IB/cm: Add basic performance counters
authorSean Hefty <sean.hefty@intel.com>
Tue, 17 Jul 2007 04:49:35 +0000 (21:49 -0700)
committerRoland Dreier <rolandd@cisco.com>
Fri, 25 Jan 2008 22:15:30 +0000 (14:15 -0800)
Add performance/debug counters to track sent/received messages, retries,
and duplicates.  Counters are tracked per CM message type, per port.

The counters are always enabled, so intrusive state tracking is not done.

Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/core/cm.c
drivers/infiniband/core/ucm.c

index 2e39236..c015014 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2004-2006 Intel Corporation.  All rights reserved.
+ * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
 
 #include <linux/completion.h>
 #include <linux/dma-mapping.h>
+#include <linux/device.h>
 #include <linux/err.h>
 #include <linux/idr.h>
 #include <linux/interrupt.h>
 #include <linux/random.h>
 #include <linux/rbtree.h>
 #include <linux/spinlock.h>
+#include <linux/sysfs.h>
 #include <linux/workqueue.h>
 
 #include <rdma/ib_cache.h>
@@ -78,17 +80,94 @@ static struct ib_cm {
        struct workqueue_struct *wq;
 } cm;
 
+/* Counter indexes ordered by attribute ID */
+enum {
+       CM_REQ_COUNTER,
+       CM_MRA_COUNTER,
+       CM_REJ_COUNTER,
+       CM_REP_COUNTER,
+       CM_RTU_COUNTER,
+       CM_DREQ_COUNTER,
+       CM_DREP_COUNTER,
+       CM_SIDR_REQ_COUNTER,
+       CM_SIDR_REP_COUNTER,
+       CM_LAP_COUNTER,
+       CM_APR_COUNTER,
+       CM_ATTR_COUNT,
+       CM_ATTR_ID_OFFSET = 0x0010,
+};
+
+enum {
+       CM_XMIT,
+       CM_XMIT_RETRIES,
+       CM_RECV,
+       CM_RECV_DUPLICATES,
+       CM_COUNTER_GROUPS
+};
+
+static char const counter_group_names[CM_COUNTER_GROUPS]
+                                    [sizeof("cm_rx_duplicates")] = {
+       "cm_tx_msgs", "cm_tx_retries",
+       "cm_rx_msgs", "cm_rx_duplicates"
+};
+
+struct cm_counter_group {
+       struct kobject obj;
+       atomic_long_t counter[CM_ATTR_COUNT];
+};
+
+struct cm_counter_attribute {
+       struct attribute attr;
+       int index;
+};
+
+#define CM_COUNTER_ATTR(_name, _index) \
+struct cm_counter_attribute cm_##_name##_counter_attr = { \
+       .attr = { .name = __stringify(_name), .mode = 0444, .owner = THIS_MODULE }, \
+       .index = _index \
+}
+
+static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
+static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
+static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
+static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
+static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
+static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
+static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
+static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
+static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
+static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
+static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
+
+static struct attribute *cm_counter_default_attrs[] = {
+       &cm_req_counter_attr.attr,
+       &cm_mra_counter_attr.attr,
+       &cm_rej_counter_attr.attr,
+       &cm_rep_counter_attr.attr,
+       &cm_rtu_counter_attr.attr,
+       &cm_dreq_counter_attr.attr,
+       &cm_drep_counter_attr.attr,
+       &cm_sidr_req_counter_attr.attr,
+       &cm_sidr_rep_counter_attr.attr,
+       &cm_lap_counter_attr.attr,
+       &cm_apr_counter_attr.attr,
+       NULL
+};
+
 struct cm_port {
        struct cm_device *cm_dev;
        struct ib_mad_agent *mad_agent;
+       struct kobject port_obj;
        u8 port_num;
+       struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 };
 
 struct cm_device {
        struct list_head list;
        struct ib_device *device;
+       struct kobject dev_obj;
        u8 ack_delay;
-       struct cm_port port[0];
+       struct cm_port *port[0];
 };
 
 struct cm_av {
@@ -278,7 +357,7 @@ static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
        list_for_each_entry(cm_dev, &cm.device_list, list) {
                if (!ib_find_cached_gid(cm_dev->device, &path->sgid,
                                        &p, NULL)) {
-                       port = &cm_dev->port[p-1];
+                       port = cm_dev->port[p-1];
                        break;
                }
        }
@@ -1270,6 +1349,9 @@ static void cm_dup_req_handler(struct cm_work *work,
        struct ib_mad_send_buf *msg = NULL;
        int ret;
 
