IB: IP address based RDMA connection manager
Sean Hefty [Sun, 18 Jun 2006 03:37:29 +0000 (20:37 -0700)]
Kernel connection management agent over InfiniBand that connects based
on IP addresses.  The agent defines a generic RDMA connection
abstraction to support clients wanting to connect over different RDMA
devices.

The agent also handles RDMA device hotplug events on behalf of clients.

Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/core/Makefile
drivers/infiniband/core/cma.c [new file with mode: 0644]
include/rdma/rdma_cm.h [new file with mode: 0644]
include/rdma/rdma_cm_ib.h [new file with mode: 0644]

index 0509591..68e73ec 100644 (file)
@@ -1,7 +1,7 @@
-addr_trans-$(CONFIG_INFINIBAND_ADDR_TRANS) := ib_addr.o
+infiniband-$(CONFIG_INFINIBAND_ADDR_TRANS)     := ib_addr.o rdma_cm.o
 
 obj-$(CONFIG_INFINIBAND) +=            ib_core.o ib_mad.o ib_sa.o \
-                                       ib_cm.o $(addr_trans-y)
+                                       ib_cm.o $(infiniband-y)
 obj-$(CONFIG_INFINIBAND_USER_MAD) +=   ib_umad.o
 obj-$(CONFIG_INFINIBAND_USER_ACCESS) +=        ib_uverbs.o ib_ucm.o
 
@@ -14,6 +14,8 @@ ib_sa-y :=                    sa_query.o
 
 ib_cm-y :=                     cm.o
 
+rdma_cm-y :=                   cma.o
+
 ib_addr-y :=                   addr.o
 
 ib_umad-y :=                   user_mad.o
diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
new file mode 100644 (file)
index 0000000..a76834e
--- /dev/null
@@ -0,0 +1,1927 @@
+/*
+ * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
+ * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
+ * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
+ * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
+ *
+ * This Software is licensed under one of the following licenses:
+ *
+ * 1) under the terms of the "Common Public License 1.0" a copy of which is
+ *    available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/cpl.php.
+ *
+ * 2) under the terms of the "The BSD License" a copy of which is
+ *    available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/bsd-license.php.
+ *
+ * 3) under the terms of the "GNU General Public License (GPL) Version 2" a
+ *    copy of which is available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/gpl-license.php.
+ *
+ * Licensee has the right to choose one of the above licenses.
+ *
+ * Redistributions of source code must retain the above copyright
+ * notice and one of the license notices.
+ *
+ * Redistributions in binary form must reproduce both the above copyright
+ * notice, one of the license notices in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ */
+
+#include <linux/completion.h>
+#include <linux/in.h>
+#include <linux/in6.h>
+#include <linux/mutex.h>
+#include <linux/random.h>
+#include <linux/idr.h>
+
+#include <net/tcp.h>
+
+#include <rdma/rdma_cm.h>
+#include <rdma/rdma_cm_ib.h>
+#include <rdma/ib_cache.h>
+#include <rdma/ib_cm.h>
+#include <rdma/ib_sa.h>
+
+MODULE_AUTHOR("Sean Hefty");
+MODULE_DESCRIPTION("Generic RDMA CM Agent");
+MODULE_LICENSE("Dual BSD/GPL");
+
+#define CMA_CM_RESPONSE_TIMEOUT 20
+#define CMA_MAX_CM_RETRIES 3
+
+static void cma_add_one(struct ib_device *device);
+static void cma_remove_one(struct ib_device *device);
+
+static struct ib_client cma_client = {
+       .name   = "cma",
+       .add    = cma_add_one,
+       .remove = cma_remove_one
+};
+
+static LIST_HEAD(dev_list);
+static LIST_HEAD(listen_any_list);
+static DEFINE_MUTEX(lock);
+static struct workqueue_struct *cma_wq;
+static DEFINE_IDR(sdp_ps);
+static DEFINE_IDR(tcp_ps);
+
+struct cma_device {
+       struct list_head        list;
+       struct ib_device        *device;
+       __be64                  node_guid;
+       struct completion       comp;
+       atomic_t                refcount;
+       struct list_head        id_list;
+};
+
+enum cma_state {
+       CMA_IDLE,
+       CMA_ADDR_QUERY,
+       CMA_ADDR_RESOLVED,
+       CMA_ROUTE_QUERY,
+       CMA_ROUTE_RESOLVED,
+       CMA_CONNECT,
+       CMA_DISCONNECT,
+       CMA_ADDR_BOUND,
+       CMA_LISTEN,
+       CMA_DEVICE_REMOVAL,
+       CMA_DESTROYING
+};
+
+struct rdma_bind_list {
+       struct idr              *ps;
+       struct hlist_head       owners;
+       unsigned short          port;
+};
+
+/*
+ * Device removal can occur at anytime, so we need extra handling to
+ * serialize notifying the user of device removal with other callbacks.
+ * We do this by disabling removal notification while a callback is in process,
+ * and reporting it after the callback completes.
