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>
45 #include <linux/nsproxy.h>
47 #include <linux/nospec.h>
49 #include <rdma/rdma_user_cm.h>
50 #include <rdma/ib_marshall.h>
51 #include <rdma/rdma_cm.h>
52 #include <rdma/rdma_cm_ib.h>
53 #include <rdma/ib_addr.h>
56 MODULE_AUTHOR("Sean Hefty");
57 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
58 MODULE_LICENSE("Dual BSD/GPL");
60 static unsigned int max_backlog = 1024;
62 static struct ctl_table_header *ucma_ctl_table_hdr;
63 static struct ctl_table ucma_ctl_table[] = {
65 .procname = "max_backlog",
67 .maxlen = sizeof max_backlog,
69 .proc_handler = proc_dointvec,
77 struct list_head ctx_list;
78 struct list_head event_list;
79 wait_queue_head_t poll_wait;
80 struct workqueue_struct *close_wq;
85 struct completion comp;
90 struct ucma_file *file;
91 struct rdma_cm_id *cm_id;
95 struct list_head list;
96 struct list_head mc_list;
97 /* mark that device is in process of destroying the internal HW
98 * resources, protected by the global mut
101 /* sync between removal event and id destroy, protected by file mut */
103 struct work_struct close_work;
106 struct ucma_multicast {
107 struct ucma_context *ctx;
113 struct list_head list;
114 struct sockaddr_storage addr;
118 struct ucma_context *ctx;
119 struct ucma_multicast *mc;
120 struct list_head list;
121 struct rdma_cm_id *cm_id;
122 struct rdma_ucm_event_resp resp;
123 struct work_struct close_work;
126 static DEFINE_MUTEX(mut);
127 static DEFINE_IDR(ctx_idr);
128 static DEFINE_IDR(multicast_idr);
130 static const struct file_operations ucma_fops;
132 static inline struct ucma_context *_ucma_find_context(int id,
133 struct ucma_file *file)
135 struct ucma_context *ctx;
137 ctx = idr_find(&ctx_idr, id);
139 ctx = ERR_PTR(-ENOENT);
140 else if (ctx->file != file || !ctx->cm_id)
141 ctx = ERR_PTR(-EINVAL);
145 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
147 struct ucma_context *ctx;
150 ctx = _ucma_find_context(id, file);
155 atomic_inc(&ctx->ref);
161 static void ucma_put_ctx(struct ucma_context *ctx)
163 if (atomic_dec_and_test(&ctx->ref))
164 complete(&ctx->comp);
168 * Same as ucm_get_ctx but requires that ->cm_id->device is valid, eg that the
171 static struct ucma_context *ucma_get_ctx_dev(struct ucma_file *file, int id)
173 struct ucma_context *ctx = ucma_get_ctx(file, id);
177 if (!ctx->cm_id->device) {
179 return ERR_PTR(-EINVAL);
184 static void ucma_close_event_id(struct work_struct *work)
186 struct ucma_event *uevent_close = container_of(work, struct ucma_event, close_work);
188 rdma_destroy_id(uevent_close->cm_id);
192 static void ucma_close_id(struct work_struct *work)
194 struct ucma_context *ctx = container_of(work, struct ucma_context, close_work);
196 /* once all inflight tasks are finished, we close all underlying
197 * resources. The context is still alive till its explicit destryoing
201 wait_for_completion(&ctx->comp);
202 /* No new events will be generated after destroying the id. */
203 rdma_destroy_id(ctx->cm_id);
206 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
208 struct ucma_context *ctx;
210 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
214 INIT_WORK(&ctx->close_work, ucma_close_id);
215 atomic_set(&ctx->ref, 1);
216 init_completion(&ctx->comp);
217 INIT_LIST_HEAD(&ctx->mc_list);
219 mutex_init(&ctx->mutex);
222 ctx->id = idr_alloc(&ctx_idr, ctx, 0, 0, GFP_KERNEL);
227 list_add_tail(&ctx->list, &file->ctx_list);
235 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
237 struct ucma_multicast *mc;
239 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
244 mc->id = idr_alloc(&multicast_idr, NULL, 0, 0, GFP_KERNEL);
250 list_add_tail(&mc->list, &ctx->mc_list);
258 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
259 struct rdma_conn_param *src)
261 if (src->private_data_len)
262 memcpy(dst->private_data, src->private_data,
263 src->private_data_len);
264 dst->private_data_len = src->private_data_len;
265 dst->responder_resources =src->responder_resources;
266 dst->initiator_depth = src->initiator_depth;
267 dst->flow_control = src->flow_control;
268 dst->retry_count = src->retry_count;
269 dst->rnr_retry_count = src->rnr_retry_count;
271 dst->qp_num = src->qp_num;
274 static void ucma_copy_ud_event(struct ib_device *device,
275 struct rdma_ucm_ud_param *dst,
276 struct rdma_ud_param *src)
278 if (src->private_data_len)
279 memcpy(dst->private_data, src->private_data,
280 src->private_data_len);
281 dst->private_data_len = src->private_data_len;
282 ib_copy_ah_attr_to_user(device, &dst->ah_attr, &src->ah_attr);
283 dst->qp_num = src->qp_num;
284 dst->qkey = src->qkey;
287 static void ucma_set_event_context(struct ucma_context *ctx,
288 struct rdma_cm_event *event,
289 struct ucma_event *uevent)
292 switch (event->event) {
293 case RDMA_CM_EVENT_MULTICAST_JOIN:
294 case RDMA_CM_EVENT_MULTICAST_ERROR:
295 uevent->mc = (struct ucma_multicast *)
296 event->param.ud.private_data;
297 uevent->resp.uid = uevent->mc->uid;
298 uevent->resp.id = uevent->mc->id;
301 uevent->resp.uid = ctx->uid;
302 uevent->resp.id = ctx->id;
307 /* Called with file->mut locked for the relevant context. */
308 static void ucma_removal_event_handler(struct rdma_cm_id *cm_id)
310 struct ucma_context *ctx = cm_id->context;
311 struct ucma_event *con_req_eve;
317 /* only if context is pointing to cm_id that it owns it and can be
318 * queued to be closed, otherwise that cm_id is an inflight one that
319 * is part of that context event list pending to be detached and
320 * reattached to its new context as part of ucma_get_event,
321 * handled separately below.