+       atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                       counter[CM_REQ_COUNTER]);
+
        /* Quick state check to discard duplicate REQs. */
        if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
                return;
@@ -1616,6 +1698,8 @@ static void cm_dup_rep_handler(struct cm_work *work)
        if (!cm_id_priv)
                return;
 
+       atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                       counter[CM_REP_COUNTER]);
        ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
        if (ret)
                goto deref;
@@ -1781,6 +1865,8 @@ static int cm_rtu_handler(struct cm_work *work)
        if (cm_id_priv->id.state != IB_CM_REP_SENT &&
            cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
                spin_unlock_irq(&cm_id_priv->lock);
+               atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                               counter[CM_RTU_COUNTER]);
                goto out;
        }
        cm_id_priv->id.state = IB_CM_ESTABLISHED;
@@ -1958,6 +2044,8 @@ static int cm_dreq_handler(struct cm_work *work)
        cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
                                   dreq_msg->local_comm_id);
        if (!cm_id_priv) {
+               atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                               counter[CM_DREQ_COUNTER]);
                cm_issue_drep(work->port, work->mad_recv_wc);
                return -EINVAL;
        }
@@ -1977,6 +2065,8 @@ static int cm_dreq_handler(struct cm_work *work)
        case IB_CM_MRA_REP_RCVD:
                break;
        case IB_CM_TIMEWAIT:
+               atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                               counter[CM_DREQ_COUNTER]);
                if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
                        goto unlock;
 
@@ -1988,6 +2078,10 @@ static int cm_dreq_handler(struct cm_work *work)
                if (ib_post_send_mad(msg, NULL))
                        cm_free_msg(msg);
                goto deref;
+       case IB_CM_DREQ_RCVD:
+               atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                               counter[CM_DREQ_COUNTER]);
+               goto unlock;
        default:
                goto unlock;
        }
@@ -2339,10 +2433,20 @@ static int cm_mra_handler(struct cm_work *work)
                if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
                    cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
                    ib_modify_mad(cm_id_priv->av.port->mad_agent,
-                                 cm_id_priv->msg, timeout))
+                                 cm_id_priv->msg, timeout)) {
+                       if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
+                               atomic_long_inc(&work->port->
+                                               counter_group[CM_RECV_DUPLICATES].
+                                               counter[CM_MRA_COUNTER]);
                        goto out;
+               }
                cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
                break;
+       case IB_CM_MRA_REQ_RCVD:
+       case IB_CM_MRA_REP_RCVD:
+               atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                               counter[CM_MRA_COUNTER]);
+               /* fall through */
        default:
                goto out;
        }
@@ -2502,6 +2606,8 @@ static int cm_lap_handler(struct cm_work *work)
        case IB_CM_LAP_IDLE:
                break;
        case IB_CM_MRA_LAP_SENT:
+               atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                               counter[CM_LAP_COUNTER]);
                if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
                        goto unlock;
 
@@ -2515,6 +2621,10 @@ static int cm_lap_handler(struct cm_work *work)
                if (ib_post_send_mad(msg, NULL))
                        cm_free_msg(msg);
                goto deref;
+       case IB_CM_LAP_RCVD:
+               atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                               counter[CM_LAP_COUNTER]);
+               goto unlock;
        default:
                goto unlock;
        }
@@ -2796,6 +2906,8 @@ static int cm_sidr_req_handler(struct cm_work *work)
        cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
        if (cur_cm_id_priv) {
                spin_unlock_irq(&cm.lock);
+               atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
+                               counter[CM_SIDR_REQ_COUNTER]);
                goto out; /* Duplicate message. */
        }
        cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
@@ -2990,6 +3102,27 @@ static void cm_send_handler(struct ib_mad_agent *mad_agent,
                            struct ib_mad_send_wc *mad_send_wc)
 {
        struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
+       struct cm_port *port;
+       u16 attr_index;
+
+       port = mad_agent->context;
+       attr_index = be16_to_cpu(((struct ib_mad_hdr *)
+                                 msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
+
+       /*
+        * If the send was in response to a received message (context[0] is not
+        * set to a cm_id), and is not a REJ, then it is a send that was
+        * manually retried.
+        */
+       if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
+               msg->retries = 1;
+
+       atomic_long_add(1 + msg->retries,
+                       &port->counter_group[CM_XMIT].counter[attr_index]);
+       if (msg->retries)
+               atomic_long_add(msg->retries,
+                               &port->counter_group[CM_XMIT_RETRIES].
+                               counter[attr_index]);
 