+ */
+struct rdma_id_private {
+       struct rdma_cm_id       id;
+
+       struct rdma_bind_list   *bind_list;
+       struct hlist_node       node;
+       struct list_head        list;
+       struct list_head        listen_list;
+       struct cma_device       *cma_dev;
+
+       enum cma_state          state;
+       spinlock_t              lock;
+       struct completion       comp;
+       atomic_t                refcount;
+       wait_queue_head_t       wait_remove;
+       atomic_t                dev_remove;
+
+       int                     backlog;
+       int                     timeout_ms;
+       struct ib_sa_query      *query;
+       int                     query_id;
+       union {
+               struct ib_cm_id *ib;
+       } cm_id;
+
+       u32                     seq_num;
+       u32                     qp_num;
+       enum ib_qp_type         qp_type;
+       u8                      srq;
+};
+
+struct cma_work {
+       struct work_struct      work;
+       struct rdma_id_private  *id;
+       enum cma_state          old_state;
+       enum cma_state          new_state;
+       struct rdma_cm_event    event;
+};
+
+union cma_ip_addr {
+       struct in6_addr ip6;
+       struct {
+               __u32 pad[3];
+               __u32 addr;
+       } ip4;
+};
+
+struct cma_hdr {
+       u8 cma_version;
+       u8 ip_version;  /* IP version: 7:4 */
+       __u16 port;
+       union cma_ip_addr src_addr;
+       union cma_ip_addr dst_addr;
+};
+
+struct sdp_hh {
+       u8 bsdh[16];
+       u8 sdp_version; /* Major version: 7:4 */
+       u8 ip_version;  /* IP version: 7:4 */
+       u8 sdp_specific1[10];
+       __u16 port;
+       __u16 sdp_specific2;
+       union cma_ip_addr src_addr;
+       union cma_ip_addr dst_addr;
+};
+
+struct sdp_hah {
+       u8 bsdh[16];
+       u8 sdp_version;
+};
+
+#define CMA_VERSION 0x00
+#define SDP_MAJ_VERSION 0x2
+
+static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
+{
+       unsigned long flags;
+       int ret;
+
+       spin_lock_irqsave(&id_priv->lock, flags);
+       ret = (id_priv->state == comp);
+       spin_unlock_irqrestore(&id_priv->lock, flags);
+       return ret;
+}
+
+static int cma_comp_exch(struct rdma_id_private *id_priv,
+                        enum cma_state comp, enum cma_state exch)
+{
+       unsigned long flags;
+       int ret;
+
+       spin_lock_irqsave(&id_priv->lock, flags);
+       if ((ret = (id_priv->state == comp)))
+               id_priv->state = exch;
+       spin_unlock_irqrestore(&id_priv->lock, flags);
+       return ret;
+}
+
+static enum cma_state cma_exch(struct rdma_id_private *id_priv,
+                              enum cma_state exch)
+{
+       unsigned long flags;
+       enum cma_state old;
+
+       spin_lock_irqsave(&id_priv->lock, flags);
+       old = id_priv->state;
+       id_priv->state = exch;
+       spin_unlock_irqrestore(&id_priv->lock, flags);
+       return old;
+}
+
+static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
+{
+       return hdr->ip_version >> 4;
+}
+
+static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
+{
+       hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
+}
+
+static inline u8 sdp_get_majv(u8 sdp_version)
+{
+       return sdp_version >> 4;
+}
+
+static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
+{
+       return hh->ip_version >> 4;
+}
+
+static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
+{
+       hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
+}
+
+static void cma_attach_to_dev(struct rdma_id_private *id_priv,
+                             struct cma_device *cma_dev)
+{
+       atomic_inc(&cma_dev->refcount);
+       id_priv->cma_dev = cma_dev;
+       id_priv->id.device = cma_dev->device;
+       list_add_tail(&id_priv->list, &cma_dev->id_list);
+}
+
+static inline void cma_deref_dev(struct cma_device *cma_dev)
+{
+       if (atomic_dec_and_test(&cma_dev->refcount))
+               complete(&cma_dev->comp);
+}
+
+static void cma_detach_from_dev(struct rdma_id_private *id_priv)
+{
+       list_del(&id_priv->list);
+       cma_deref_dev(id_priv->cma_dev);
+       id_priv->cma_dev = NULL;
+}
+
+static int cma_acquire_ib_dev(struct rdma_id_private *id_priv)
+{
+       struct cma_device *cma_dev;
+       union ib_gid *gid;
+       int ret = -ENODEV;
+
+       gid = ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr);
+
+       mutex_lock(&lock);
+       list_for_each_entry(cma_dev, &dev_list, list) {
+               ret = ib_find_cached_gid(cma_dev->device, gid,
+                                        &id_priv->id.port_num, NULL);
+               if (!ret) {
+                       cma_attach_to_dev(id_priv, cma_dev);
+                       break;
+               }
+       }
+       mutex_unlock(&lock);
+       return ret;
+}
+
+static int cma_acquire_dev(struct rdma_id_private *id_priv)
+{
+       switch (id_priv->id.route.addr.dev_addr.dev_type) {
+       case IB_NODE_CA:
+               return cma_acquire_ib_dev(id_priv);
+       default:
+               return -ENODEV;
+       }
+}
+
+static void cma_deref_id(struct rdma_id_private *id_priv)
+{
+       if (atomic_dec_and_test(&id_priv->refcount))
+               complete(&id_priv->comp);
+}
+
+static void cma_release_remove(struct rdma_id_private *id_priv)
+{
+       if (atomic_dec_and_test(&id_priv->dev_remove))
+               wake_up(&id_priv->wait_remove);
+}
+
+struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
+                                 void *context, enum rdma_port_space ps)
+{
+       struct rdma_id_private *id_priv;
+
+       id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
+       if (!id_priv)
+               return ERR_PTR(-ENOMEM);
+
+       id_priv->state = CMA_IDLE;
+       id_priv->id.context = context;
+       id_priv->id.event_handler = event_handler;
+       id_priv->id.ps = ps;
+       spin_lock_init(&id_priv->lock);
+       init_completion(&id_priv->comp);
+       atomic_set(&id_priv->refcount, 1);
+       init_waitqueue_head(&id_priv->wait_remove);
+       atomic_set(&id_priv->dev_remove, 0);
+       INIT_LIST_HEAD(&id_priv->listen_list);
+       get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
+
+       return &id_priv->id;
+}
+EXPORT_SYMBOL(rdma_create_id);
+
+static int cma_init_ib_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
+{
+       struct ib_qp_attr qp_attr;
+       struct rdma_dev_addr *dev_addr;
+       int ret;
+
+       dev_addr = &id_priv->id.route.addr.dev_addr;
+       ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
+                                 ib_addr_get_pkey(dev_addr),
+                                 &qp_attr.pkey_index);
+       if (ret)
+               return ret;
+
+       qp_attr.qp_state = IB_QPS_INIT;
+       qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
+       qp_attr.port_num = id_priv->id.port_num;
+       return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS |
+                                         IB_QP_PKEY_INDEX | IB_QP_PORT);
+}
+
+int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
+                  struct ib_qp_init_attr *qp_init_attr)
+{
+       struct rdma_id_private *id_priv;
+       struct ib_qp *qp;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (id->device != pd->device)
+               return -EINVAL;
+
+       qp = ib_create_qp(pd, qp_init_attr);
+       if (IS_ERR(qp))
+               return PTR_ERR(qp);
+
+       switch (id->device->node_type) {
+       case IB_NODE_CA:
+               ret = cma_init_ib_qp(id_priv, qp);
+               break;
+       default:
+               ret = -ENOSYS;
+               break;
+       }
