RDMA/cma: Export rdma cm interface to userspace
authorSean Hefty <sean.hefty@intel.com>
Fri, 1 Dec 2006 00:53:41 +0000 (16:53 -0800)
committerRoland Dreier <rolandd@cisco.com>
Tue, 12 Dec 2006 19:50:22 +0000 (11:50 -0800)
Export the rdma cm interfaces to userspace via a misc device.

Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/core/Makefile
drivers/infiniband/core/ucma.c [new file with mode: 0644]
drivers/infiniband/core/uverbs_marshall.c
include/rdma/ib_marshall.h
include/rdma/rdma_user_cm.h [new file with mode: 0644]

index 163d991..50fb1cd 100644 (file)
@@ -1,9 +1,11 @@
 infiniband-$(CONFIG_INFINIBAND_ADDR_TRANS)     := ib_addr.o rdma_cm.o
+user_access-$(CONFIG_INFINIBAND_ADDR_TRANS)    := rdma_ucm.o
 
 obj-$(CONFIG_INFINIBAND) +=            ib_core.o ib_mad.o ib_sa.o \
                                        ib_cm.o iw_cm.o $(infiniband-y)
 obj-$(CONFIG_INFINIBAND_USER_MAD) +=   ib_umad.o
-obj-$(CONFIG_INFINIBAND_USER_ACCESS) +=        ib_uverbs.o ib_ucm.o
+obj-$(CONFIG_INFINIBAND_USER_ACCESS) +=        ib_uverbs.o ib_ucm.o \
+                                       $(user_access-y)
 
 ib_core-y :=                   packer.o ud_header.o verbs.o sysfs.o \
                                device.o fmr_pool.o cache.o
@@ -18,6 +20,8 @@ iw_cm-y :=                    iwcm.o
 
 rdma_cm-y :=                   cma.o
 
+rdma_ucm-y :=                  ucma.o
+
 ib_addr-y :=                   addr.o
 
 ib_umad-y :=                   user_mad.o
diff --git a/drivers/infiniband/core/ucma.c b/drivers/infiniband/core/ucma.c
new file mode 100644 (file)
index 0000000..81a5cdc
--- /dev/null
@@ -0,0 +1,874 @@
+/*
+ * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *     copyright notice, this list of conditions and the following
+ *     disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *     copyright notice, this list of conditions and the following
+ *     disclaimer in the documentation and/or other materials
+ *     provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/completion.h>
+#include <linux/mutex.h>
+#include <linux/poll.h>
+#include <linux/idr.h>
+#include <linux/in.h>
+#include <linux/in6.h>
+#include <linux/miscdevice.h>
+
+#include <rdma/rdma_user_cm.h>
+#include <rdma/ib_marshall.h>
+#include <rdma/rdma_cm.h>
+
+MODULE_AUTHOR("Sean Hefty");
+MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
+MODULE_LICENSE("Dual BSD/GPL");
+
+enum {
+       UCMA_MAX_BACKLOG        = 128
+};
+
+struct ucma_file {
+       struct mutex            mut;
+       struct file             *filp;
+       struct list_head        ctx_list;
+       struct list_head        event_list;
+       wait_queue_head_t       poll_wait;
+};
+
+struct ucma_context {
+       int                     id;
+       struct completion       comp;
+       atomic_t                ref;
+       int                     events_reported;
+       int                     backlog;
+
+       struct ucma_file        *file;
+       struct rdma_cm_id       *cm_id;
+       u64                     uid;
+
+       struct list_head        list;
+};
+
+struct ucma_event {
+       struct ucma_context     *ctx;
+       struct list_head        list;
+       struct rdma_cm_id       *cm_id;
+       struct rdma_ucm_event_resp resp;
+};
+
+static DEFINE_MUTEX(mut);
+static DEFINE_IDR(ctx_idr);
+
+static inline struct ucma_context *_ucma_find_context(int id,
+                                                     struct ucma_file *file)
+{
+       struct ucma_context *ctx;
+
+       ctx = idr_find(&ctx_idr, id);
+       if (!ctx)
+               ctx = ERR_PTR(-ENOENT);
+       else if (ctx->file != file)
+               ctx = ERR_PTR(-EINVAL);
+       return ctx;
+}
+
+static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
+{
+       struct ucma_context *ctx;
+
+       mutex_lock(&mut);
+       ctx = _ucma_find_context(id, file);