323 if (ctx->cm_id == cm_id) {
327 queue_work(ctx->file->close_wq, &ctx->close_work);
331 list_for_each_entry(con_req_eve, &ctx->file->event_list, list) {
332 if (con_req_eve->cm_id == cm_id &&
333 con_req_eve->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
334 list_del(&con_req_eve->list);
335 INIT_WORK(&con_req_eve->close_work, ucma_close_event_id);
336 queue_work(ctx->file->close_wq, &con_req_eve->close_work);
342 pr_err("ucma_removal_event_handler: warning: connect request event wasn't found\n");
345 static int ucma_event_handler(struct rdma_cm_id *cm_id,
346 struct rdma_cm_event *event)
348 struct ucma_event *uevent;
349 struct ucma_context *ctx = cm_id->context;
352 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
354 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
356 mutex_lock(&ctx->file->mut);
357 uevent->cm_id = cm_id;
358 ucma_set_event_context(ctx, event, uevent);
359 uevent->resp.event = event->event;
360 uevent->resp.status = event->status;
361 if (cm_id->qp_type == IB_QPT_UD)
362 ucma_copy_ud_event(cm_id->device, &uevent->resp.param.ud,
365 ucma_copy_conn_event(&uevent->resp.param.conn,
368 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
375 } else if (!ctx->uid || ctx->cm_id != cm_id) {
377 * We ignore events for new connections until userspace has set
378 * their context. This can only happen if an error occurs on a
379 * new connection before the user accepts it. This is okay,
380 * since the accept will just fail later. However, we do need
381 * to release the underlying HW resources in case of a device
384 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
385 ucma_removal_event_handler(cm_id);
391 list_add_tail(&uevent->list, &ctx->file->event_list);
392 wake_up_interruptible(&ctx->file->poll_wait);
393 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
394 ucma_removal_event_handler(cm_id);
396 mutex_unlock(&ctx->file->mut);
400 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
401 int in_len, int out_len)
403 struct ucma_context *ctx;
404 struct rdma_ucm_get_event cmd;
405 struct ucma_event *uevent;
409 * Old 32 bit user space does not send the 4 byte padding in the
410 * reserved field. We don't care, allow it to keep working.
412 if (out_len < sizeof(uevent->resp) - sizeof(uevent->resp.reserved))
415 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
418 mutex_lock(&file->mut);
419 while (list_empty(&file->event_list)) {
420 mutex_unlock(&file->mut);
422 if (file->filp->f_flags & O_NONBLOCK)
425 if (wait_event_interruptible(file->poll_wait,
426 !list_empty(&file->event_list)))
429 mutex_lock(&file->mut);
432 uevent = list_entry(file->event_list.next, struct ucma_event, list);
434 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
435 ctx = ucma_alloc_ctx(file);
440 uevent->ctx->backlog++;
441 ctx->cm_id = uevent->cm_id;
442 ctx->cm_id->context = ctx;
443 uevent->resp.id = ctx->id;
446 if (copy_to_user(u64_to_user_ptr(cmd.response),
448 min_t(size_t, out_len, sizeof(uevent->resp)))) {
453 list_del(&uevent->list);
454 uevent->ctx->events_reported++;
456 uevent->mc->events_reported++;
459 mutex_unlock(&file->mut);
463 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
467 *qp_type = IB_QPT_RC;
471 *qp_type = IB_QPT_UD;
474 *qp_type = cmd->qp_type;
481 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
482 int in_len, int out_len)
484 struct rdma_ucm_create_id cmd;
485 struct rdma_ucm_create_id_resp resp;
486 struct ucma_context *ctx;
487 struct rdma_cm_id *cm_id;
488 enum ib_qp_type qp_type;
491 if (out_len < sizeof(resp))
494 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
497 ret = ucma_get_qp_type(&cmd, &qp_type);
501 mutex_lock(&file->mut);
502 ctx = ucma_alloc_ctx(file);
503 mutex_unlock(&file->mut);
508 cm_id = __rdma_create_id(current->nsproxy->net_ns,
509 ucma_event_handler, ctx, cmd.ps, qp_type, NULL);
511 ret = PTR_ERR(cm_id);
516 if (copy_to_user(u64_to_user_ptr(cmd.response),
517 &resp, sizeof(resp))) {
526 rdma_destroy_id(cm_id);
529 idr_remove(&ctx_idr, ctx->id);
531 mutex_lock(&file->mut);
532 list_del(&ctx->list);
533 mutex_unlock(&file->mut);
538 static void ucma_cleanup_multicast(struct ucma_context *ctx)
540 struct ucma_multicast *mc, *tmp;
543 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
545 idr_remove(&multicast_idr, mc->id);
551 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
553 struct ucma_event *uevent, *tmp;
555 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
556 if (uevent->mc != mc)
559 list_del(&uevent->list);
565 * ucma_free_ctx is called after the underlying rdma CM-ID is destroyed. At
566 * this point, no new events will be reported from the hardware. However, we
567 * still need to cleanup the UCMA context for this ID. Specifically, there
568 * might be events that have not yet been consumed by the user space software.