+
+       if (ret)
+               goto err;
+
+       id->qp = qp;
+       id_priv->qp_num = qp->qp_num;
+       id_priv->qp_type = qp->qp_type;
+       id_priv->srq = (qp->srq != NULL);
+       return 0;
+err:
+       ib_destroy_qp(qp);
+       return ret;
+}
+EXPORT_SYMBOL(rdma_create_qp);
+
+void rdma_destroy_qp(struct rdma_cm_id *id)
+{
+       ib_destroy_qp(id->qp);
+}
+EXPORT_SYMBOL(rdma_destroy_qp);
+
+static int cma_modify_qp_rtr(struct rdma_cm_id *id)
+{
+       struct ib_qp_attr qp_attr;
+       int qp_attr_mask, ret;
+
+       if (!id->qp)
+               return 0;
+
+       /* Need to update QP attributes from default values. */
+       qp_attr.qp_state = IB_QPS_INIT;
+       ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
+       if (ret)
+               return ret;
+
+       ret = ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
+       if (ret)
+               return ret;
+
+       qp_attr.qp_state = IB_QPS_RTR;
+       ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
+       if (ret)
+               return ret;
+
+       return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
+}
+
+static int cma_modify_qp_rts(struct rdma_cm_id *id)
+{
+       struct ib_qp_attr qp_attr;
+       int qp_attr_mask, ret;
+
+       if (!id->qp)
+               return 0;
+
+       qp_attr.qp_state = IB_QPS_RTS;
+       ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
+       if (ret)
+               return ret;
+
+       return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
+}
+
+static int cma_modify_qp_err(struct rdma_cm_id *id)
+{
+       struct ib_qp_attr qp_attr;
+
+       if (!id->qp)
+               return 0;
+
+       qp_attr.qp_state = IB_QPS_ERR;
+       return ib_modify_qp(id->qp, &qp_attr, IB_QP_STATE);
+}
+
+int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
+                      int *qp_attr_mask)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       switch (id_priv->id.device->node_type) {
+       case IB_NODE_CA:
+               ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
+                                        qp_attr_mask);
+               if (qp_attr->qp_state == IB_QPS_RTR)
+                       qp_attr->rq_psn = id_priv->seq_num;
+               break;
+       default:
+               ret = -ENOSYS;
+               break;
+       }
+
+       return ret;
+}
+EXPORT_SYMBOL(rdma_init_qp_attr);
+
+static inline int cma_zero_addr(struct sockaddr *addr)
+{
+       struct in6_addr *ip6;
+
+       if (addr->sa_family == AF_INET)
+               return ZERONET(((struct sockaddr_in *) addr)->sin_addr.s_addr);
+       else {
+               ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
+               return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
+                       ip6->s6_addr32[3] | ip6->s6_addr32[4]) == 0;
+       }
+}
+
+static inline int cma_loopback_addr(struct sockaddr *addr)
+{
+       return LOOPBACK(((struct sockaddr_in *) addr)->sin_addr.s_addr);
+}
+
+static inline int cma_any_addr(struct sockaddr *addr)
+{
+       return cma_zero_addr(addr) || cma_loopback_addr(addr);
+}
+
+static inline int cma_any_port(struct sockaddr *addr)
+{
+       return !((struct sockaddr_in *) addr)->sin_port;
+}
+
+static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
+                           u8 *ip_ver, __u16 *port,
+                           union cma_ip_addr **src, union cma_ip_addr **dst)
+{
+       switch (ps) {
+       case RDMA_PS_SDP:
+               if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
+                   SDP_MAJ_VERSION)
+                       return -EINVAL;
+
+               *ip_ver = sdp_get_ip_ver(hdr);
+               *port   = ((struct sdp_hh *) hdr)->port;
+               *src    = &((struct sdp_hh *) hdr)->src_addr;
+               *dst    = &((struct sdp_hh *) hdr)->dst_addr;
+               break;
+       default:
+               if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
+                       return -EINVAL;
+
+               *ip_ver = cma_get_ip_ver(hdr);
+               *port   = ((struct cma_hdr *) hdr)->port;
+               *src    = &((struct cma_hdr *) hdr)->src_addr;
+               *dst    = &((struct cma_hdr *) hdr)->dst_addr;
+               break;
+       }
+
+       if (*ip_ver != 4 && *ip_ver != 6)
+               return -EINVAL;
+       return 0;
+}
+
+static void cma_save_net_info(struct rdma_addr *addr,
+                             struct rdma_addr *listen_addr,
+                             u8 ip_ver, __u16 port,
+                             union cma_ip_addr *src, union cma_ip_addr *dst)
+{
+       struct sockaddr_in *listen4, *ip4;
+       struct sockaddr_in6 *listen6, *ip6;
+
+       switch (ip_ver) {
+       case 4:
+               listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
+               ip4 = (struct sockaddr_in *) &addr->src_addr;
+               ip4->sin_family = listen4->sin_family;
+               ip4->sin_addr.s_addr = dst->ip4.addr;
+               ip4->sin_port = listen4->sin_port;
+
+               ip4 = (struct sockaddr_in *) &addr->dst_addr;
+               ip4->sin_family = listen4->sin_family;
+               ip4->sin_addr.s_addr = src->ip4.addr;
+               ip4->sin_port = port;
+               break;
+       case 6:
+               listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
+               ip6 = (struct sockaddr_in6 *) &addr->src_addr;
+               ip6->sin6_family = listen6->sin6_family;
+               ip6->sin6_addr = dst->ip6;
+               ip6->sin6_port = listen6->sin6_port;
+
+               ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
+               ip6->sin6_family = listen6->sin6_family;
+               ip6->sin6_addr = src->ip6;
+               ip6->sin6_port = port;
+               break;
+       default:
+               break;
+       }
+}
+
+static inline int cma_user_data_offset(enum rdma_port_space ps)
+{
+       switch (ps) {
+       case RDMA_PS_SDP:
+               return 0;
+       default:
+               return sizeof(struct cma_hdr);
+       }
+}
+
+static int cma_notify_user(struct rdma_id_private *id_priv,
+                          enum rdma_cm_event_type type, int status,
+                          void *data, u8 data_len)
+{
+       struct rdma_cm_event event;
+
+       event.event = type;
+       event.status = status;
+       event.private_data = data;
+       event.private_data_len = data_len;
+
+       return id_priv->id.event_handler(&id_priv->id, &event);
+}
+
+static void cma_cancel_route(struct rdma_id_private *id_priv)
+{
+       switch (id_priv->id.device->node_type) {
+       case IB_NODE_CA:
+               if (id_priv->query)
+                       ib_sa_cancel_query(id_priv->query_id, id_priv->query);
+               break;
+       default:
+               break;
+       }
+}
+
+static inline int cma_internal_listen(struct rdma_id_private *id_priv)
+{
+       return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
+              cma_any_addr(&id_priv->id.route.addr.src_addr);
+}
+
+static void cma_destroy_listen(struct rdma_id_private *id_priv)
+{
+       cma_exch(id_priv, CMA_DESTROYING);
+
+       if (id_priv->cma_dev) {
+               switch (id_priv->id.device->node_type) {
+               case IB_NODE_CA:
+                       if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
+                               ib_destroy_cm_id(id_priv->cm_id.ib);
+                       break;
+               default:
+                       break;
+               }
+               cma_detach_from_dev(id_priv);
+       }
+       list_del(&id_priv->listen_list);
