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 || ctx->cm_id != cm_id) {
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)
659 resp->num_paths = route->num_paths;
660 switch (route->num_paths) {
662 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
663 (union ib_gid *)&resp->ib_route[0].dgid);
664 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
665 (union ib_gid *)&resp->ib_route[0].sgid);
666 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
669 ib_copy_path_rec_to_user(&resp->ib_route[1],
670 &route->path_rec[1]);
673 ib_copy_path_rec_to_user(&resp->ib_route[0],
674 &route->path_rec[0]);
681 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
682 struct rdma_route *route)
684 struct rdma_dev_addr *dev_addr;
686 dev_addr = &route->addr.dev_addr;
687 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
688 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
691 static ssize_t ucma_query_route(struct ucma_file *file,
692 const char __user *inbuf,
693 int in_len, int out_len)
695 struct rdma_ucm_query cmd;
696 struct rdma_ucm_query_route_resp resp;
697 struct ucma_context *ctx;
698 struct sockaddr *addr;
701 if (out_len < sizeof(resp))
704 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
707 ctx = ucma_get_ctx(file, cmd.id);
711 memset(&resp, 0, sizeof resp);
712 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
713 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
714 sizeof(struct sockaddr_in) :
715 sizeof(struct sockaddr_in6));
716 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
717 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
718 sizeof(struct sockaddr_in) :
719 sizeof(struct sockaddr_in6));
720 if (!ctx->cm_id->device)
723 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
724 resp.port_num = ctx->cm_id->port_num;
725 switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
726 case RDMA_TRANSPORT_IB:
727 switch (rdma_port_get_link_layer(ctx->cm_id->device,
728 ctx->cm_id->port_num)) {
729 case IB_LINK_LAYER_INFINIBAND:
730 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
732 case IB_LINK_LAYER_ETHERNET:
733 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
739 case RDMA_TRANSPORT_IWARP:
740 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
747 if (copy_to_user((void __user *)(unsigned long)cmd.response,
748 &resp, sizeof(resp)))
755 static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
756 struct rdma_ucm_query_addr_resp *resp)
761 resp->node_guid = (__force __u64) cm_id->device->node_guid;
762 resp->port_num = cm_id->port_num;
763 resp->pkey = (__force __u16) cpu_to_be16(
764 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
767 static ssize_t ucma_query_addr(struct ucma_context *ctx,
768 void __user *response, int out_len)
770 struct rdma_ucm_query_addr_resp resp;
771 struct sockaddr *addr;
774 if (out_len < sizeof(resp))
777 memset(&resp, 0, sizeof resp);
779 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
780 resp.src_size = rdma_addr_size(addr);
781 memcpy(&resp.src_addr, addr, resp.src_size);
783 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
784 resp.dst_size = rdma_addr_size(addr);
785 memcpy(&resp.dst_addr, addr, resp.dst_size);
787 ucma_query_device_addr(ctx->cm_id, &resp);
789 if (copy_to_user(response, &resp, sizeof(resp)))
795 static ssize_t ucma_query_path(struct ucma_context *ctx,
796 void __user *response, int out_len)
798 struct rdma_ucm_query_path_resp *resp;
801 if (out_len < sizeof(*resp))
804 resp = kzalloc(out_len, GFP_KERNEL);
808 resp->num_paths = ctx->cm_id->route.num_paths;
809 for (i = 0, out_len -= sizeof(*resp);
810 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
811 i++, out_len -= sizeof(struct ib_path_rec_data)) {
813 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
814 IB_PATH_BIDIRECTIONAL;
815 ib_sa_pack_path(&ctx->cm_id->route.path_rec[i],
816 &resp->path_data[i].path_rec);
819 if (copy_to_user(response, resp,
820 sizeof(*resp) + (i * sizeof(struct ib_path_rec_data))))
827 static ssize_t ucma_query_gid(struct ucma_context *ctx,
828 void __user *response, int out_len)
830 struct rdma_ucm_query_addr_resp resp;
831 struct sockaddr_ib *addr;
834 if (out_len < sizeof(resp))
837 memset(&resp, 0, sizeof resp);
839 ucma_query_device_addr(ctx->cm_id, &resp);
841 addr = (struct sockaddr_ib *) &resp.src_addr;
842 resp.src_size = sizeof(*addr);
843 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
844 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
846 addr->sib_family = AF_IB;
847 addr->sib_pkey = (__force __be16) resp.pkey;
848 rdma_addr_get_sgid(&ctx->cm_id->route.addr.dev_addr,
849 (union ib_gid *) &addr->sib_addr);
850 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
851 &ctx->cm_id->route.addr.src_addr);
854 addr = (struct sockaddr_ib *) &resp.dst_addr;
855 resp.dst_size = sizeof(*addr);
856 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
857 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
859 addr->sib_family = AF_IB;
860 addr->sib_pkey = (__force __be16) resp.pkey;
861 rdma_addr_get_dgid(&ctx->cm_id->route.addr.dev_addr,
862 (union ib_gid *) &addr->sib_addr);
863 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
864 &ctx->cm_id->route.addr.dst_addr);