569 * These might include pending connect requests which we have not completed
570 * processing. We cannot call rdma_destroy_id while holding the lock of the
571 * context (file->mut), as it might cause a deadlock. We therefore extract all
572 * relevant events from the context pending events list while holding the
573 * mutex. After that we release them as needed.
575 static int ucma_free_ctx(struct ucma_context *ctx)
578 struct ucma_event *uevent, *tmp;
582 ucma_cleanup_multicast(ctx);
584 /* Cleanup events not yet reported to the user. */
585 mutex_lock(&ctx->file->mut);
586 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
587 if (uevent->ctx == ctx)
588 list_move_tail(&uevent->list, &list);
590 list_del(&ctx->list);
591 events_reported = ctx->events_reported;
592 mutex_unlock(&ctx->file->mut);
594 list_for_each_entry_safe(uevent, tmp, &list, list) {
595 list_del(&uevent->list);
596 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
597 rdma_destroy_id(uevent->cm_id);
601 mutex_destroy(&ctx->mutex);
603 return events_reported;
606 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
607 int in_len, int out_len)
609 struct rdma_ucm_destroy_id cmd;
610 struct rdma_ucm_destroy_id_resp resp;
611 struct ucma_context *ctx;
614 if (out_len < sizeof(resp))
617 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
621 ctx = _ucma_find_context(cmd.id, file);
623 idr_remove(&ctx_idr, ctx->id);
629 mutex_lock(&ctx->file->mut);
631 mutex_unlock(&ctx->file->mut);
633 flush_workqueue(ctx->file->close_wq);
634 /* At this point it's guaranteed that there is no inflight
640 wait_for_completion(&ctx->comp);
641 rdma_destroy_id(ctx->cm_id);
646 resp.events_reported = ucma_free_ctx(ctx);
647 if (copy_to_user(u64_to_user_ptr(cmd.response),
648 &resp, sizeof(resp)))
654 static ssize_t ucma_bind_ip(struct ucma_file *file, const char __user *inbuf,
655 int in_len, int out_len)
657 struct rdma_ucm_bind_ip cmd;
658 struct ucma_context *ctx;
661 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
664 if (!rdma_addr_size_in6(&cmd.addr))
667 ctx = ucma_get_ctx(file, cmd.id);
671 mutex_lock(&ctx->mutex);
672 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
673 mutex_unlock(&ctx->mutex);
679 static ssize_t ucma_bind(struct ucma_file *file, const char __user *inbuf,
680 int in_len, int out_len)
682 struct rdma_ucm_bind cmd;
683 struct ucma_context *ctx;
686 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
689 if (cmd.reserved || !cmd.addr_size ||
690 cmd.addr_size != rdma_addr_size_kss(&cmd.addr))
693 ctx = ucma_get_ctx(file, cmd.id);
697 mutex_lock(&ctx->mutex);
698 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
699 mutex_unlock(&ctx->mutex);
704 static ssize_t ucma_resolve_ip(struct ucma_file *file,
705 const char __user *inbuf,
706 int in_len, int out_len)
708 struct rdma_ucm_resolve_ip cmd;
709 struct ucma_context *ctx;
712 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
715 if ((cmd.src_addr.sin6_family && !rdma_addr_size_in6(&cmd.src_addr)) ||
716 !rdma_addr_size_in6(&cmd.dst_addr))
719 ctx = ucma_get_ctx(file, cmd.id);
723 mutex_lock(&ctx->mutex);
724 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
725 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
726 mutex_unlock(&ctx->mutex);
731 static ssize_t ucma_resolve_addr(struct ucma_file *file,
732 const char __user *inbuf,
733 int in_len, int out_len)
735 struct rdma_ucm_resolve_addr cmd;
736 struct ucma_context *ctx;
739 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
743 (cmd.src_size && (cmd.src_size != rdma_addr_size_kss(&cmd.src_addr))) ||
744 !cmd.dst_size || (cmd.dst_size != rdma_addr_size_kss(&cmd.dst_addr)))
747 ctx = ucma_get_ctx(file, cmd.id);
751 mutex_lock(&ctx->mutex);
752 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
753 (struct sockaddr *) &cmd.dst_addr, cmd.timeout_ms);
754 mutex_unlock(&ctx->mutex);
759 static ssize_t ucma_resolve_route(struct ucma_file *file,
760 const char __user *inbuf,
761 int in_len, int out_len)
763 struct rdma_ucm_resolve_route cmd;
764 struct ucma_context *ctx;
767 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
770 ctx = ucma_get_ctx_dev(file, cmd.id);
774 mutex_lock(&ctx->mutex);
775 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
776 mutex_unlock(&ctx->mutex);
781 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
782 struct rdma_route *route)
784 struct rdma_dev_addr *dev_addr;
786 resp->num_paths = route->num_paths;
787 switch (route->num_paths) {
789 dev_addr = &route->addr.dev_addr;
790 rdma_addr_get_dgid(dev_addr,
791 (union ib_gid *) &resp->ib_route[0].dgid);