+       if (!IS_ERR(ctx))
+               atomic_inc(&ctx->ref);
+       mutex_unlock(&mut);
+       return ctx;
+}
+
+static void ucma_put_ctx(struct ucma_context *ctx)
+{
+       if (atomic_dec_and_test(&ctx->ref))
+               complete(&ctx->comp);
+}
+
+static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
+{
+       struct ucma_context *ctx;
+       int ret;
+
+       ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
+       if (!ctx)
+               return NULL;
+
+       atomic_set(&ctx->ref, 1);
+       init_completion(&ctx->comp);
+       ctx->file = file;
+
+       do {
+               ret = idr_pre_get(&ctx_idr, GFP_KERNEL);
+               if (!ret)
+                       goto error;
+
+               mutex_lock(&mut);
+               ret = idr_get_new(&ctx_idr, ctx, &ctx->id);
+               mutex_unlock(&mut);
+       } while (ret == -EAGAIN);
+
+       if (ret)
+               goto error;
+
+       list_add_tail(&ctx->list, &file->ctx_list);
+       return ctx;
+
+error:
+       kfree(ctx);
+       return NULL;
+}
+
+static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
+                                struct rdma_conn_param *src)
+{
+       if (src->private_data_len)
+               memcpy(dst->private_data, src->private_data,
+                      src->private_data_len);
+       dst->private_data_len = src->private_data_len;
+       dst->responder_resources =src->responder_resources;
+       dst->initiator_depth = src->initiator_depth;
+       dst->flow_control = src->flow_control;
+       dst->retry_count = src->retry_count;
+       dst->rnr_retry_count = src->rnr_retry_count;
+       dst->srq = src->srq;
+       dst->qp_num = src->qp_num;
+}
+
+static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
+                              struct rdma_ud_param *src)
+{
+       if (src->private_data_len)
+               memcpy(dst->private_data, src->private_data,
+                      src->private_data_len);
+       dst->private_data_len = src->private_data_len;
+       ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
+       dst->qp_num = src->qp_num;
+       dst->qkey = src->qkey;
+}
+
+static void ucma_set_event_context(struct ucma_context *ctx,
+                                  struct rdma_cm_event *event,
+                                  struct ucma_event *uevent)
+{
+       uevent->ctx = ctx;
+       uevent->resp.uid = ctx->uid;
+       uevent->resp.id = ctx->id;
+}
+
+static int ucma_event_handler(struct rdma_cm_id *cm_id,
+                             struct rdma_cm_event *event)
+{
+       struct ucma_event *uevent;
+       struct ucma_context *ctx = cm_id->context;
+       int ret = 0;
+
+       uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
+       if (!uevent)
+               return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
+
+       uevent->cm_id = cm_id;
+       ucma_set_event_context(ctx, event, uevent);
+       uevent->resp.event = event->event;
+       uevent->resp.status = event->status;
+       if (cm_id->ps == RDMA_PS_UDP)
+               ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
+       else
+               ucma_copy_conn_event(&uevent->resp.param.conn,
+                                    &event->param.conn);
+
+       mutex_lock(&ctx->file->mut);
+       if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
+               if (!ctx->backlog) {
+                       ret = -EDQUOT;
+                       goto out;
+               }
+               ctx->backlog--;
+       }
+       list_add_tail(&uevent->list, &ctx->file->event_list);
+       wake_up_interruptible(&ctx->file->poll_wait);
+out:
+       mutex_unlock(&ctx->file->mut);
+       return ret;
+}
+
+static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
+                             int in_len, int out_len)
+{
+       struct ucma_context *ctx;