        switch (mad_send_wc->status) {
        case IB_WC_SUCCESS:
@@ -3148,8 +3281,10 @@ EXPORT_SYMBOL(ib_cm_notify);
 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
                            struct ib_mad_recv_wc *mad_recv_wc)
 {
+       struct cm_port *port = mad_agent->context;
        struct cm_work *work;
        enum ib_cm_event_type event;
+       u16 attr_id;
        int paths = 0;
 
        switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
@@ -3194,6 +3329,10 @@ static void cm_recv_handler(struct ib_mad_agent *mad_agent,
                return;
        }
 
+       attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
+       atomic_long_inc(&port->counter_group[CM_RECV].
+                       counter[attr_id - CM_ATTR_ID_OFFSET]);
+
        work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
                       GFP_KERNEL);
        if (!work) {
@@ -3204,7 +3343,7 @@ static void cm_recv_handler(struct ib_mad_agent *mad_agent,
        INIT_DELAYED_WORK(&work->work, cm_work_handler);
        work->cm_event.event = event;
        work->mad_recv_wc = mad_recv_wc;
-       work->port = (struct cm_port *)mad_agent->context;
+       work->port = port;
        queue_delayed_work(cm.wq, &work->work, 0);
 }
 
@@ -3379,6 +3518,108 @@ static void cm_get_ack_delay(struct cm_device *cm_dev)
                cm_dev->ack_delay = attr.local_ca_ack_delay;
 }
 
+static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
+                              char *buf)
+{
+       struct cm_counter_group *group;
+       struct cm_counter_attribute *cm_attr;
+
+       group = container_of(obj, struct cm_counter_group, obj);
+       cm_attr = container_of(attr, struct cm_counter_attribute, attr);
+
+       return sprintf(buf, "%ld\n",
+                      atomic_long_read(&group->counter[cm_attr->index]));
+}
+
+static struct sysfs_ops cm_counter_ops = {
+       .show = cm_show_counter
+};
+
+static struct kobj_type cm_counter_obj_type = {
+       .sysfs_ops = &cm_counter_ops,
+       .default_attrs = cm_counter_default_attrs
+};
+
+static void cm_release_port_obj(struct kobject *obj)
+{
+       struct cm_port *cm_port;
+
+       printk(KERN_ERR "free cm port\n");
+
+       cm_port = container_of(obj, struct cm_port, port_obj);
+       kfree(cm_port);
+}
+
+static struct kobj_type cm_port_obj_type = {
+       .release = cm_release_port_obj
+};
+
+static void cm_release_dev_obj(struct kobject *obj)
+{
+       struct cm_device *cm_dev;
+
+       printk(KERN_ERR "free cm dev\n");
+
+       cm_dev = container_of(obj, struct cm_device, dev_obj);
+       kfree(cm_dev);
+}
+
+static struct kobj_type cm_dev_obj_type = {
+       .release = cm_release_dev_obj
+};
+
+struct class cm_class = {
+       .name    = "infiniband_cm",
+};
+EXPORT_SYMBOL(cm_class);
+
+static void cm_remove_fs_obj(struct kobject *obj)
+{
+       kobject_put(obj->parent);
+       kobject_put(obj);
+}
+
+static int cm_create_port_fs(struct cm_port *port)
+{
+       int i, ret;
+
+       ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
+                                  kobject_get(&port->cm_dev->dev_obj),
+                                  "%d", port->port_num);
+       if (ret) {
+               kfree(port);
+               return ret;
+       }
+
+       for (i = 0; i < CM_COUNTER_GROUPS; i++) {
+               ret = kobject_init_and_add(&port->counter_group[i].obj,
+                                          &cm_counter_obj_type,
+                                          kobject_get(&port->port_obj),
+                                          "%s", counter_group_names[i]);
+               if (ret)
+                       goto error;
+       }
+
+       return 0;
+
+error:
+       while (i--)
+               cm_remove_fs_obj(&port->counter_group[i].obj);
+       cm_remove_fs_obj(&port->port_obj);
+       return ret;
+
+}
+
+static void cm_remove_port_fs(struct cm_port *port)
+{
+       int i;
+
+       for (i = 0; i < CM_COUNTER_GROUPS; i++)
+               cm_remove_fs_obj(&port->counter_group[i].obj);
+
+       cm_remove_fs_obj(&port->port_obj);
+}
+
 static void cm_add_one(struct ib_device *device)
 {
        struct cm_device *cm_dev;
@@ -3397,7 +3638,7 @@ static void cm_add_one(struct ib_device *device)
        if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
                return;
 