792 rdma_addr_get_sgid(dev_addr,
793 (union ib_gid *) &resp->ib_route[0].sgid);
794 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
797 ib_copy_path_rec_to_user(&resp->ib_route[1],
798 &route->path_rec[1]);
801 ib_copy_path_rec_to_user(&resp->ib_route[0],
802 &route->path_rec[0]);
809 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
810 struct rdma_route *route)
813 resp->num_paths = route->num_paths;
814 switch (route->num_paths) {
816 rdma_ip2gid((struct sockaddr *)&route->addr.dst_addr,
817 (union ib_gid *)&resp->ib_route[0].dgid);
818 rdma_ip2gid((struct sockaddr *)&route->addr.src_addr,
819 (union ib_gid *)&resp->ib_route[0].sgid);
820 resp->ib_route[0].pkey = cpu_to_be16(0xffff);
823 ib_copy_path_rec_to_user(&resp->ib_route[1],
824 &route->path_rec[1]);
827 ib_copy_path_rec_to_user(&resp->ib_route[0],
828 &route->path_rec[0]);
835 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
836 struct rdma_route *route)
838 struct rdma_dev_addr *dev_addr;
840 dev_addr = &route->addr.dev_addr;
841 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
842 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
845 static ssize_t ucma_query_route(struct ucma_file *file,
846 const char __user *inbuf,
847 int in_len, int out_len)
849 struct rdma_ucm_query cmd;
850 struct rdma_ucm_query_route_resp resp;
851 struct ucma_context *ctx;
852 struct sockaddr *addr;
855 if (out_len < sizeof(resp))
858 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
861 ctx = ucma_get_ctx(file, cmd.id);
865 mutex_lock(&ctx->mutex);
866 memset(&resp, 0, sizeof resp);
867 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
868 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
869 sizeof(struct sockaddr_in) :
870 sizeof(struct sockaddr_in6));
871 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
872 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
873 sizeof(struct sockaddr_in) :
874 sizeof(struct sockaddr_in6));
875 if (!ctx->cm_id->device)
878 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
879 resp.port_num = ctx->cm_id->port_num;
881 if (rdma_cap_ib_sa(ctx->cm_id->device, ctx->cm_id->port_num))
882 ucma_copy_ib_route(&resp, &ctx->cm_id->route);
883 else if (rdma_protocol_roce(ctx->cm_id->device, ctx->cm_id->port_num))
884 ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
885 else if (rdma_protocol_iwarp(ctx->cm_id->device, ctx->cm_id->port_num))
886 ucma_copy_iw_route(&resp, &ctx->cm_id->route);
889 mutex_unlock(&ctx->mutex);
890 if (copy_to_user(u64_to_user_ptr(cmd.response),
891 &resp, sizeof(resp)))
898 static void ucma_query_device_addr(struct rdma_cm_id *cm_id,
899 struct rdma_ucm_query_addr_resp *resp)
904 resp->node_guid = (__force __u64) cm_id->device->node_guid;
905 resp->port_num = cm_id->port_num;
906 resp->pkey = (__force __u16) cpu_to_be16(
907 ib_addr_get_pkey(&cm_id->route.addr.dev_addr));
910 static ssize_t ucma_query_addr(struct ucma_context *ctx,
911 void __user *response, int out_len)
913 struct rdma_ucm_query_addr_resp resp;
914 struct sockaddr *addr;
917 if (out_len < sizeof(resp))
920 memset(&resp, 0, sizeof resp);
922 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
923 resp.src_size = rdma_addr_size(addr);
924 memcpy(&resp.src_addr, addr, resp.src_size);
926 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
927 resp.dst_size = rdma_addr_size(addr);
928 memcpy(&resp.dst_addr, addr, resp.dst_size);
930 ucma_query_device_addr(ctx->cm_id, &resp);
932 if (copy_to_user(response, &resp, sizeof(resp)))
938 static ssize_t ucma_query_path(struct ucma_context *ctx,
939 void __user *response, int out_len)
941 struct rdma_ucm_query_path_resp *resp;
944 if (out_len < sizeof(*resp))
947 resp = kzalloc(out_len, GFP_KERNEL);
951 resp->num_paths = ctx->cm_id->route.num_paths;
952 for (i = 0, out_len -= sizeof(*resp);
953 i < resp->num_paths && out_len > sizeof(struct ib_path_rec_data);
954 i++, out_len -= sizeof(struct ib_path_rec_data)) {
955 struct sa_path_rec *rec = &ctx->cm_id->route.path_rec[i];
957 resp->path_data[i].flags = IB_PATH_GMP | IB_PATH_PRIMARY |
958 IB_PATH_BIDIRECTIONAL;
959 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
960 struct sa_path_rec ib;
962 sa_convert_path_opa_to_ib(&ib, rec);
963 ib_sa_pack_path(&ib, &resp->path_data[i].path_rec);
966 ib_sa_pack_path(rec, &resp->path_data[i].path_rec);
970 if (copy_to_user(response, resp,
971 sizeof(*resp) + (i * sizeof(struct ib_path_rec_data))))
978 static ssize_t ucma_query_gid(struct ucma_context *ctx,
979 void __user *response, int out_len)
981 struct rdma_ucm_query_addr_resp resp;