+       struct rdma_ucm_get_event cmd;
+       struct ucma_event *uevent;
+       int ret = 0;
+       DEFINE_WAIT(wait);
+
+       if (out_len < sizeof uevent->resp)
+               return -ENOSPC;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       mutex_lock(&file->mut);
+       while (list_empty(&file->event_list)) {
+               if (file->filp->f_flags & O_NONBLOCK) {
+                       ret = -EAGAIN;
+                       break;
+               }
+
+               if (signal_pending(current)) {
+                       ret = -ERESTARTSYS;
+                       break;
+               }
+
+               prepare_to_wait(&file->poll_wait, &wait, TASK_INTERRUPTIBLE);
+               mutex_unlock(&file->mut);
+               schedule();
+               mutex_lock(&file->mut);
+               finish_wait(&file->poll_wait, &wait);
+       }
+
+       if (ret)
+               goto done;
+
+       uevent = list_entry(file->event_list.next, struct ucma_event, list);
+
+       if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
+               ctx = ucma_alloc_ctx(file);
+               if (!ctx) {
+                       ret = -ENOMEM;
+                       goto done;
+               }
+               uevent->ctx->backlog++;
+               ctx->cm_id = uevent->cm_id;
+               ctx->cm_id->context = ctx;
+               uevent->resp.id = ctx->id;
+       }
+
+       if (copy_to_user((void __user *)(unsigned long)cmd.response,
+                        &uevent->resp, sizeof uevent->resp)) {
+               ret = -EFAULT;
+               goto done;
+       }
+
+       list_del(&uevent->list);
+       uevent->ctx->events_reported++;
+       kfree(uevent);
+done:
+       mutex_unlock(&file->mut);
+       return ret;
+}
+
+static ssize_t ucma_create_id(struct ucma_file *file,
+                               const char __user *inbuf,
+                               int in_len, int out_len)
+{
+       struct rdma_ucm_create_id cmd;
+       struct rdma_ucm_create_id_resp resp;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (out_len < sizeof(resp))
+               return -ENOSPC;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       mutex_lock(&file->mut);
+       ctx = ucma_alloc_ctx(file);
+       mutex_unlock(&file->mut);
+       if (!ctx)
+               return -ENOMEM;
+
+       ctx->uid = cmd.uid;
+       ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps);
+       if (IS_ERR(ctx->cm_id)) {
+               ret = PTR_ERR(ctx->cm_id);
+               goto err1;
+       }
+
+       resp.id = ctx->id;
+       if (copy_to_user((void __user *)(unsigned long)cmd.response,
+                        &resp, sizeof(resp))) {
+               ret = -EFAULT;
+               goto err2;
+       }
+       return 0;
+
+err2:
+       rdma_destroy_id(ctx->cm_id);
+err1:
+       mutex_lock(&mut);
+       idr_remove(&ctx_idr, ctx->id);
+       mutex_unlock(&mut);
+       kfree(ctx);
+       return ret;
+}
+
+static void ucma_cleanup_events(struct ucma_context *ctx)
+{
+       struct ucma_event *uevent, *tmp;
+
+       list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
+               if (uevent->ctx != ctx)
+                       continue;
+
+               list_del(&uevent->list);
+
+               /* clear incoming connections. */
+               if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
+                       rdma_destroy_id(uevent->cm_id);
+
+               kfree(uevent);
+       }
+}
+
+static int ucma_free_ctx(struct ucma_context *ctx)
+{
+       int events_reported;
+
+       /* No new events will be generated after destroying the id. */
+       rdma_destroy_id(ctx->cm_id);
+
+       /* Cleanup events not yet reported to the user. */
+       mutex_lock(&ctx->file->mut);
+       ucma_cleanup_events(ctx);
+       list_del(&ctx->list);