867 if (copy_to_user(response, &resp, sizeof(resp)))
873 static ssize_t ucma_query(struct ucma_file *file,
874 const char __user *inbuf,
875 int in_len, int out_len)
877 struct rdma_ucm_query cmd;
878 struct ucma_context *ctx;
879 void __user *response;
882 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
885 response = (void __user *)(unsigned long) cmd.response;
886 ctx = ucma_get_ctx(file, cmd.id);
890 switch (cmd.option) {
891 case RDMA_USER_CM_QUERY_ADDR:
892 ret = ucma_query_addr(ctx, response, out_len);
894 case RDMA_USER_CM_QUERY_PATH:
895 ret = ucma_query_path(ctx, response, out_len);
897 case RDMA_USER_CM_QUERY_GID:
898 ret = ucma_query_gid(ctx, response, out_len);
909 static void ucma_copy_conn_param(struct rdma_cm_id *id,
910 struct rdma_conn_param *dst,
911 struct rdma_ucm_conn_param *src)
913 dst->private_data = src->private_data;
914 dst->private_data_len = src->private_data_len;
915 dst->responder_resources =src->responder_resources;
916 dst->initiator_depth = src->initiator_depth;
917 dst->flow_control = src->flow_control;
918 dst->retry_count = src->retry_count;
919 dst->rnr_retry_count = src->rnr_retry_count;
921 dst->qp_num = src->qp_num;
922 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
925 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
926 int in_len, int out_len)
928 struct rdma_ucm_connect cmd;
929 struct rdma_conn_param conn_param;
930 struct ucma_context *ctx;
933 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
936 if (!cmd.conn_param.valid)
939 ctx = ucma_get_ctx(file, cmd.id);
943 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
944 ret = rdma_connect(ctx->cm_id, &conn_param);
949 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
950 int in_len, int out_len)
952 struct rdma_ucm_listen cmd;
953 struct ucma_context *ctx;
956 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
959 ctx = ucma_get_ctx(file, cmd.id);
963 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
964 cmd.backlog : max_backlog;
965 ret = rdma_listen(ctx->cm_id, ctx->backlog);
970 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
971 int in_len, int out_len)
973 struct rdma_ucm_accept cmd;
974 struct rdma_conn_param conn_param;
975 struct ucma_context *ctx;
978 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
981 ctx = ucma_get_ctx(file, cmd.id);
985 if (cmd.conn_param.valid) {
986 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
987 mutex_lock(&file->mut);
988 ret = rdma_accept(ctx->cm_id, &conn_param);
991 mutex_unlock(&file->mut);
993 ret = rdma_accept(ctx->cm_id, NULL);
999 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1000 int in_len, int out_len)
1002 struct rdma_ucm_reject cmd;
1003 struct ucma_context *ctx;
1006 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1009 ctx = ucma_get_ctx(file, cmd.id);
1011 return PTR_ERR(ctx);
1013 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
1018 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1019 int in_len, int out_len)
1021 struct rdma_ucm_disconnect cmd;
1022 struct ucma_context *ctx;
1025 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1028 ctx = ucma_get_ctx(file, cmd.id);
1030 return PTR_ERR(ctx);
1032 ret = rdma_disconnect(ctx->cm_id);
1037 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1038 const char __user *inbuf,
1039 int in_len, int out_len)
1041 struct rdma_ucm_init_qp_attr cmd;
1042 struct ib_uverbs_qp_attr resp;
1043 struct ucma_context *ctx;
1044 struct ib_qp_attr qp_attr;
1047 if (out_len < sizeof(resp))
1050 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1053 ctx = ucma_get_ctx(file, cmd.id);
1055 return PTR_ERR(ctx);
1057 resp.qp_attr_mask = 0;
1058 memset(&qp_attr, 0, sizeof qp_attr);
1059 qp_attr.qp_state = cmd.qp_state;
1060 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1064 ib_copy_qp_attr_to_user(&resp, &qp_attr);
1065 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1066 &resp, sizeof(resp)))
1074 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1075 void *optval, size_t optlen)
1080 case RDMA_OPTION_ID_TOS:
1081 if (optlen != sizeof(u8)) {
1085 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1087 case RDMA_OPTION_ID_REUSEADDR:
1088 if (optlen != sizeof(int)) {
1092 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1094 case RDMA_OPTION_ID_AFONLY:
1095 if (optlen != sizeof(int)) {
1099 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1108 static int ucma_set_ib_path(struct ucma_context *ctx,
1109 struct ib_path_rec_data *path_data, size_t optlen)
1111 struct ib_sa_path_rec sa_path;
1112 struct rdma_cm_event event;
1115 if (optlen % sizeof(*path_data))
1118 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1119 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1120 IB_PATH_BIDIRECTIONAL))
1127 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1128 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
1132 memset(&event, 0, sizeof event);
1133 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1134 return ucma_event_handler(ctx->cm_id, &event);
1137 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1138 void *optval, size_t optlen)
1143 case RDMA_OPTION_IB_PATH:
1144 ret = ucma_set_ib_path(ctx, optval, optlen);