982 struct sockaddr_ib *addr;
985 if (out_len < sizeof(resp))
988 memset(&resp, 0, sizeof resp);
990 ucma_query_device_addr(ctx->cm_id, &resp);
992 addr = (struct sockaddr_ib *) &resp.src_addr;
993 resp.src_size = sizeof(*addr);
994 if (ctx->cm_id->route.addr.src_addr.ss_family == AF_IB) {
995 memcpy(addr, &ctx->cm_id->route.addr.src_addr, resp.src_size);
997 addr->sib_family = AF_IB;
998 addr->sib_pkey = (__force __be16) resp.pkey;
999 rdma_read_gids(ctx->cm_id, (union ib_gid *)&addr->sib_addr,
1001 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
1002 &ctx->cm_id->route.addr.src_addr);
1005 addr = (struct sockaddr_ib *) &resp.dst_addr;
1006 resp.dst_size = sizeof(*addr);
1007 if (ctx->cm_id->route.addr.dst_addr.ss_family == AF_IB) {
1008 memcpy(addr, &ctx->cm_id->route.addr.dst_addr, resp.dst_size);
1010 addr->sib_family = AF_IB;
1011 addr->sib_pkey = (__force __be16) resp.pkey;
1012 rdma_read_gids(ctx->cm_id, NULL,
1013 (union ib_gid *)&addr->sib_addr);
1014 addr->sib_sid = rdma_get_service_id(ctx->cm_id, (struct sockaddr *)
1015 &ctx->cm_id->route.addr.dst_addr);
1018 if (copy_to_user(response, &resp, sizeof(resp)))
1024 static ssize_t ucma_query(struct ucma_file *file,
1025 const char __user *inbuf,
1026 int in_len, int out_len)
1028 struct rdma_ucm_query cmd;
1029 struct ucma_context *ctx;
1030 void __user *response;
1033 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1036 response = u64_to_user_ptr(cmd.response);
1037 ctx = ucma_get_ctx(file, cmd.id);
1039 return PTR_ERR(ctx);
1041 mutex_lock(&ctx->mutex);
1042 switch (cmd.option) {
1043 case RDMA_USER_CM_QUERY_ADDR:
1044 ret = ucma_query_addr(ctx, response, out_len);
1046 case RDMA_USER_CM_QUERY_PATH:
1047 ret = ucma_query_path(ctx, response, out_len);
1049 case RDMA_USER_CM_QUERY_GID:
1050 ret = ucma_query_gid(ctx, response, out_len);
1056 mutex_unlock(&ctx->mutex);
1062 static void ucma_copy_conn_param(struct rdma_cm_id *id,
1063 struct rdma_conn_param *dst,
1064 struct rdma_ucm_conn_param *src)
1066 dst->private_data = src->private_data;
1067 dst->private_data_len = src->private_data_len;
1068 dst->responder_resources =src->responder_resources;
1069 dst->initiator_depth = src->initiator_depth;
1070 dst->flow_control = src->flow_control;
1071 dst->retry_count = src->retry_count;
1072 dst->rnr_retry_count = src->rnr_retry_count;
1073 dst->srq = src->srq;
1074 dst->qp_num = src->qp_num;
1075 dst->qkey = (id->route.addr.src_addr.ss_family == AF_IB) ? src->qkey : 0;
1078 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
1079 int in_len, int out_len)
1081 struct rdma_ucm_connect cmd;
1082 struct rdma_conn_param conn_param;
1083 struct ucma_context *ctx;
1086 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1089 if (!cmd.conn_param.valid)
1092 ctx = ucma_get_ctx_dev(file, cmd.id);
1094 return PTR_ERR(ctx);
1096 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1097 mutex_lock(&ctx->mutex);
1098 ret = rdma_connect(ctx->cm_id, &conn_param);
1099 mutex_unlock(&ctx->mutex);
1104 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
1105 int in_len, int out_len)
1107 struct rdma_ucm_listen cmd;
1108 struct ucma_context *ctx;
1111 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1114 ctx = ucma_get_ctx(file, cmd.id);
1116 return PTR_ERR(ctx);
1118 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
1119 cmd.backlog : max_backlog;
1120 mutex_lock(&ctx->mutex);
1121 ret = rdma_listen(ctx->cm_id, ctx->backlog);
1122 mutex_unlock(&ctx->mutex);
1127 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
1128 int in_len, int out_len)
1130 struct rdma_ucm_accept cmd;
1131 struct rdma_conn_param conn_param;
1132 struct ucma_context *ctx;
1135 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1138 ctx = ucma_get_ctx_dev(file, cmd.id);
1140 return PTR_ERR(ctx);
1142 if (cmd.conn_param.valid) {
1143 ucma_copy_conn_param(ctx->cm_id, &conn_param, &cmd.conn_param);
1144 mutex_lock(&file->mut);
1145 mutex_lock(&ctx->mutex);
1146 ret = __rdma_accept(ctx->cm_id, &conn_param, NULL);
1147 mutex_unlock(&ctx->mutex);
1150 mutex_unlock(&file->mut);
1152 mutex_lock(&ctx->mutex);
1153 ret = __rdma_accept(ctx->cm_id, NULL, NULL);
1154 mutex_unlock(&ctx->mutex);
1160 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
1161 int in_len, int out_len)
1163 struct rdma_ucm_reject cmd;
1164 struct ucma_context *ctx;
1167 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1170 ctx = ucma_get_ctx_dev(file, cmd.id);
1172 return PTR_ERR(ctx);
1174 mutex_lock(&ctx->mutex);
1175 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