+       mutex_unlock(&ctx->file->mut);
+
+       events_reported = ctx->events_reported;
+       kfree(ctx);
+       return events_reported;
+}
+
+static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
+                              int in_len, int out_len)
+{
+       struct rdma_ucm_destroy_id cmd;
+       struct rdma_ucm_destroy_id_resp resp;
+       struct ucma_context *ctx;
+       int ret = 0;
+
+       if (out_len < sizeof(resp))
+               return -ENOSPC;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       mutex_lock(&mut);
+       ctx = _ucma_find_context(cmd.id, file);
+       if (!IS_ERR(ctx))
+               idr_remove(&ctx_idr, ctx->id);
+       mutex_unlock(&mut);
+
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ucma_put_ctx(ctx);
+       wait_for_completion(&ctx->comp);
+       resp.events_reported = ucma_free_ctx(ctx);
+
+       if (copy_to_user((void __user *)(unsigned long)cmd.response,
+                        &resp, sizeof(resp)))
+               ret = -EFAULT;
+
+       return ret;
+}
+
+static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf,
+                             int in_len, int out_len)
+{
+       struct rdma_ucm_bind_addr cmd;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t ucma_resolve_addr(struct ucma_file *file,
+                                const char __user *inbuf,
+                                int in_len, int out_len)
+{
+       struct rdma_ucm_resolve_addr cmd;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
+                               (struct sockaddr *) &cmd.dst_addr,
+                               cmd.timeout_ms);
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t ucma_resolve_route(struct ucma_file *file,
+                                 const char __user *inbuf,
+                                 int in_len, int out_len)
+{
+       struct rdma_ucm_resolve_route cmd;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
+                              struct rdma_route *route)
+{
+       struct rdma_dev_addr *dev_addr;
+
+       resp->num_paths = route->num_paths;
+       switch (route->num_paths) {
+       case 0:
+               dev_addr = &route->addr.dev_addr;
+               ib_addr_get_dgid(dev_addr,
+                                (union ib_gid *) &resp->ib_route[0].dgid);
+               ib_addr_get_sgid(dev_addr,
+                                (union ib_gid *) &resp->ib_route[0].sgid);
+               resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
+               break;
+       case 2:
+               ib_copy_path_rec_to_user(&resp->ib_route[1],
+                                        &route->path_rec[1]);
+               /* fall through */
+       case 1:
+               ib_copy_path_rec_to_user(&resp->ib_route[0],
+                                        &route->path_rec[0]);
+               break;
+       default:
+               break;
+       }
+}
+
+static ssize_t ucma_query_route(struct ucma_file *file,
+                               const char __user *inbuf,
+                               int in_len, int out_len)
+{
+       struct rdma_ucm_query_route cmd;
+       struct rdma_ucm_query_route_resp resp;
+       struct ucma_context *ctx;
+       struct sockaddr *addr;
+       int ret = 0;
+
+       if (out_len < sizeof(resp))
+               return -ENOSPC;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       memset(&resp, 0, sizeof resp);
+       addr = &ctx->cm_id->route.addr.src_addr;
+       memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
+                                    sizeof(struct sockaddr_in) :
+                                    sizeof(struct sockaddr_in6));
+       addr = &ctx->cm_id->route.addr.dst_addr;
+       memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
+                                    sizeof(struct sockaddr_in) :