-       cm_dev = kmalloc(sizeof(*cm_dev) + sizeof(*port) *
+       cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
                         device->phys_port_cnt, GFP_KERNEL);
        if (!cm_dev)
                return;
@@ -3405,11 +3646,27 @@ static void cm_add_one(struct ib_device *device)
        cm_dev->device = device;
        cm_get_ack_delay(cm_dev);
 
+       ret = kobject_init_and_add(&cm_dev->dev_obj, &cm_dev_obj_type,
+                                  &cm_class.subsys.kobj, "%s", device->name);
+       if (ret) {
+               kfree(cm_dev);
+               return;
+       }
+
        set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
        for (i = 1; i <= device->phys_port_cnt; i++) {
-               port = &cm_dev->port[i-1];
+               port = kzalloc(sizeof *port, GFP_KERNEL);
+               if (!port)
+                       goto error1;
+
+               cm_dev->port[i-1] = port;
                port->cm_dev = cm_dev;
                port->port_num = i;
+
+               ret = cm_create_port_fs(port);
+               if (ret)
+                       goto error1;
+
                port->mad_agent = ib_register_mad_agent(device, i,
                                                        IB_QPT_GSI,
                                                        &reg_req,
@@ -3418,11 +3675,11 @@ static void cm_add_one(struct ib_device *device)
                                                        cm_recv_handler,
                                                        port);
                if (IS_ERR(port->mad_agent))
-                       goto error1;
+                       goto error2;
 
                ret = ib_modify_port(device, i, 0, &port_modify);
                if (ret)
-                       goto error2;
+                       goto error3;
        }
        ib_set_client_data(device, &cm_client, cm_dev);
 
@@ -3431,17 +3688,20 @@ static void cm_add_one(struct ib_device *device)
        write_unlock_irqrestore(&cm.device_lock, flags);
        return;
 
-error2:
+error3:
        ib_unregister_mad_agent(port->mad_agent);
+error2:
+       cm_remove_port_fs(port);
 error1:
        port_modify.set_port_cap_mask = 0;
        port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
        while (--i) {
-               port = &cm_dev->port[i-1];
+               port = cm_dev->port[i-1];
                ib_modify_port(device, port->port_num, 0, &port_modify);
                ib_unregister_mad_agent(port->mad_agent);
+               cm_remove_port_fs(port);
        }
-       kfree(cm_dev);
+       cm_remove_fs_obj(&cm_dev->dev_obj);
 }
 
 static void cm_remove_one(struct ib_device *device)
@@ -3463,11 +3723,12 @@ static void cm_remove_one(struct ib_device *device)
        write_unlock_irqrestore(&cm.device_lock, flags);
 
        for (i = 1; i <= device->phys_port_cnt; i++) {
-               port = &cm_dev->port[i-1];
+               port = cm_dev->port[i-1];
                ib_modify_port(device, port->port_num, 0, &port_modify);
                ib_unregister_mad_agent(port->mad_agent);
+               cm_remove_port_fs(port);
        }
-       kfree(cm_dev);
+       cm_remove_fs_obj(&cm_dev->dev_obj);
 }
 
 static int __init ib_cm_init(void)
@@ -3488,17 +3749,25 @@ static int __init ib_cm_init(void)
        idr_pre_get(&cm.local_id_table, GFP_KERNEL);
        INIT_LIST_HEAD(&cm.timewait_list);
 
-       cm.wq = create_workqueue("ib_cm");
-       if (!cm.wq)
+       ret = class_register(&cm_class);
+       if (ret)
                return -ENOMEM;
 