+
+       cma_deref_id(id_priv);
+       wait_for_completion(&id_priv->comp);
+
+       kfree(id_priv);
+}
+
+static void cma_cancel_listens(struct rdma_id_private *id_priv)
+{
+       struct rdma_id_private *dev_id_priv;
+
+       mutex_lock(&lock);
+       list_del(&id_priv->list);
+
+       while (!list_empty(&id_priv->listen_list)) {
+               dev_id_priv = list_entry(id_priv->listen_list.next,
+                                        struct rdma_id_private, listen_list);
+               cma_destroy_listen(dev_id_priv);
+       }
+       mutex_unlock(&lock);
+}
+
+static void cma_cancel_operation(struct rdma_id_private *id_priv,
+                                enum cma_state state)
+{
+       switch (state) {
+       case CMA_ADDR_QUERY:
+               rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
+               break;
+       case CMA_ROUTE_QUERY:
+               cma_cancel_route(id_priv);
+               break;
+       case CMA_LISTEN:
+               if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
+                   !id_priv->cma_dev)
+                       cma_cancel_listens(id_priv);
+               break;
+       default:
+               break;
+       }
+}
+
+static void cma_release_port(struct rdma_id_private *id_priv)
+{
+       struct rdma_bind_list *bind_list = id_priv->bind_list;
+
+       if (!bind_list)
+               return;
+
+       mutex_lock(&lock);
+       hlist_del(&id_priv->node);
+       if (hlist_empty(&bind_list->owners)) {
+               idr_remove(bind_list->ps, bind_list->port);
+               kfree(bind_list);
+       }
+       mutex_unlock(&lock);
+}
+
+void rdma_destroy_id(struct rdma_cm_id *id)
+{
+       struct rdma_id_private *id_priv;
+       enum cma_state state;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       state = cma_exch(id_priv, CMA_DESTROYING);
+       cma_cancel_operation(id_priv, state);
+
+       if (id_priv->cma_dev) {
+               switch (id->device->node_type) {
+               case IB_NODE_CA:
+                       if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
+                               ib_destroy_cm_id(id_priv->cm_id.ib);
+                       break;
+               default:
+                       break;
+               }
+               mutex_lock(&lock);
+               cma_detach_from_dev(id_priv);
+               mutex_unlock(&lock);
+       }
+
+       cma_release_port(id_priv);
+       cma_deref_id(id_priv);
+       wait_for_completion(&id_priv->comp);
+
+       kfree(id_priv->id.route.path_rec);
+       kfree(id_priv);
+}
+EXPORT_SYMBOL(rdma_destroy_id);
+
+static int cma_rep_recv(struct rdma_id_private *id_priv)
+{
+       int ret;
+
+       ret = cma_modify_qp_rtr(&id_priv->id);
+       if (ret)
+               goto reject;
+
+       ret = cma_modify_qp_rts(&id_priv->id);
+       if (ret)
+               goto reject;
+
+       ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
+       if (ret)
+               goto reject;
+
+       return 0;
+reject:
+       cma_modify_qp_err(&id_priv->id);
+       ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
+                      NULL, 0, NULL, 0);
+       return ret;
+}
+
+static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
+{
+       if (id_priv->id.ps == RDMA_PS_SDP &&
+           sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
+           SDP_MAJ_VERSION)
+               return -EINVAL;
+
+       return 0;
+}
+
+static int cma_rtu_recv(struct rdma_id_private *id_priv)
+{
+       int ret;
+
+       ret = cma_modify_qp_rts(&id_priv->id);
+       if (ret)
+               goto reject;
+
+       return 0;
+reject:
+       cma_modify_qp_err(&id_priv->id);
+       ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
+                      NULL, 0, NULL, 0);
+       return ret;
+}
+
+static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
+{
+       struct rdma_id_private *id_priv = cm_id->context;
+       enum rdma_cm_event_type event;
+       u8 private_data_len = 0;
+       int ret = 0, status = 0;
+
+       atomic_inc(&id_priv->dev_remove);
+       if (!cma_comp(id_priv, CMA_CONNECT))
+               goto out;
+
+       switch (ib_event->event) {
+       case IB_CM_REQ_ERROR:
+       case IB_CM_REP_ERROR:
+               event = RDMA_CM_EVENT_UNREACHABLE;
+               status = -ETIMEDOUT;
+               break;
+       case IB_CM_REP_RECEIVED:
+               status = cma_verify_rep(id_priv, ib_event->private_data);
+               if (status)
+                       event = RDMA_CM_EVENT_CONNECT_ERROR;
+               else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
+                       status = cma_rep_recv(id_priv);
+                       event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
+                                        RDMA_CM_EVENT_ESTABLISHED;
+               } else
+                       event = RDMA_CM_EVENT_CONNECT_RESPONSE;
+               private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
+               break;
+       case IB_CM_RTU_RECEIVED:
+               status = cma_rtu_recv(id_priv);
+               event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
+                                RDMA_CM_EVENT_ESTABLISHED;
+               break;
+       case IB_CM_DREQ_ERROR:
+               status = -ETIMEDOUT; /* fall through */
+       case IB_CM_DREQ_RECEIVED:
+       case IB_CM_DREP_RECEIVED:
+               if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
+                       goto out;
+               event = RDMA_CM_EVENT_DISCONNECTED;
+               break;
+       case IB_CM_TIMEWAIT_EXIT:
+       case IB_CM_MRA_RECEIVED:
+               /* ignore event */
+               goto out;
+       case IB_CM_REJ_RECEIVED:
+               cma_modify_qp_err(&id_priv->id);
+               status = ib_event->param.rej_rcvd.reason;
+               event = RDMA_CM_EVENT_REJECTED;
+               break;
+       default:
+               printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
+                      ib_event->event);
+               goto out;
+       }
+
+       ret = cma_notify_user(id_priv, event, status, ib_event->private_data,
+                             private_data_len);
+       if (ret) {
+               /* Destroy the CM ID by returning a non-zero value. */
+               id_priv->cm_id.ib = NULL;
+               cma_exch(id_priv, CMA_DESTROYING);
+               cma_release_remove(id_priv);
+               rdma_destroy_id(&id_priv->id);
+               return ret;
+       }
+out:
+       cma_release_remove(id_priv);
+       return ret;
+}
+
+static struct rdma_id_private *cma_new_id(struct rdma_cm_id *listen_id,
+                                         struct ib_cm_event *ib_event)
+{
+       struct rdma_id_private *id_priv;
+       struct rdma_cm_id *id;
+       struct rdma_route *rt;
+       union cma_ip_addr *src, *dst;
+       __u16 port;
+       u8 ip_ver;
+
+       id = rdma_create_id(listen_id->event_handler, listen_id->context,
+                           listen_id->ps);
+       if (IS_ERR(id))
+               return NULL;
+
+       rt = &id->route;
+       rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
+       rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, GFP_KERNEL);
+       if (!rt->path_rec)
+               goto err;
+
+       if (cma_get_net_info(ib_event->private_data, listen_id->ps,
+                            &ip_ver, &port, &src, &dst))
+               goto err;
+
+       cma_save_net_info(&id->route.addr, &listen_id->route.addr,
+                         ip_ver, port, src, dst);