+                                    sizeof(struct sockaddr_in6));
+       if (!ctx->cm_id->device)
+               goto out;
+
+       resp.node_guid = ctx->cm_id->device->node_guid;
+       resp.port_num = ctx->cm_id->port_num;
+       switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
+       case RDMA_TRANSPORT_IB:
+               ucma_copy_ib_route(&resp, &ctx->cm_id->route);
+               break;
+       default:
+               break;
+       }
+
+out:
+       if (copy_to_user((void __user *)(unsigned long)cmd.response,
+                        &resp, sizeof(resp)))
+               ret = -EFAULT;
+
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static void ucma_copy_conn_param(struct rdma_conn_param *dst,
+                                struct rdma_ucm_conn_param *src)
+{
+       dst->private_data = src->private_data;
+       dst->private_data_len = src->private_data_len;
+       dst->responder_resources =src->responder_resources;
+       dst->initiator_depth = src->initiator_depth;
+       dst->flow_control = src->flow_control;
+       dst->retry_count = src->retry_count;
+       dst->rnr_retry_count = src->rnr_retry_count;
+       dst->srq = src->srq;
+       dst->qp_num = src->qp_num;
+}
+
+static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
+                           int in_len, int out_len)
+{
+       struct rdma_ucm_connect cmd;
+       struct rdma_conn_param conn_param;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       if (!cmd.conn_param.valid)
+               return -EINVAL;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ucma_copy_conn_param(&conn_param, &cmd.conn_param);
+       ret = rdma_connect(ctx->cm_id, &conn_param);
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
+                          int in_len, int out_len)
+{
+       struct rdma_ucm_listen cmd;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ctx->backlog = cmd.backlog > 0 && cmd.backlog < UCMA_MAX_BACKLOG ?
+                      cmd.backlog : UCMA_MAX_BACKLOG;
+       ret = rdma_listen(ctx->cm_id, ctx->backlog);
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
+                          int in_len, int out_len)
+{
+       struct rdma_ucm_accept cmd;
+       struct rdma_conn_param conn_param;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       if (cmd.conn_param.valid) {
+               ctx->uid = cmd.uid;
+               ucma_copy_conn_param(&conn_param, &cmd.conn_param);
+               ret = rdma_accept(ctx->cm_id, &conn_param);
+       } else
+               ret = rdma_accept(ctx->cm_id, NULL);
+
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
+                          int in_len, int out_len)
+{
+       struct rdma_ucm_reject cmd;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
+                              int in_len, int out_len)
+{
+       struct rdma_ucm_disconnect cmd;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ret = rdma_disconnect(ctx->cm_id);
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t ucma_init_qp_attr(struct ucma_file *file,
+                                const char __user *inbuf,
+                                int in_len, int out_len)
+{
+       struct rdma_ucm_init_qp_attr cmd;
+       struct ib_uverbs_qp_attr resp;
+       struct ucma_context *ctx;
+       struct ib_qp_attr qp_attr;
+       int ret;
+
+       if (out_len < sizeof(resp))
+               return -ENOSPC;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       resp.qp_attr_mask = 0;
+       memset(&qp_attr, 0, sizeof qp_attr);
+       qp_attr.qp_state = cmd.qp_state;
+       ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
+       if (ret)
+               goto out;
+
+       ib_copy_qp_attr_to_user(&resp, &qp_attr);
+       if (copy_to_user((void __user *)(unsigned long)cmd.response,
+                        &resp, sizeof(resp)))
