_resources;
926 dst->initiator_depth = src->initiator_depth;
927 dst->flow_control = src->flow_control;
928 dst->retry_count = src->retry_count;
929 dst->rnr_retry_count = src->rnr_retry_count;
931 dst->qp_num = src->qp_num;
932 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
935 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
936 int in_len, int out_len)
938 struct rdma_ucm_connect cmd;
939 struct rdma_conn_param conn_param;
940 struct ucma_context *ctx;
943 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
946 if (!cmd.conn_param.valid)
949 ctx = ucma_get_ctx(file, cmd.id);
953 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
954 ret = rdma_connect(ctx->cm_id, &conn_param);
959 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
960 int in_len, int out_len)
962 struct rdma_ucm_listen cmd;
963 struct ucma_context *ctx;
966 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
969 ctx = ucma_get_ctx(file, cmd.id);
973 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
974 cmd.backlog : max_backlog;
975 ret = rdma_listen(ctx->cm_id, ctx->backlog);
980 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
981 int in_len, int out_len)
983 struct rdma_ucm_accept cmd;
984 struct rdma_conn_param conn_param;
985 struct ucma_context *ctx;
988 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
991 ctx = ucma_get_ctx(file, cmd.id);
995 if (cmd.conn_param.valid) {
996 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
997 mutex_lock(&file->mut);
998 ret = rdma_accept(ctx->cm_id, &conn_param);
1001 mutex_unlock(&file->mut);
1003 ret = rdma_accept(ctx->cm_id, NULL);
1009 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1010 int in_len, int out_len)
1012 struct rdma_ucm_reject cmd;
1013 struct ucma_context *ctx;
1016 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1019 ctx = ucma_get_ctx(file, cmd.id);
1021 return PTR_ERR(ctx);
1023 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
1028 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1029 int in_len, int out_len)
1031 struct rdma_ucm_disconnect cmd;
1032 struct ucma_context *ctx;
1035 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1038 ctx = ucma_get_ctx(file, cmd.id);
1040 return PTR_ERR(ctx);
1042 ret = rdma_disconnect(ctx->cm_id);
1047 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1048 const char __user *inbuf,
1049 int in_len, int out_len)
1051 struct rdma_ucm_init_qp_attr cmd;
1052 struct ib_uverbs_qp_attr resp;
1053 struct ucma_context *ctx;
1054 struct ib_qp_attr qp_attr;
1057 if (out_len < sizeof(resp))
1060 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1063 ctx = ucma_get_ctx(file, cmd.id);
1065 return PTR_ERR(ctx);
1067 resp.qp_attr_mask = 0;
1068 memset(&qp_attr, 0, sizeof qp_attr);
1069 qp_attr.qp_state = cmd.qp_state;
1070 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1074 ib_copy_qp_attr_to_user(&resp, &qp_attr);
1075 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1076 &resp, sizeof(resp)))
1084 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1085 void *optval, size_t optlen)
1090 case RDMA_OPTION_ID_TOS:
1091 if (optlen != sizeof(u8)) {
1095 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1097 case RDMA_OPTION_ID_REUSEADDR:
1098 if (optlen != sizeof(int)) {
1102 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1104 case RDMA_OPTION_ID_AFONLY:
1105 if (optlen != sizeof(int)) {
1109 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1118 static int ucma_set_ib_path(struct ucma_context *ctx,
1119 struct ib_path_rec_data *path_data, size_t optlen)
1121 struct ib_sa_path_rec sa_path;
1122 struct rdma_cm_event event;
1125 if (optlen % sizeof(*path_data))
1128 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1129 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1130 IB_PATH_BIDIRECTIONAL))
1137 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1138 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
1142 memset(&event, 0, sizeof event);
1143 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1144 return ucma_event_handler(ctx->cm_id, &event);
1147 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1148 void *optval, size_t optlen)
1153 case RDMA_OPTION_IB_PATH:
1154 ret = ucma_set_ib_path(ctx, optval, optlen);
1163 static int ucma_set_option_level(struct ucma_context *ctx, int level,
1164 int optname, void *optval, size_t optlen)
1169 case RDMA_OPTION_ID:
1170 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1172 case RDMA_OPTION_IB:
1173 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1182 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1183 int in_len, int out_len)
1185 struct rdma_ucm_set_option cmd;
1186 struct ucma_context *ctx;