+       rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
+       if (rt->num_paths == 2)
+               rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
+
+       ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
+       ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
+       ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
+       rt->addr.dev_addr.dev_type = IB_NODE_CA;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       id_priv->state = CMA_CONNECT;
+       return id_priv;
+err:
+       rdma_destroy_id(id);
+       return NULL;
+}
+
+static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
+{
+       struct rdma_id_private *listen_id, *conn_id;
+       int offset, ret;
+
+       listen_id = cm_id->context;
+       atomic_inc(&listen_id->dev_remove);
+       if (!cma_comp(listen_id, CMA_LISTEN)) {
+               ret = -ECONNABORTED;
+               goto out;
+       }
+
+       conn_id = cma_new_id(&listen_id->id, ib_event);
+       if (!conn_id) {
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       atomic_inc(&conn_id->dev_remove);
+       ret = cma_acquire_ib_dev(conn_id);
+       if (ret) {
+               ret = -ENODEV;
+               cma_release_remove(conn_id);
+               rdma_destroy_id(&conn_id->id);
+               goto out;
+       }
+
+       conn_id->cm_id.ib = cm_id;
+       cm_id->context = conn_id;
+       cm_id->cm_handler = cma_ib_handler;
+
+       offset = cma_user_data_offset(listen_id->id.ps);
+       ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
+                             ib_event->private_data + offset,
+                             IB_CM_REQ_PRIVATE_DATA_SIZE - offset);
+       if (ret) {
+               /* Destroy the CM ID by returning a non-zero value. */
+               conn_id->cm_id.ib = NULL;
+               cma_exch(conn_id, CMA_DESTROYING);
+               cma_release_remove(conn_id);
+               rdma_destroy_id(&conn_id->id);
+       }
+out:
+       cma_release_remove(listen_id);
+       return ret;
+}
+
+static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
+{
+       return cpu_to_be64(((u64)ps << 16) +
+              be16_to_cpu(((struct sockaddr_in *) addr)->sin_port));
+}
+
+static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
+                                struct ib_cm_compare_data *compare)
+{
+       struct cma_hdr *cma_data, *cma_mask;
+       struct sdp_hh *sdp_data, *sdp_mask;
+       __u32 ip4_addr;
+       struct in6_addr ip6_addr;
+
+       memset(compare, 0, sizeof *compare);
+       cma_data = (void *) compare->data;
+       cma_mask = (void *) compare->mask;
+       sdp_data = (void *) compare->data;
+       sdp_mask = (void *) compare->mask;
+
+       switch (addr->sa_family) {
+       case AF_INET:
+               ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
+               if (ps == RDMA_PS_SDP) {
+                       sdp_set_ip_ver(sdp_data, 4);
+                       sdp_set_ip_ver(sdp_mask, 0xF);
+                       sdp_data->dst_addr.ip4.addr = ip4_addr;
+                       sdp_mask->dst_addr.ip4.addr = ~0;
+               } else {
+                       cma_set_ip_ver(cma_data, 4);
+                       cma_set_ip_ver(cma_mask, 0xF);
+                       cma_data->dst_addr.ip4.addr = ip4_addr;
+                       cma_mask->dst_addr.ip4.addr = ~0;
+               }
+               break;
+       case AF_INET6:
+               ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
+               if (ps == RDMA_PS_SDP) {
+                       sdp_set_ip_ver(sdp_data, 6);
+                       sdp_set_ip_ver(sdp_mask, 0xF);
+                       sdp_data->dst_addr.ip6 = ip6_addr;
+                       memset(&sdp_mask->dst_addr.ip6, 0xFF,
+                              sizeof sdp_mask->dst_addr.ip6);
+               } else {
+                       cma_set_ip_ver(cma_data, 6);
+                       cma_set_ip_ver(cma_mask, 0xF);
+                       cma_data->dst_addr.ip6 = ip6_addr;
+                       memset(&cma_mask->dst_addr.ip6, 0xFF,
+                              sizeof cma_mask->dst_addr.ip6);
+               }
+               break;
+       default:
+               break;
+       }
+}
+
+static int cma_ib_listen(struct rdma_id_private *id_priv)
+{
+       struct ib_cm_compare_data compare_data;
+       struct sockaddr *addr;
+       __be64 svc_id;
+       int ret;
+
+       id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
+                                           id_priv);
+       if (IS_ERR(id_priv->cm_id.ib))
+               return PTR_ERR(id_priv->cm_id.ib);
+
+       addr = &id_priv->id.route.addr.src_addr;
+       svc_id = cma_get_service_id(id_priv->id.ps, addr);
+       if (cma_any_addr(addr))
+               ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
+       else {
+               cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
+               ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
+       }
+
+       if (ret) {
+               ib_destroy_cm_id(id_priv->cm_id.ib);
+               id_priv->cm_id.ib = NULL;
+       }
+
+       return ret;
+}
+
+static int cma_listen_handler(struct rdma_cm_id *id,
+                             struct rdma_cm_event *event)
+{
+       struct rdma_id_private *id_priv = id->context;
+
+       id->context = id_priv->id.context;
+       id->event_handler = id_priv->id.event_handler;
+       return id_priv->id.event_handler(id, event);
+}
+
+static void cma_listen_on_dev(struct rdma_id_private *id_priv,
+                             struct cma_device *cma_dev)
+{
+       struct rdma_id_private *dev_id_priv;
+       struct rdma_cm_id *id;
+       int ret;
+
+       id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
+       if (IS_ERR(id))
+               return;
+
+       dev_id_priv = container_of(id, struct rdma_id_private, id);
+
+       dev_id_priv->state = CMA_ADDR_BOUND;
+       memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
+              ip_addr_size(&id_priv->id.route.addr.src_addr));
+
+       cma_attach_to_dev(dev_id_priv, cma_dev);
+       list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
+
+       ret = rdma_listen(id, id_priv->backlog);
+       if (ret)
+               goto err;
+
+       return;
+err:
+       cma_destroy_listen(dev_id_priv);
+}
+
+static void cma_listen_on_all(struct rdma_id_private *id_priv)
+{
+       struct cma_device *cma_dev;
+
+       mutex_lock(&lock);
+       list_add_tail(&id_priv->list, &listen_any_list);
+       list_for_each_entry(cma_dev, &dev_list, list)
+               cma_listen_on_dev(id_priv, cma_dev);
+       mutex_unlock(&lock);
+}
+
+static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
+{
+       struct sockaddr_in addr_in;
+
+       memset(&addr_in, 0, sizeof addr_in);
+       addr_in.sin_family = af;
+       return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
+}
+
+int rdma_listen(struct rdma_cm_id *id, int backlog)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (id_priv->state == CMA_IDLE) {
+               ret = cma_bind_any(id, AF_INET);
+               if (ret)
+                       return ret;
+       }
+
+       if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
+               return -EINVAL;
+
+       id_priv->backlog = backlog;
+       if (id->device) {
+               switch (id->device->node_type) {
+               case IB_NODE_CA:
+                       ret = cma_ib_listen(id_priv);
+                       if (ret)
+                               goto err;
+                       break;
+               default:
+                       ret = -ENOSYS;