1153 static int ucma_set_option_level(struct ucma_context *ctx, int level,
1154 int optname, void *optval, size_t optlen)
1159 case RDMA_OPTION_ID:
1160 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1162 case RDMA_OPTION_IB:
1163 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1172 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1173 int in_len, int out_len)
1175 struct rdma_ucm_set_option cmd;
1176 struct ucma_context *ctx;
1180 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1183 ctx = ucma_get_ctx(file, cmd.id);
1185 return PTR_ERR(ctx);
1187 optval = memdup_user((void __user *) (unsigned long) cmd.optval,
1189 if (IS_ERR(optval)) {
1190 ret = PTR_ERR(optval);
1194 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1203 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1204 int in_len, int out_len)
1206 struct rdma_ucm_notify cmd;
1207 struct ucma_context *ctx;
1210 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1213 ctx = ucma_get_ctx(file, cmd.id);
1215 return PTR_ERR(ctx);
1217 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1222 static ssize_t ucma_process_join(struct ucma_file *file,
1223 struct rdma_ucm_join_mcast *cmd, int out_len)
1225 struct rdma_ucm_create_id_resp resp;
1226 struct ucma_context *ctx;
1227 struct ucma_multicast *mc;
1228 struct sockaddr *addr;
1231 if (out_len < sizeof(resp))
1234 addr = (struct sockaddr *) &cmd->addr;
1235 if (cmd->reserved || !cmd->addr_size || (cmd->addr_size != rdma_addr_size(addr)))
1238 ctx = ucma_get_ctx(file, cmd->id);
1240 return PTR_ERR(ctx);
1242 mutex_lock(&file->mut);
1243 mc = ucma_alloc_multicast(ctx);
1250 memcpy(&mc->addr, addr, cmd->addr_size);
1251 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1256 if (copy_to_user((void __user *)(unsigned long) cmd->response,
1257 &resp, sizeof(resp))) {
1262 mutex_unlock(&file->mut);
1267 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1268 ucma_cleanup_mc_events(mc);
1271 idr_remove(&multicast_idr, mc->id);
1273 list_del(&mc->list);
1276 mutex_unlock(&file->mut);
1281 static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1282 const char __user *inbuf,
1283 int in_len, int out_len)
1285 struct rdma_ucm_join_ip_mcast cmd;
1286 struct rdma_ucm_join_mcast join_cmd;
1288 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1291 join_cmd.response = cmd.response;
1292 join_cmd.uid = cmd.uid;
1293 join_cmd.id = cmd.id;
1294 join_cmd.addr_size = rdma_addr_size((struct sockaddr *) &cmd.addr);
1295 join_cmd.reserved = 0;
1296 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1298 return ucma_process_join(file, &join_cmd, out_len);
1301 static ssize_t ucma_join_multicast(struct ucma_file *file,
1302 const char __user *inbuf,
1303 int in_len, int out_len)
1305 struct rdma_ucm_join_mcast cmd;
1307 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1310 return ucma_process_join(file, &cmd, out_len);
1313 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1314 const char __user *inbuf,
1315 int in_len, int out_len)
1317 struct rdma_ucm_destroy_id cmd;
1318 struct rdma_ucm_destroy_id_resp resp;
1319 struct ucma_multicast *mc;
1322 if (out_len < sizeof(resp))
1325 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1329 mc = idr_find(&multicast_idr, cmd.id);
1331 mc = ERR_PTR(-ENOENT);
1332 else if (mc->ctx->file != file)
1333 mc = ERR_PTR(-EINVAL);
1335 idr_remove(&multicast_idr, mc->id);
1336 atomic_inc(&mc->ctx->ref);
1345 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1346 mutex_lock(&mc->ctx->file->mut);
1347 ucma_cleanup_mc_events(mc);
1348 list_del(&mc->list);
1349 mutex_unlock(&mc->ctx->file->mut);
1351 ucma_put_ctx(mc->ctx);
1352 resp.events_reported = mc->events_reported;
1355 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1356 &resp, sizeof(resp)))
1362 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1364 /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1365 if (file1 < file2) {
1366 mutex_lock(&file1->mut);
1367 mutex_lock(&file2->mut);
1369 mutex_lock(&file2->mut);
1370 mutex_lock(&file1->mut);
1374 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1376 if (file1 < file2) {
1377 mutex_unlock(&file2->mut);
1378 mutex_unlock(&file1->mut);
1380 mutex_unlock(&file1->mut);
1381 mutex_unlock(&file2->mut);
1385 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1387 struct ucma_event *uevent, *tmp;
1389 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1390 if (uevent->ctx == ctx)
1391 list_move_tail(&uevent->list, &file->event_list);
1394 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1395 const char __user *inbuf,
1396 int in_len, int out_len)
1398 struct rdma_ucm_migrate_id cmd;
1399 struct rdma_ucm_migrate_resp resp;
1400 struct ucma_context *ctx;
1402 struct ucma_file *cur_file;
1405 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1408 /* Get current fd to protect against it being closed */
1413 /* Validate current fd and prevent destruction of id. */
1414 ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1420 cur_file = ctx->file;
1421 if (cur_file == new_file) {
1422 resp.events_reported = ctx->events_reported;
1427 * Migrate events between fd's, maintaining order, and avoiding new
1428 * events being added before existing events.