1190 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1193 ctx = ucma_get_ctx(file, cmd.id);
1195 return PTR_ERR(ctx);
1197 optval = memdup_user((void __user *) (unsigned long) cmd.optval,
1199 if (IS_ERR(optval)) {
1200 ret = PTR_ERR(optval);
1204 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1213 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1214 int in_len, int out_len)
1216 struct rdma_ucm_notify cmd;
1217 struct ucma_context *ctx;
1220 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1223 ctx = ucma_get_ctx(file, cmd.id);
1225 return PTR_ERR(ctx);
1227 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1232 static ssize_t ucma_process_join(struct ucma_file *file,
1233 struct rdma_ucm_join_mcast *cmd, int out_len)
1235 struct rdma_ucm_create_id_resp resp;
1236 struct ucma_context *ctx;
1237 struct ucma_multicast *mc;
1238 struct sockaddr *addr;
1241 if (out_len < sizeof(resp))
1244 addr = (struct sockaddr *) &cmd->addr;
1245 if (cmd->reserved || !cmd->addr_size || (cmd->addr_size != rdma_addr_size(addr)))
1248 ctx = ucma_get_ctx(file, cmd->id);
1250 return PTR_ERR(ctx);
1252 mutex_lock(&file->mut);
1253 mc = ucma_alloc_multicast(ctx);
1260 memcpy(&mc->addr, addr, cmd->addr_size);
1261 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1266 if (copy_to_user((void __user *)(unsigned long) cmd->response,
1267 &resp, sizeof(resp))) {
1272 mutex_unlock(&file->mut);
1277 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1278 ucma_cleanup_mc_events(mc);
1281 idr_remove(&multicast_idr, mc->id);
1283 list_del(&mc->list);
1286 mutex_unlock(&file->mut);
1291 static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1292 const char __user *inbuf,
1293 int in_len, int out_len)
1295 struct rdma_ucm_join_ip_mcast cmd;
1296 struct rdma_ucm_join_mcast join_cmd;
1298 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1301 join_cmd.response = cmd.response;
1302 join_cmd.uid = cmd.uid;
1303 join_cmd.id = cmd.id;
1304 join_cmd.addr_size = rdma_addr_size((struct sockaddr *) &cmd.addr);
1305 join_cmd.reserved = 0;
1306 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1308 return ucma_process_join(file, &join_cmd, out_len);
1311 static ssize_t ucma_join_multicast(struct ucma_file *file,
1312 const char __user *inbuf,
1313 int in_len, int out_len)
1315 struct rdma_ucm_join_mcast cmd;
1317 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1320 return ucma_process_join(file, &cmd, out_len);
1323 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1324 const char __user *inbuf,
1325 int in_len, int out_len)
1327 struct rdma_ucm_destroy_id cmd;
1328 struct rdma_ucm_destroy_id_resp resp;
1329 struct ucma_multicast *mc;
1332 if (out_len < sizeof(resp))
1335 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1339 mc = idr_find(&multicast_idr, cmd.id);
1341 mc = ERR_PTR(-ENOENT);
1342 else if (mc->ctx->file != file)
1343 mc = ERR_PTR(-EINVAL);
1345 idr_remove(&multicast_idr, mc->id);
1346 atomic_inc(&mc->ctx->ref);
1355 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1356 mutex_lock(&mc->ctx->file->mut);
1357 ucma_cleanup_mc_events(mc);
1358 list_del(&mc->list);
1359 mutex_unlock(&mc->ctx->file->mut);
1361 ucma_put_ctx(mc->ctx);
1362 resp.events_reported = mc->events_reported;
1365 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1366 &resp, sizeof(resp)))
1372 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1374 /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1375 if (file1 < file2) {
1376 mutex_lock(&file1->mut);
1377 mutex_lock(&file2->mut);
1379 mutex_lock(&file2->mut);
1380 mutex_lock(&file1->mut);
1384 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1386 if (file1 < file2) {
1387 mutex_unlock(&file2->mut);
1388 mutex_unlock(&file1->mut);
1390 mutex_unlock(&file1->mut);
1391 mutex_unlock(&file2->mut);
1395 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1397 struct ucma_event *uevent, *tmp;
1399 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1400 if (uevent->ctx == ctx)
1401 list_move_tail(&uevent->list, &file->event_list);
1404 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1405 const char __user *inbuf,
1406 int in_len, int out_len)
1408 struct rdma_ucm_migrate_id cmd;
1409 struct rdma_ucm_migrate_resp resp;
1410 struct ucma_context *ctx;
1412 struct ucma_file *cur_file;
1415 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1418 /* Get current fd to protect against it being closed */
1423 /* Validate current fd and prevent destruction of id. */
1424 ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1430 cur_file = ctx->file;
1431 if (cur_file == new_file) {
1432 resp.events_reported = ctx->events_reported;
1437 * Migrate events between fd's, maintaining order, and avoiding new
1438 * events being added before existing events.