+               ret = -EFAULT;
+
+out:
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
+                          int in_len, int out_len)
+{
+       struct rdma_ucm_notify cmd;
+       struct ucma_context *ctx;
+       int ret;
+
+       if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
+               return -EFAULT;
+
+       ctx = ucma_get_ctx(file, cmd.id);
+       if (IS_ERR(ctx))
+               return PTR_ERR(ctx);
+
+       ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
+       ucma_put_ctx(ctx);
+       return ret;
+}
+
+static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
+                                  const char __user *inbuf,
+                                  int in_len, int out_len) = {
+       [RDMA_USER_CM_CMD_CREATE_ID]    = ucma_create_id,
+       [RDMA_USER_CM_CMD_DESTROY_ID]   = ucma_destroy_id,
+       [RDMA_USER_CM_CMD_BIND_ADDR]    = ucma_bind_addr,
+       [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr,
+       [RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route,
+       [RDMA_USER_CM_CMD_QUERY_ROUTE]  = ucma_query_route,
+       [RDMA_USER_CM_CMD_CONNECT]      = ucma_connect,
+       [RDMA_USER_CM_CMD_LISTEN]       = ucma_listen,
+       [RDMA_USER_CM_CMD_ACCEPT]       = ucma_accept,
+       [RDMA_USER_CM_CMD_REJECT]       = ucma_reject,
+       [RDMA_USER_CM_CMD_DISCONNECT]   = ucma_disconnect,
+       [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr,
+       [RDMA_USER_CM_CMD_GET_EVENT]    = ucma_get_event,
+       [RDMA_USER_CM_CMD_GET_OPTION]   = NULL,
+       [RDMA_USER_CM_CMD_SET_OPTION]   = NULL,
+       [RDMA_USER_CM_CMD_NOTIFY]       = ucma_notify,
+};
+
+static ssize_t ucma_write(struct file *filp, const char __user *buf,
+                         size_t len, loff_t *pos)
+{
+       struct ucma_file *file = filp->private_data;
+       struct rdma_ucm_cmd_hdr hdr;
+       ssize_t ret;
+
+       if (len < sizeof(hdr))
+               return -EINVAL;
+
+       if (copy_from_user(&hdr, buf, sizeof(hdr)))
+               return -EFAULT;
+
+       if (hdr.cmd < 0 || hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
+               return -EINVAL;
+
+       if (hdr.in + sizeof(hdr) > len)
+               return -EINVAL;
+
+       if (!ucma_cmd_table[hdr.cmd])
+               return -ENOSYS;
+
+       ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
+       if (!ret)
+               ret = len;
+
+       return ret;
+}
+
+static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
+{
+       struct ucma_file *file = filp->private_data;
+       unsigned int mask = 0;
+
+       poll_wait(filp, &file->poll_wait, wait);
+
+       if (!list_empty(&file->event_list))
+               mask = POLLIN | POLLRDNORM;
+
+       return mask;
+}
+
+static int ucma_open(struct inode *inode, struct file *filp)
+{
+       struct ucma_file *file;
+
+       file = kmalloc(sizeof *file, GFP_KERNEL);
+       if (!file)
+               return -ENOMEM;
+
+       INIT_LIST_HEAD(&file->event_list);
+       INIT_LIST_HEAD(&file->ctx_list);
+       init_waitqueue_head(&file->poll_wait);
+       mutex_init(&file->mut);
+
+       filp->private_data = file;
+       file->filp = filp;
+       return 0;
+}
+
+static int ucma_close(struct inode *inode, struct file *filp)
+{
+       struct ucma_file *file = filp->private_data;
+       struct ucma_context *ctx, *tmp;
+
+       mutex_lock(&file->mut);
+       list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
+               mutex_unlock(&file->mut);
+
+               mutex_lock(&mut);
+               idr_remove(&ctx_idr, ctx->id);
+               mutex_unlock(&mut);
+
+               ucma_free_ctx(ctx);
+               mutex_lock(&file->mut);
+       }
+       mutex_unlock(&file->mut);
+       kfree(file);
+       return 0;
+}
+
+static struct file_operations ucma_fops = {