1430 ucma_lock_files(cur_file, new_file);
1433 list_move_tail(&ctx->list, &new_file->ctx_list);
1434 ucma_move_events(ctx, new_file);
1435 ctx->file = new_file;
1436 resp.events_reported = ctx->events_reported;
1439 ucma_unlock_files(cur_file, new_file);
1442 if (copy_to_user((void __user *)(unsigned long)cmd.response,
1443 &resp, sizeof(resp)))
1452 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1453 const char __user *inbuf,
1454 int in_len, int out_len) = {
1455 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1456 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1457 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1458 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1459 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1460 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1461 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1462 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1463 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1464 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1465 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1466 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1467 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1468 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1469 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1470 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1471 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1472 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1473 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1474 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1475 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1476 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1477 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1480 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1481 size_t len, loff_t *pos)
1483 struct ucma_file *file = filp->private_data;
1484 struct rdma_ucm_cmd_hdr hdr;
1487 if (len < sizeof(hdr))
1490 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1493 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1496 if (hdr.in + sizeof(hdr) > len)
1499 if (!ucma_cmd_table[hdr.cmd])
1502 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1509 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1511 struct ucma_file *file = filp->private_data;
1512 unsigned int mask = 0;
1514 poll_wait(filp, &file->poll_wait, wait);
1516 if (!list_empty(&file->event_list))
1517 mask = POLLIN | POLLRDNORM;
1523 * ucma_open() does not need the BKL:
1525 * - no global state is referred to;
1526 * - there is no ioctl method to race against;
1527 * - no further module initialization is required for open to work
1528 * after the device is registered.
1530 static int ucma_open(struct inode *inode, struct file *filp)
1532 struct ucma_file *file;
1534 file = kmalloc(sizeof *file, GFP_KERNEL);
1538 INIT_LIST_HEAD(&file->event_list);
1539 INIT_LIST_HEAD(&file->ctx_list);
1540 init_waitqueue_head(&file->poll_wait);
1541 mutex_init(&file->mut);
1543 filp->private_data = file;
1546 return nonseekable_open(inode, filp);
1549 static int ucma_close(struct inode *inode, struct file *filp)
1551 struct ucma_file *file = filp->private_data;
1552 struct ucma_context *ctx, *tmp;
1554 mutex_lock(&file->mut);
1555 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1556 mutex_unlock(&file->mut);
1559 idr_remove(&ctx_idr, ctx->id);
1563 mutex_lock(&file->mut);
1565 mutex_unlock(&file->mut);
1570 static const struct file_operations ucma_fops = {
1571 .owner = THIS_MODULE,
1573 .release = ucma_close,
1574 .write = ucma_write,
1576 .llseek = no_llseek,
1579 static struct miscdevice ucma_misc = {
1580 .minor = MISC_DYNAMIC_MINOR,
1582 .nodename = "infiniband/rdma_cm",
1587 static ssize_t show_abi_version(struct device *dev,
1588 struct device_attribute *attr,
1591 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1593 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1595 static int __init ucma_init(void)
1599 ret = misc_register(&ucma_misc);
1603 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1605 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1609 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1610 if (!ucma_ctl_table_hdr) {
1611 printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n");
1617 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1619 misc_deregister(&ucma_misc);
1623 static void __exit ucma_cleanup(void)
1625 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1626 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1627 misc_deregister(&ucma_misc);
1628 idr_destroy(&ctx_idr);
1631 module_init(ucma_init);
1632 module_exit(ucma_cleanup);