+                       goto err;
+               }
+       } else
+               cma_listen_on_all(id_priv);
+
+       return 0;
+err:
+       id_priv->backlog = 0;
+       cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
+       return ret;
+}
+EXPORT_SYMBOL(rdma_listen);
+
+static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
+                             void *context)
+{
+       struct cma_work *work = context;
+       struct rdma_route *route;
+
+       route = &work->id->id.route;
+
+       if (!status) {
+               route->num_paths = 1;
+               *route->path_rec = *path_rec;
+       } else {
+               work->old_state = CMA_ROUTE_QUERY;
+               work->new_state = CMA_ADDR_RESOLVED;
+               work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
+       }
+
+       queue_work(cma_wq, &work->work);
+}
+
+static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
+                             struct cma_work *work)
+{
+       struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
+       struct ib_sa_path_rec path_rec;
+
+       memset(&path_rec, 0, sizeof path_rec);
+       path_rec.sgid = *ib_addr_get_sgid(addr);
+       path_rec.dgid = *ib_addr_get_dgid(addr);
+       path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
+       path_rec.numb_path = 1;
+
+       id_priv->query_id = ib_sa_path_rec_get(id_priv->id.device,
+                               id_priv->id.port_num, &path_rec,
+                               IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
+                               IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
+                               timeout_ms, GFP_KERNEL,
+                               cma_query_handler, work, &id_priv->query);
+
+       return (id_priv->query_id < 0) ? id_priv->query_id : 0;
+}
+
+static void cma_work_handler(void *data)
+{
+       struct cma_work *work = data;
+       struct rdma_id_private *id_priv = work->id;
+       int destroy = 0;
+
+       atomic_inc(&id_priv->dev_remove);
+       if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
+               goto out;
+
+       if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
+               cma_exch(id_priv, CMA_DESTROYING);
+               destroy = 1;
+       }
+out:
+       cma_release_remove(id_priv);
+       cma_deref_id(id_priv);
+       if (destroy)
+               rdma_destroy_id(&id_priv->id);
+       kfree(work);
+}
+
+static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
+{
+       struct rdma_route *route = &id_priv->id.route;
+       struct cma_work *work;
+       int ret;
+
+       work = kzalloc(sizeof *work, GFP_KERNEL);
+       if (!work)
+               return -ENOMEM;
+
+       work->id = id_priv;
+       INIT_WORK(&work->work, cma_work_handler, work);
+       work->old_state = CMA_ROUTE_QUERY;
+       work->new_state = CMA_ROUTE_RESOLVED;
+       work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
+
+       route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
+       if (!route->path_rec) {
+               ret = -ENOMEM;
+               goto err1;
+       }
+
+       ret = cma_query_ib_route(id_priv, timeout_ms, work);
+       if (ret)
+               goto err2;
+
+       return 0;
+err2:
+       kfree(route->path_rec);
+       route->path_rec = NULL;
+err1:
+       kfree(work);
+       return ret;
+}
+
+int rdma_set_ib_paths(struct rdma_cm_id *id,
+                     struct ib_sa_path_rec *path_rec, int num_paths)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
+               return -EINVAL;
+
+       id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
+       if (!id->route.path_rec) {
+               ret = -ENOMEM;
+               goto err;
+       }
+
+       memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
+       return 0;
+err:
+       cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
+       return ret;
+}
+EXPORT_SYMBOL(rdma_set_ib_paths);
+
+int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
+               return -EINVAL;
+
+       atomic_inc(&id_priv->refcount);
+       switch (id->device->node_type) {
+       case IB_NODE_CA:
+               ret = cma_resolve_ib_route(id_priv, timeout_ms);
+               break;
+       default:
+               ret = -ENOSYS;
+               break;
+       }
+       if (ret)
+               goto err;
+
+       return 0;
+err:
+       cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
+       cma_deref_id(id_priv);
+       return ret;
+}
+EXPORT_SYMBOL(rdma_resolve_route);
+
+static int cma_bind_loopback(struct rdma_id_private *id_priv)
+{
+       struct cma_device *cma_dev;
+       struct ib_port_attr port_attr;
+       union ib_gid *gid;
+       u16 pkey;
+       int ret;
+       u8 p;
+
+       mutex_lock(&lock);
+       list_for_each_entry(cma_dev, &dev_list, list)
+               for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
+                       if (!ib_query_port (cma_dev->device, p, &port_attr) &&
+                           port_attr.state == IB_PORT_ACTIVE)
+                               goto port_found;
+
+       if (!list_empty(&dev_list)) {
+               p = 1;
+               cma_dev = list_entry(dev_list.next, struct cma_device, list);
+       } else {
+               ret = -ENODEV;
+               goto out;
+       }
+
+port_found:
+       gid = ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr);
+       ret = ib_get_cached_gid(cma_dev->device, p, 0, gid);
+       if (ret)
+               goto out;
+
+       ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
+       if (ret)
+               goto out;
+
+       ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
+       id_priv->id.port_num = p;
+       cma_attach_to_dev(id_priv, cma_dev);
+out:
+       mutex_unlock(&lock);
+       return ret;
+}
+
+static void addr_handler(int status, struct sockaddr *src_addr,
+                        struct rdma_dev_addr *dev_addr, void *context)
+{
+       struct rdma_id_private *id_priv = context;
+       enum rdma_cm_event_type event;
+
+       atomic_inc(&id_priv->dev_remove);
+       if (!id_priv->cma_dev && !status)
+               status = cma_acquire_dev(id_priv);
+
+       if (status) {
+               if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND))
+                       goto out;
+               event = RDMA_CM_EVENT_ADDR_ERROR;
+       } else {
+               if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED))
+                       goto out;
+               memcpy(&id_priv->id.route.addr.src_addr, src_addr,
+                      ip_addr_size(src_addr));
+               event = RDMA_CM_EVENT_ADDR_RESOLVED;
+       }
+
+       if (cma_notify_user(id_priv, event, status, NULL, 0)) {
+               cma_exch(id_priv, CMA_DESTROYING);
+               cma_release_remove(id_priv);
+               cma_deref_id(id_priv);
+               rdma_destroy_id(&id_priv->id);
+               return;
+       }
+out:
+       cma_release_remove(id_priv);
+       cma_deref_id(id_priv);
+}
+
+static int cma_resolve_loopback(struct rdma_id_private *id_priv)
+{
+       struct cma_work *work;
+       struct sockaddr_in *src_in, *dst_in;
+       int ret;
+
+       work = kzalloc(sizeof *work, GFP_KERNEL);
+       if (!work)
+               return -ENOMEM;
+
+       if (!id_priv->cma_dev) {
+               ret = cma_bind_loopback(id_priv);
+               if (ret)
+                       goto err;
+       }
+
+       ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr,
+                        ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr));
+