1176 mutex_unlock(&ctx->mutex);
1181 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
1182 int in_len, int out_len)
1184 struct rdma_ucm_disconnect cmd;
1185 struct ucma_context *ctx;
1188 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1191 ctx = ucma_get_ctx_dev(file, cmd.id);
1193 return PTR_ERR(ctx);
1195 mutex_lock(&ctx->mutex);
1196 ret = rdma_disconnect(ctx->cm_id);
1197 mutex_unlock(&ctx->mutex);
1202 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
1203 const char __user *inbuf,
1204 int in_len, int out_len)
1206 struct rdma_ucm_init_qp_attr cmd;
1207 struct ib_uverbs_qp_attr resp;
1208 struct ucma_context *ctx;
1209 struct ib_qp_attr qp_attr;
1212 if (out_len < sizeof(resp))
1215 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1218 if (cmd.qp_state > IB_QPS_ERR)
1221 ctx = ucma_get_ctx_dev(file, cmd.id);
1223 return PTR_ERR(ctx);
1225 resp.qp_attr_mask = 0;
1226 memset(&qp_attr, 0, sizeof qp_attr);
1227 qp_attr.qp_state = cmd.qp_state;
1228 mutex_lock(&ctx->mutex);
1229 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
1230 mutex_unlock(&ctx->mutex);
1234 ib_copy_qp_attr_to_user(ctx->cm_id->device, &resp, &qp_attr);
1235 if (copy_to_user(u64_to_user_ptr(cmd.response),
1236 &resp, sizeof(resp)))
1244 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
1245 void *optval, size_t optlen)
1250 case RDMA_OPTION_ID_TOS:
1251 if (optlen != sizeof(u8)) {
1255 rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
1257 case RDMA_OPTION_ID_REUSEADDR:
1258 if (optlen != sizeof(int)) {
1262 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
1264 case RDMA_OPTION_ID_AFONLY:
1265 if (optlen != sizeof(int)) {
1269 ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
1278 static int ucma_set_ib_path(struct ucma_context *ctx,
1279 struct ib_path_rec_data *path_data, size_t optlen)
1281 struct sa_path_rec sa_path;
1282 struct rdma_cm_event event;
1285 if (optlen % sizeof(*path_data))
1288 for (; optlen; optlen -= sizeof(*path_data), path_data++) {
1289 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
1290 IB_PATH_BIDIRECTIONAL))
1297 if (!ctx->cm_id->device)
1300 memset(&sa_path, 0, sizeof(sa_path));
1302 sa_path.rec_type = SA_PATH_REC_TYPE_IB;
1303 ib_sa_unpack_path(path_data->path_rec, &sa_path);
1305 if (rdma_cap_opa_ah(ctx->cm_id->device, ctx->cm_id->port_num)) {
1306 struct sa_path_rec opa;
1308 sa_convert_path_ib_to_opa(&opa, &sa_path);
1309 mutex_lock(&ctx->mutex);
1310 ret = rdma_set_ib_path(ctx->cm_id, &opa);
1311 mutex_unlock(&ctx->mutex);
1313 mutex_lock(&ctx->mutex);
1314 ret = rdma_set_ib_path(ctx->cm_id, &sa_path);
1315 mutex_unlock(&ctx->mutex);
1320 memset(&event, 0, sizeof event);
1321 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1322 return ucma_event_handler(ctx->cm_id, &event);
1325 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
1326 void *optval, size_t optlen)
1331 case RDMA_OPTION_IB_PATH:
1332 ret = ucma_set_ib_path(ctx, optval, optlen);
1341 static int ucma_set_option_level(struct ucma_context *ctx, int level,
1342 int optname, void *optval, size_t optlen)
1347 case RDMA_OPTION_ID:
1348 mutex_lock(&ctx->mutex);
1349 ret = ucma_set_option_id(ctx, optname, optval, optlen);
1350 mutex_unlock(&ctx->mutex);
1352 case RDMA_OPTION_IB:
1353 ret = ucma_set_option_ib(ctx, optname, optval, optlen);
1362 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
1363 int in_len, int out_len)
1365 struct rdma_ucm_set_option cmd;
1366 struct ucma_context *ctx;
1370 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1373 if (unlikely(cmd.optlen > KMALLOC_MAX_SIZE))
1376 ctx = ucma_get_ctx(file, cmd.id);
1378 return PTR_ERR(ctx);
1380 optval = memdup_user(u64_to_user_ptr(cmd.optval),
1382 if (IS_ERR(optval)) {
1383 ret = PTR_ERR(optval);
1387 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1396 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1397 int in_len, int out_len)
1399 struct rdma_ucm_notify cmd;
1400 struct ucma_context *ctx;
1403 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1406 ctx = ucma_get_ctx(file, cmd.id);
1408 return PTR_ERR(ctx);
1410 mutex_lock(&ctx->mutex);
1411 if (ctx->cm_id->device)
1412 ret = rdma_notify(ctx->cm_id, (enum ib_event_type)cmd.event);
1413 mutex_unlock(&ctx->mutex);
1419 static ssize_t ucma_process_join(struct ucma_file *file,
1420 struct rdma_ucm_join_mcast *cmd, int out_len)
1422 struct rdma_ucm_create_id_resp resp;
1423 struct ucma_context *ctx;
1424 struct ucma_multicast *mc;
1425 struct sockaddr *addr;
1429 if (out_len < sizeof(resp))
1432 addr = (struct sockaddr *) &cmd->addr;