+       .owner   = THIS_MODULE,
+       .open    = ucma_open,
+       .release = ucma_close,
+       .write   = ucma_write,
+       .poll    = ucma_poll,
+};
+
+static struct miscdevice ucma_misc = {
+       .minor  = MISC_DYNAMIC_MINOR,
+       .name   = "rdma_cm",
+       .fops   = &ucma_fops,
+};
+
+static ssize_t show_abi_version(struct device *dev,
+                               struct device_attribute *attr,
+                               char *buf)
+{
+       return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
+}
+static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
+
+static int __init ucma_init(void)
+{
+       int ret;
+
+       ret = misc_register(&ucma_misc);
+       if (ret)
+               return ret;
+
+       ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
+       if (ret) {
+               printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
+               goto err;
+       }
+       return 0;
+err:
+       misc_deregister(&ucma_misc);
+       return ret;
+}
+
+static void __exit ucma_cleanup(void)
+{
+       device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
+       misc_deregister(&ucma_misc);
+       idr_destroy(&ctx_idr);
+}
+
+module_init(ucma_init);
+module_exit(ucma_cleanup);
index ce46b13..5440da0 100644 (file)
@@ -32,8 +32,8 @@
 
 #include <rdma/ib_marshall.h>
 
-static void ib_copy_ah_attr_to_user(struct ib_uverbs_ah_attr *dst,
-                                   struct ib_ah_attr *src)
+void ib_copy_ah_attr_to_user(struct ib_uverbs_ah_attr *dst,
+                            struct ib_ah_attr *src)
 {
        memcpy(dst->grh.dgid, src->grh.dgid.raw, sizeof src->grh.dgid);
        dst->grh.flow_label        = src->grh.flow_label;
@@ -47,6 +47,7 @@ static void ib_copy_ah_attr_to_user(struct ib_uverbs_ah_attr *dst,
        dst->is_global             = src->ah_flags & IB_AH_GRH ? 1 : 0;
        dst->port_num              = src->port_num;
 }
+EXPORT_SYMBOL(ib_copy_ah_attr_to_user);
 
 void ib_copy_qp_attr_to_user(struct ib_uverbs_qp_attr *dst,
                             struct ib_qp_attr *src)
index 66bf4d7..db03720 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2005 Intel Corporation.  All rights reserved.
+ * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
  *
  * This software is available to you under a choice of one of two
  * licenses.  You may choose to be licensed under the terms of the GNU
@@ -41,6 +41,9 @@
 void ib_copy_qp_attr_to_user(struct ib_uverbs_qp_attr *dst,
                             struct ib_qp_attr *src);
 
+void ib_copy_ah_attr_to_user(struct ib_uverbs_ah_attr *dst,
+                            struct ib_ah_attr *src);
+
 void ib_copy_path_rec_to_user(struct ib_user_path_rec *dst,
                              struct ib_sa_path_rec *src);
 
diff --git a/include/rdma/rdma_user_cm.h b/include/rdma/rdma_user_cm.h
new file mode 100644 (file)
index 0000000..9572ab8
--- /dev/null
@@ -0,0 +1,206 @@
+/*
+ * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef RDMA_USER_CM_H
+#define RDMA_USER_CM_H
+
+#include <linux/types.h>
+#include <linux/in6.h>
+#include <rdma/ib_user_verbs.h>
+#include <rdma/ib_user_sa.h>
+
+#define RDMA_USER_CM_ABI_VERSION       3
+
+#define RDMA_MAX_PRIVATE_DATA          256
+
+enum {
+       RDMA_USER_CM_CMD_CREATE_ID,
+       RDMA_USER_CM_CMD_DESTROY_ID,
+       RDMA_USER_CM_CMD_BIND_ADDR,
+       RDMA_USER_CM_CMD_RESOLVE_ADDR,
+       RDMA_USER_CM_CMD_RESOLVE_ROUTE,
+       RDMA_USER_CM_CMD_QUERY_ROUTE,
+       RDMA_USER_CM_CMD_CONNECT,
+       RDMA_USER_CM_CMD_LISTEN,
+       RDMA_USER_CM_CMD_ACCEPT,
+       RDMA_USER_CM_CMD_REJECT,
+       RDMA_USER_CM_CMD_DISCONNECT,
+       RDMA_USER_CM_CMD_INIT_QP_ATTR,
+       RDMA_USER_CM_CMD_GET_EVENT,
+       RDMA_USER_CM_CMD_GET_OPTION,
+       RDMA_USER_CM_CMD_SET_OPTION,
+       RDMA_USER_CM_CMD_NOTIFY
+};
+
+/*
+ * command ABI structures.