+       if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
+               src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
+               dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
+               src_in->sin_family = dst_in->sin_family;
+               src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
+       }
+
+       work->id = id_priv;
+       INIT_WORK(&work->work, cma_work_handler, work);
+       work->old_state = CMA_ADDR_QUERY;
+       work->new_state = CMA_ADDR_RESOLVED;
+       work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
+       queue_work(cma_wq, &work->work);
+       return 0;
+err:
+       kfree(work);
+       return ret;
+}
+
+static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
+                        struct sockaddr *dst_addr)
+{
+       if (src_addr && src_addr->sa_family)
+               return rdma_bind_addr(id, src_addr);
+       else
+               return cma_bind_any(id, dst_addr->sa_family);
+}
+
+int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
+                     struct sockaddr *dst_addr, int timeout_ms)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (id_priv->state == CMA_IDLE) {
+               ret = cma_bind_addr(id, src_addr, dst_addr);
+               if (ret)
+                       return ret;
+       }
+
+       if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
+               return -EINVAL;
+
+       atomic_inc(&id_priv->refcount);
+       memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
+       if (cma_any_addr(dst_addr))
+               ret = cma_resolve_loopback(id_priv);
+       else
+               ret = rdma_resolve_ip(&id->route.addr.src_addr, dst_addr,
+                                     &id->route.addr.dev_addr,
+                                     timeout_ms, addr_handler, id_priv);
+       if (ret)
+               goto err;
+
+       return 0;
+err:
+       cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
+       cma_deref_id(id_priv);
+       return ret;
+}
+EXPORT_SYMBOL(rdma_resolve_addr);
+
+static void cma_bind_port(struct rdma_bind_list *bind_list,
+                         struct rdma_id_private *id_priv)
+{
+       struct sockaddr_in *sin;
+
+       sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
+       sin->sin_port = htons(bind_list->port);
+       id_priv->bind_list = bind_list;
+       hlist_add_head(&id_priv->node, &bind_list->owners);
+}
+
+static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
+                         unsigned short snum)
+{
+       struct rdma_bind_list *bind_list;
+       int port, start, ret;
+
+       bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
+       if (!bind_list)
+               return -ENOMEM;
+
+       start = snum ? snum : sysctl_local_port_range[0];
+
+       do {
+               ret = idr_get_new_above(ps, bind_list, start, &port);
+       } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
+
+       if (ret)
+               goto err;
+
+       if ((snum && port != snum) ||
+           (!snum && port > sysctl_local_port_range[1])) {
+               idr_remove(ps, port);
+               ret = -EADDRNOTAVAIL;
+               goto err;
+       }
+
+       bind_list->ps = ps;
+       bind_list->port = (unsigned short) port;
+       cma_bind_port(bind_list, id_priv);
+       return 0;
+err:
+       kfree(bind_list);
+       return ret;
+}
+
+static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
+{
+       struct rdma_id_private *cur_id;
+       struct sockaddr_in *sin, *cur_sin;
+       struct rdma_bind_list *bind_list;
+       struct hlist_node *node;
+       unsigned short snum;
+
+       sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
+       snum = ntohs(sin->sin_port);
+       if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
+               return -EACCES;
+
+       bind_list = idr_find(ps, snum);
+       if (!bind_list)
+               return cma_alloc_port(ps, id_priv, snum);
+
+       /*
+        * We don't support binding to any address if anyone is bound to
+        * a specific address on the same port.
+        */
+       if (cma_any_addr(&id_priv->id.route.addr.src_addr))
+               return -EADDRNOTAVAIL;
+
+       hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
+               if (cma_any_addr(&cur_id->id.route.addr.src_addr))
+                       return -EADDRNOTAVAIL;
+               
+               cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
+               if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
+                       return -EADDRINUSE;
+       }
+
+       cma_bind_port(bind_list, id_priv);
+       return 0;
+}
+
+static int cma_get_port(struct rdma_id_private *id_priv)
+{
+       struct idr *ps;
+       int ret;
+
+       switch (id_priv->id.ps) {
+       case RDMA_PS_SDP:
+               ps = &sdp_ps;
+               break;
+       case RDMA_PS_TCP:
+               ps = &tcp_ps;
+               break;
+       default:
+               return -EPROTONOSUPPORT;
+       }
+
+       mutex_lock(&lock);
+       if (cma_any_port(&id_priv->id.route.addr.src_addr))
+               ret = cma_alloc_port(ps, id_priv, 0);
+       else
+               ret = cma_use_port(ps, id_priv);
+       mutex_unlock(&lock);
+
+       return ret;
+}
+
+int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       if (addr->sa_family != AF_INET)
+               return -EAFNOSUPPORT;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
+               return -EINVAL;
+
+       if (!cma_any_addr(addr)) {
+               ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
+               if (!ret)
+                       ret = cma_acquire_dev(id_priv);
+               if (ret)
+                       goto err;
+       }
+
+       memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
+       ret = cma_get_port(id_priv);
+       if (ret)
+               goto err;
+
+       return 0;
+err:
+       cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
+       return ret;
+}
+EXPORT_SYMBOL(rdma_bind_addr);
+
+static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
+                         struct rdma_route *route)
+{
+       struct sockaddr_in *src4, *dst4;
+       struct cma_hdr *cma_hdr;
+       struct sdp_hh *sdp_hdr;
+
+       src4 = (struct sockaddr_in *) &route->addr.src_addr;
+       dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
+
+       switch (ps) {
+       case RDMA_PS_SDP:
+               sdp_hdr = hdr;
+               if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
+                       return -EINVAL;
+               sdp_set_ip_ver(sdp_hdr, 4);
+               sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
+               sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
+               sdp_hdr->port = src4->sin_port;
+               break;
+       default:
+               cma_hdr = hdr;
+               cma_hdr->cma_version = CMA_VERSION;
+               cma_set_ip_ver(cma_hdr, 4);
+               cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
+               cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
+               cma_hdr->port = src4->sin_port;
+               break;
+       }
+       return 0;
+}
+
+static int cma_connect_ib(struct rdma_id_private *id_priv,
+                         struct rdma_conn_param *conn_param)
+{
+       struct ib_cm_req_param req;
+       struct rdma_route *route;
+       void *private_data;
+       int offset, ret;
+
+       memset(&req, 0, sizeof req);
+       offset = cma_user_data_offset(id_priv->id.ps);
+       req.private_data_len = offset + conn_param->private_data_len;