1433 if (cmd->addr_size != rdma_addr_size(addr))
1436 if (cmd->join_flags == RDMA_MC_JOIN_FLAG_FULLMEMBER)
1437 join_state = BIT(FULLMEMBER_JOIN);
1438 else if (cmd->join_flags == RDMA_MC_JOIN_FLAG_SENDONLY_FULLMEMBER)
1439 join_state = BIT(SENDONLY_FULLMEMBER_JOIN);
1443 ctx = ucma_get_ctx_dev(file, cmd->id);
1445 return PTR_ERR(ctx);
1447 mutex_lock(&file->mut);
1448 mc = ucma_alloc_multicast(ctx);
1453 mc->join_state = join_state;
1455 memcpy(&mc->addr, addr, cmd->addr_size);
1456 mutex_lock(&ctx->mutex);
1457 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *)&mc->addr,
1459 mutex_unlock(&ctx->mutex);
1464 if (copy_to_user(u64_to_user_ptr(cmd->response),
1465 &resp, sizeof(resp))) {
1471 idr_replace(&multicast_idr, mc, mc->id);
1474 mutex_unlock(&file->mut);
1479 mutex_lock(&ctx->mutex);
1480 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1481 mutex_unlock(&ctx->mutex);
1482 ucma_cleanup_mc_events(mc);
1485 idr_remove(&multicast_idr, mc->id);
1487 list_del(&mc->list);
1490 mutex_unlock(&file->mut);
1495 static ssize_t ucma_join_ip_multicast(struct ucma_file *file,
1496 const char __user *inbuf,
1497 int in_len, int out_len)
1499 struct rdma_ucm_join_ip_mcast cmd;
1500 struct rdma_ucm_join_mcast join_cmd;
1502 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1505 join_cmd.response = cmd.response;
1506 join_cmd.uid = cmd.uid;
1507 join_cmd.id = cmd.id;
1508 join_cmd.addr_size = rdma_addr_size_in6(&cmd.addr);
1509 if (!join_cmd.addr_size)
1512 join_cmd.join_flags = RDMA_MC_JOIN_FLAG_FULLMEMBER;
1513 memcpy(&join_cmd.addr, &cmd.addr, join_cmd.addr_size);
1515 return ucma_process_join(file, &join_cmd, out_len);
1518 static ssize_t ucma_join_multicast(struct ucma_file *file,
1519 const char __user *inbuf,
1520 int in_len, int out_len)
1522 struct rdma_ucm_join_mcast cmd;
1524 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1527 if (!rdma_addr_size_kss(&cmd.addr))
1530 return ucma_process_join(file, &cmd, out_len);
1533 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1534 const char __user *inbuf,
1535 int in_len, int out_len)
1537 struct rdma_ucm_destroy_id cmd;
1538 struct rdma_ucm_destroy_id_resp resp;
1539 struct ucma_multicast *mc;
1542 if (out_len < sizeof(resp))
1545 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1549 mc = idr_find(&multicast_idr, cmd.id);
1551 mc = ERR_PTR(-ENOENT);
1552 else if (mc->ctx->file != file)
1553 mc = ERR_PTR(-EINVAL);
1554 else if (!atomic_inc_not_zero(&mc->ctx->ref))
1555 mc = ERR_PTR(-ENXIO);
1557 idr_remove(&multicast_idr, mc->id);
1565 mutex_lock(&mc->ctx->mutex);
1566 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1567 mutex_unlock(&mc->ctx->mutex);
1569 mutex_lock(&mc->ctx->file->mut);
1570 ucma_cleanup_mc_events(mc);
1571 list_del(&mc->list);
1572 mutex_unlock(&mc->ctx->file->mut);
1574 ucma_put_ctx(mc->ctx);
1575 resp.events_reported = mc->events_reported;
1578 if (copy_to_user(u64_to_user_ptr(cmd.response),
1579 &resp, sizeof(resp)))
1585 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1587 /* Acquire mutex's based on pointer comparison to prevent deadlock. */
1588 if (file1 < file2) {
1589 mutex_lock(&file1->mut);
1590 mutex_lock_nested(&file2->mut, SINGLE_DEPTH_NESTING);
1592 mutex_lock(&file2->mut);
1593 mutex_lock_nested(&file1->mut, SINGLE_DEPTH_NESTING);
1597 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1599 if (file1 < file2) {
1600 mutex_unlock(&file2->mut);
1601 mutex_unlock(&file1->mut);
1603 mutex_unlock(&file1->mut);
1604 mutex_unlock(&file2->mut);
1608 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1610 struct ucma_event *uevent, *tmp;
1612 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1613 if (uevent->ctx == ctx)
1614 list_move_tail(&uevent->list, &file->event_list);
1617 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1618 const char __user *inbuf,
1619 int in_len, int out_len)
1621 struct rdma_ucm_migrate_id cmd;
1622 struct rdma_ucm_migrate_resp resp;
1623 struct ucma_context *ctx;
1625 struct ucma_file *cur_file;
1628 if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1631 /* Get current fd to protect against it being closed */
1635 if (f.file->f_op != &ucma_fops) {
1640 /* Validate current fd and prevent destruction of id. */
1641 ctx = ucma_get_ctx(f.file->private_data, cmd.id);
1647 cur_file = ctx->file;
1648 if (cur_file == new_file) {
1649 mutex_lock(&cur_file->mut);
1650 resp.events_reported = ctx->events_reported;
1651 mutex_unlock(&cur_file->mut);
1656 * Migrate events between fd's, maintaining order, and avoiding new
1657 * events being added before existing events.