+ */
+struct rdma_ucm_cmd_hdr {
+       __u32 cmd;
+       __u16 in;
+       __u16 out;
+};
+
+struct rdma_ucm_create_id {
+       __u64 uid;
+       __u64 response;
+       __u16 ps;
+       __u8  reserved[6];
+};
+
+struct rdma_ucm_create_id_resp {
+       __u32 id;
+};
+
+struct rdma_ucm_destroy_id {
+       __u64 response;
+       __u32 id;
+       __u32 reserved;
+};
+
+struct rdma_ucm_destroy_id_resp {
+       __u32 events_reported;
+};
+
+struct rdma_ucm_bind_addr {
+       __u64 response;
+       struct sockaddr_in6 addr;
+       __u32 id;
+};
+
+struct rdma_ucm_resolve_addr {
+       struct sockaddr_in6 src_addr;
+       struct sockaddr_in6 dst_addr;
+       __u32 id;
+       __u32 timeout_ms;
+};
+
+struct rdma_ucm_resolve_route {
+       __u32 id;
+       __u32 timeout_ms;
+};
+
+struct rdma_ucm_query_route {
+       __u64 response;
+       __u32 id;
+       __u32 reserved;
+};
+
+struct rdma_ucm_query_route_resp {
+       __u64 node_guid;
+       struct ib_user_path_rec ib_route[2];
+       struct sockaddr_in6 src_addr;
+       struct sockaddr_in6 dst_addr;
+       __u32 num_paths;
+       __u8 port_num;
+       __u8 reserved[3];
+};
+
+struct rdma_ucm_conn_param {
+       __u32 qp_num;
+       __u32 reserved;
+       __u8  private_data[RDMA_MAX_PRIVATE_DATA];
+       __u8  private_data_len;
+       __u8  srq;
+       __u8  responder_resources;
+       __u8  initiator_depth;
+       __u8  flow_control;
+       __u8  retry_count;
+       __u8  rnr_retry_count;
+       __u8  valid;
+};
+
+struct rdma_ucm_ud_param {
+       __u32 qp_num;
+       __u32 qkey;
+       struct ib_uverbs_ah_attr ah_attr;
+       __u8  private_data[RDMA_MAX_PRIVATE_DATA];
+       __u8  private_data_len;
+       __u8  reserved[7];
+};
+
+struct rdma_ucm_connect {
+       struct rdma_ucm_conn_param conn_param;
+       __u32 id;
+       __u32 reserved;
+};
+
+struct rdma_ucm_listen {
+       __u32 id;
+       __u32 backlog;
+};
+
+struct rdma_ucm_accept {
+       __u64 uid;
+       struct rdma_ucm_conn_param conn_param;
+       __u32 id;
+       __u32 reserved;
+};
+
+struct rdma_ucm_reject {
+       __u32 id;
+       __u8  private_data_len;
+       __u8  reserved[3];
+       __u8  private_data[RDMA_MAX_PRIVATE_DATA];
+};
+
+struct rdma_ucm_disconnect {
+       __u32 id;
+};
+
+struct rdma_ucm_init_qp_attr {
+       __u64 response;
+       __u32 id;
+       __u32 qp_state;
+};
+
+struct rdma_ucm_notify {
+       __u32 id;
+       __u32 event;
+};
+
+struct rdma_ucm_get_event {
+       __u64 response;
+};
+
+struct rdma_ucm_event_resp {
+       __u64 uid;
+       __u32 id;
+       __u32 event;
+       __u32 status;
+       union {
+               struct rdma_ucm_conn_param conn;
+               struct rdma_ucm_ud_param   ud;
+       } param;
+};
+
+#endif /* RDMA_USER_CM_H */