+       cm.wq = create_workqueue("ib_cm");
+       if (!cm.wq) {
+               ret = -ENOMEM;
+               goto error1;
+       }
+
        ret = ib_register_client(&cm_client);
        if (ret)
-               goto error;
+               goto error2;
 
        return 0;
-error:
+error2:
        destroy_workqueue(cm.wq);
+error1:
+       class_unregister(&cm_class);
        return ret;
 }
 
@@ -3519,6 +3788,7 @@ static void __exit ib_cm_cleanup(void)
        }
 
        ib_unregister_client(&cm_client);
+       class_unregister(&cm_class);
        idr_destroy(&cm.local_id_table);
 }
 
index 424983f..4291ab4 100644 (file)
@@ -106,6 +106,9 @@ enum {
        IB_UCM_MAX_DEVICES = 32
 };
 
+/* ib_cm and ib_user_cm modules share /sys/class/infiniband_cm */
+extern struct class cm_class;
+
 #define IB_UCM_BASE_DEV MKDEV(IB_UCM_MAJOR, IB_UCM_BASE_MINOR)
 
 static void ib_ucm_add_one(struct ib_device *device);
@@ -1199,7 +1202,7 @@ static int ib_ucm_close(struct inode *inode, struct file *filp)
        return 0;
 }
 
-static void ib_ucm_release_class_dev(struct class_device *class_dev)
+static void ucm_release_class_dev(struct class_device *class_dev)
 {
        struct ib_ucm_device *dev;
 
@@ -1217,11 +1220,6 @@ static const struct file_operations ucm_fops = {
        .poll    = ib_ucm_poll,
 };
 
-static struct class ucm_class = {
-       .name    = "infiniband_cm",
-       .release = ib_ucm_release_class_dev
-};
-
 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
 {
        struct ib_ucm_device *dev;
@@ -1257,9 +1255,10 @@ static void ib_ucm_add_one(struct ib_device *device)
        if (cdev_add(&ucm_dev->dev, IB_UCM_BASE_DEV + ucm_dev->devnum, 1))
                goto err;
 
-       ucm_dev->class_dev.class = &ucm_class;
+       ucm_dev->class_dev.class = &cm_class;
        ucm_dev->class_dev.dev = device->dma_device;
        ucm_dev->class_dev.devt = ucm_dev->dev.dev;
+       ucm_dev->class_dev.release = ucm_release_class_dev;
        snprintf(ucm_dev->class_dev.class_id, BUS_ID_SIZE, "ucm%d",
                 ucm_dev->devnum);
        if (class_device_register(&ucm_dev->class_dev))
@@ -1306,40 +1305,34 @@ static int __init ib_ucm_init(void)
                                     "infiniband_cm");
        if (ret) {
                printk(KERN_ERR "ucm: couldn't register device number\n");
-               goto err;
+               goto error1;
        }
 
-       ret = class_register(&ucm_class);
-       if (ret) {
-               printk(KERN_ERR "ucm: couldn't create class infiniband_cm\n");
-               goto err_chrdev;
-       }
-
-       ret = class_create_file(&ucm_class, &class_attr_abi_version);
+       ret = class_create_file(&cm_class, &class_attr_abi_version);
        if (ret) {
                printk(KERN_ERR "ucm: couldn't create abi_version attribute\n");
-               goto err_class;
+               goto error2;
        }
 
        ret = ib_register_client(&ucm_client);
        if (ret) {
                printk(KERN_ERR "ucm: couldn't register client\n");
-               goto err_class;
+               goto error3;
        }
        return 0;
 
-err_class:
-       class_unregister(&ucm_class);
-err_chrdev:
+error3:
+       class_remove_file(&cm_class, &class_attr_abi_version);
+error2:
        unregister_chrdev_region(IB_UCM_BASE_DEV, IB_UCM_MAX_DEVICES);
-err:
+error1:
        return ret;
 }
 
 static void __exit ib_ucm_cleanup(void)
 {
        ib_unregister_client(&ucm_client);
-       class_unregister(&ucm_class);
+       class_remove_file(&cm_class, &class_attr_abi_version);
        unregister_chrdev_region(IB_UCM_BASE_DEV, IB_UCM_MAX_DEVICES);
        idr_destroy(&ctx_id_table);
 }