1440 ucma_lock_files(cur_file, new_file);
1443 list_move_tail(&ctx->list, &new_file->ctx_list);
1444 ucma_move_events(ctx, new_file);
1445 ctx->file = new_file;
1446 resp.events_reported = ctx->events_reported;
1449 ucma_unlock_files(cur_file, new_file);
1452 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1453 &resp, sizeof(resp)))
1462 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1463 const char __user *inbuf,
1464 int in_len, int out_len) = {
1465 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1466 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1467 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1468 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1469 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1470 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1471 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1472 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1473 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1474 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1475 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1476 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1477 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1478 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1479 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1480 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1481 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1482 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1483 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1484 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1485 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1486 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1487 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1490 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1491 size_t len, loff_t *pos)
1493 struct ucma_file *file = filp->private_data;
1494 struct rdma_ucm_cmd_hdr hdr;
1497 if (len < sizeof(hdr))
1500 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1503 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1506 if (hdr.in + sizeof(hdr) > len)
1509 if (!ucma_cmd_table[hdr.cmd])
1512 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1519 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1521 struct ucma_file *file = filp->private_data;
1522 unsigned int mask = 0;
1524 poll_wait(filp, &file->poll_wait, wait);
1526 if (!list_empty(&file->event_list))
1527 mask = POLLIN | POLLRDNORM;
1533 * ucma_open() does not need the BKL:
1535 * - no global state is referred to;
1536 * - there is no ioctl method to race against;
1537 * - no further module initialization is required for open to work
1538 * after the device is registered.
1540 static int ucma_open(struct inode *inode, struct file *filp)
1542 struct ucma_file *file;
1544 file = kmalloc(sizeof *file, GFP_KERNEL);
1548 INIT_LIST_HEAD(&file->event_list);
1549 INIT_LIST_HEAD(&file->ctx_list);
1550 init_waitqueue_head(&file->poll_wait);
1551 mutex_init(&file->mut);
1553 filp->private_data = file;
1556 return nonseekable_open(inode, filp);
1559 static int ucma_close(struct inode *inode, struct file *filp)
1561 struct ucma_file *file = filp->private_data;
1562 struct ucma_context *ctx, *tmp;
1564 mutex_lock(&file->mut);
1565 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1566 mutex_unlock(&file->mut);
1569 idr_remove(&ctx_idr, ctx->id);
1573 mutex_lock(&file->mut);
1575 mutex_unlock(&file->mut);
1580 static const struct file_operations ucma_fops = {
1581 .owner = THIS_MODULE,
1583 .release = ucma_close,
1584 .write = ucma_write,
1586 .llseek = no_llseek,
1589 static struct miscdevice ucma_misc = {
1590 .minor = MISC_DYNAMIC_MINOR,
1592 .nodename = "infiniband/rdma_cm",
1597 static ssize_t show_abi_version(struct device *dev,
1598 struct device_attribute *attr,
1601 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1603 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1605 static int __init ucma_init(void)
1609 ret = misc_register(&ucma_misc);
1613 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1615 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1619 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1620 if (!ucma_ctl_table_hdr) {
1621 printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n");
1627 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1629 misc_deregister(&ucma_misc);
1633 static void __exit ucma_cleanup(void)
1635 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1636 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1637 misc_deregister(&ucma_misc);
1638 idr_destroy(&ctx_idr);
1641 module_init(ucma_init);
1642 module_exit(ucma_cleanup);