1659 ucma_lock_files(cur_file, new_file);
1662 list_move_tail(&ctx->list, &new_file->ctx_list);
1663 ucma_move_events(ctx, new_file);
1664 ctx->file = new_file;
1665 resp.events_reported = ctx->events_reported;
1668 ucma_unlock_files(cur_file, new_file);
1671 if (copy_to_user(u64_to_user_ptr(cmd.response),
1672 &resp, sizeof(resp)))
1681 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1682 const char __user *inbuf,
1683 int in_len, int out_len) = {
1684 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id,
1685 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id,
1686 [RDMA_USER_CM_CMD_BIND_IP] = ucma_bind_ip,
1687 [RDMA_USER_CM_CMD_RESOLVE_IP] = ucma_resolve_ip,
1688 [RDMA_USER_CM_CMD_RESOLVE_ROUTE] = ucma_resolve_route,
1689 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route,
1690 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect,
1691 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen,
1692 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept,
1693 [RDMA_USER_CM_CMD_REJECT] = ucma_reject,
1694 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect,
1695 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
1696 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event,
1697 [RDMA_USER_CM_CMD_GET_OPTION] = NULL,
1698 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option,
1699 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify,
1700 [RDMA_USER_CM_CMD_JOIN_IP_MCAST] = ucma_join_ip_multicast,
1701 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast,
1702 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id,
1703 [RDMA_USER_CM_CMD_QUERY] = ucma_query,
1704 [RDMA_USER_CM_CMD_BIND] = ucma_bind,
1705 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
1706 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast
1709 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1710 size_t len, loff_t *pos)
1712 struct ucma_file *file = filp->private_data;
1713 struct rdma_ucm_cmd_hdr hdr;
1716 if (!ib_safe_file_access(filp)) {
1717 pr_err_once("ucma_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
1718 task_tgid_vnr(current), current->comm);
1722 if (len < sizeof(hdr))
1725 if (copy_from_user(&hdr, buf, sizeof(hdr)))
1728 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1730 hdr.cmd = array_index_nospec(hdr.cmd, ARRAY_SIZE(ucma_cmd_table));
1732 if (hdr.in + sizeof(hdr) > len)
1735 if (!ucma_cmd_table[hdr.cmd])
1738 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1745 static __poll_t ucma_poll(struct file *filp, struct poll_table_struct *wait)
1747 struct ucma_file *file = filp->private_data;
1750 poll_wait(filp, &file->poll_wait, wait);
1752 if (!list_empty(&file->event_list))
1753 mask = EPOLLIN | EPOLLRDNORM;
1759 * ucma_open() does not need the BKL:
1761 * - no global state is referred to;
1762 * - there is no ioctl method to race against;
1763 * - no further module initialization is required for open to work
1764 * after the device is registered.
1766 static int ucma_open(struct inode *inode, struct file *filp)
1768 struct ucma_file *file;
1770 file = kmalloc(sizeof *file, GFP_KERNEL);
1774 file->close_wq = alloc_ordered_workqueue("ucma_close_id",
1776 if (!file->close_wq) {
1781 INIT_LIST_HEAD(&file->event_list);
1782 INIT_LIST_HEAD(&file->ctx_list);
1783 init_waitqueue_head(&file->poll_wait);
1784 mutex_init(&file->mut);
1786 filp->private_data = file;
1789 return nonseekable_open(inode, filp);
1792 static int ucma_close(struct inode *inode, struct file *filp)
1794 struct ucma_file *file = filp->private_data;
1795 struct ucma_context *ctx, *tmp;
1797 mutex_lock(&file->mut);
1798 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1799 ctx->destroying = 1;
1800 mutex_unlock(&file->mut);
1803 idr_remove(&ctx_idr, ctx->id);
1806 flush_workqueue(file->close_wq);
1807 /* At that step once ctx was marked as destroying and workqueue
1808 * was flushed we are safe from any inflights handlers that
1809 * might put other closing task.
1812 if (!ctx->closing) {
1815 wait_for_completion(&ctx->comp);
1816 /* rdma_destroy_id ensures that no event handlers are
1817 * inflight for that id before releasing it.
1819 rdma_destroy_id(ctx->cm_id);
1825 mutex_lock(&file->mut);
1827 mutex_unlock(&file->mut);
1828 destroy_workqueue(file->close_wq);
1833 static const struct file_operations ucma_fops = {
1834 .owner = THIS_MODULE,
1836 .release = ucma_close,
1837 .write = ucma_write,
1839 .llseek = no_llseek,
1842 static struct miscdevice ucma_misc = {
1843 .minor = MISC_DYNAMIC_MINOR,
1845 .nodename = "infiniband/rdma_cm",
1850 static ssize_t show_abi_version(struct device *dev,
1851 struct device_attribute *attr,
1854 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1856 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1858 static int __init ucma_init(void)
1862 ret = misc_register(&ucma_misc);
1866 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1868 pr_err("rdma_ucm: couldn't create abi_version attr\n");
1872 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1873 if (!ucma_ctl_table_hdr) {
1874 pr_err("rdma_ucm: couldn't register sysctl paths\n");
1880 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1882 misc_deregister(&ucma_misc);
1886 static void __exit ucma_cleanup(void)
1888 unregister_net_sysctl_table(ucma_ctl_table_hdr);
1889 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1890 misc_deregister(&ucma_misc);
1891 idr_destroy(&ctx_idr);
1892 idr_destroy(&multicast_idr);
1895 module_init(ucma_init);
1896 module_exit(ucma_cleanup);