+       private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
+       if (!private_data)
+               return -ENOMEM;
+
+       if (conn_param->private_data && conn_param->private_data_len)
+               memcpy(private_data + offset, conn_param->private_data,
+                      conn_param->private_data_len);
+
+       id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
+                                           id_priv);
+       if (IS_ERR(id_priv->cm_id.ib)) {
+               ret = PTR_ERR(id_priv->cm_id.ib);
+               goto out;
+       }
+
+       route = &id_priv->id.route;
+       ret = cma_format_hdr(private_data, id_priv->id.ps, route);
+       if (ret)
+               goto out;
+       req.private_data = private_data;
+
+       req.primary_path = &route->path_rec[0];
+       if (route->num_paths == 2)
+               req.alternate_path = &route->path_rec[1];
+
+       req.service_id = cma_get_service_id(id_priv->id.ps,
+                                           &route->addr.dst_addr);
+       req.qp_num = id_priv->qp_num;
+       req.qp_type = id_priv->qp_type;
+       req.starting_psn = id_priv->seq_num;
+       req.responder_resources = conn_param->responder_resources;
+       req.initiator_depth = conn_param->initiator_depth;
+       req.flow_control = conn_param->flow_control;
+       req.retry_count = conn_param->retry_count;
+       req.rnr_retry_count = conn_param->rnr_retry_count;
+       req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
+       req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
+       req.max_cm_retries = CMA_MAX_CM_RETRIES;
+       req.srq = id_priv->srq ? 1 : 0;
+
+       ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
+out:
+       kfree(private_data);
+       return ret;
+}
+
+int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
+               return -EINVAL;
+
+       if (!id->qp) {
+               id_priv->qp_num = conn_param->qp_num;
+               id_priv->qp_type = conn_param->qp_type;
+               id_priv->srq = conn_param->srq;
+       }
+
+       switch (id->device->node_type) {
+       case IB_NODE_CA:
+               ret = cma_connect_ib(id_priv, conn_param);
+               break;
+       default:
+               ret = -ENOSYS;
+               break;
+       }
+       if (ret)
+               goto err;
+
+       return 0;
+err:
+       cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
+       return ret;
+}
+EXPORT_SYMBOL(rdma_connect);
+
+static int cma_accept_ib(struct rdma_id_private *id_priv,
+                        struct rdma_conn_param *conn_param)
+{
+       struct ib_cm_rep_param rep;
+       int ret;
+
+       ret = cma_modify_qp_rtr(&id_priv->id);
+       if (ret)
+               return ret;
+
+       memset(&rep, 0, sizeof rep);
+       rep.qp_num = id_priv->qp_num;
+       rep.starting_psn = id_priv->seq_num;
+       rep.private_data = conn_param->private_data;
+       rep.private_data_len = conn_param->private_data_len;
+       rep.responder_resources = conn_param->responder_resources;
+       rep.initiator_depth = conn_param->initiator_depth;
+       rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
+       rep.failover_accepted = 0;
+       rep.flow_control = conn_param->flow_control;
+       rep.rnr_retry_count = conn_param->rnr_retry_count;
+       rep.srq = id_priv->srq ? 1 : 0;
+
+       return ib_send_cm_rep(id_priv->cm_id.ib, &rep);
+}
+
+int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (!cma_comp(id_priv, CMA_CONNECT))
+               return -EINVAL;
+
+       if (!id->qp && conn_param) {
+               id_priv->qp_num = conn_param->qp_num;
+               id_priv->qp_type = conn_param->qp_type;
+               id_priv->srq = conn_param->srq;
+       }
+
+       switch (id->device->node_type) {
+       case IB_NODE_CA:
+               if (conn_param)
+                       ret = cma_accept_ib(id_priv, conn_param);
+               else
+                       ret = cma_rep_recv(id_priv);
+               break;
+       default:
+               ret = -ENOSYS;
+               break;
+       }
+
+       if (ret)
+               goto reject;
+
+       return 0;
+reject:
+       cma_modify_qp_err(id);
+       rdma_reject(id, NULL, 0);
+       return ret;
+}
+EXPORT_SYMBOL(rdma_accept);
+
+int rdma_reject(struct rdma_cm_id *id, const void *private_data,
+               u8 private_data_len)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (!cma_comp(id_priv, CMA_CONNECT))
+               return -EINVAL;
+
+       switch (id->device->node_type) {
+       case IB_NODE_CA:
+               ret = ib_send_cm_rej(id_priv->cm_id.ib,
+                                    IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
+                                    private_data, private_data_len);
+               break;
+       default:
+               ret = -ENOSYS;
+               break;
+       }
+       return ret;
+}
+EXPORT_SYMBOL(rdma_reject);
+
+int rdma_disconnect(struct rdma_cm_id *id)
+{
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       id_priv = container_of(id, struct rdma_id_private, id);
+       if (!cma_comp(id_priv, CMA_CONNECT) &&
+           !cma_comp(id_priv, CMA_DISCONNECT))
+               return -EINVAL;
+
+       ret = cma_modify_qp_err(id);
+       if (ret)
+               goto out;
+
+       switch (id->device->node_type) {
+       case IB_NODE_CA:
+               /* Initiate or respond to a disconnect. */
+               if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
+                       ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
+               break;
+       default:
+               break;
+       }
+out:
+       return ret;
+}
+EXPORT_SYMBOL(rdma_disconnect);
+
+static void cma_add_one(struct ib_device *device)
+{
+       struct cma_device *cma_dev;
+       struct rdma_id_private *id_priv;
+
+       cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
+       if (!cma_dev)
+               return;
+
+       cma_dev->device = device;
+       cma_dev->node_guid = device->node_guid;
+       if (!cma_dev->node_guid)
+               goto err;
+
+       init_completion(&cma_dev->comp);
+       atomic_set(&cma_dev->refcount, 1);
+       INIT_LIST_HEAD(&cma_dev->id_list);
+       ib_set_client_data(device, &cma_client, cma_dev);
+
+       mutex_lock(&lock);
+       list_add_tail(&cma_dev->list, &dev_list);
+       list_for_each_entry(id_priv, &listen_any_list, list)
+               cma_listen_on_dev(id_priv, cma_dev);
+       mutex_unlock(&lock);
+       return;
+err:
+       kfree(cma_dev);
+}
+
+static int cma_remove_id_dev(struct rdma_id_private *id_priv)
+{
+       enum cma_state state;
+
+       /* Record that we want to remove the device */
+       state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
+       if (state == CMA_DESTROYING)
+               return 0;
+
+       cma_cancel_operation(id_priv, state);
+       wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
+
+       /* Check for destruction from another callback. */
+       if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
+               return 0;
+
+       return cma_notify_user(id_priv, RDMA_CM_EVENT_DEVICE_REMOVAL,
+                              0, NULL, 0);
+}
+
+static void cma_process_remove(struct cma_device *cma_dev)
+{
+       struct list_head remove_list;
+       struct rdma_id_private *id_priv;
+       int ret;
+
+       INIT_LIST_HEAD(&remove_list);
+
+       mutex_lock(&lock);
+       while (!list_empty(&cma_dev->id_list)) {
+               id_priv = list_entry(cma_dev->id_list.next,
+                                    struct rdma_id_private, list);
+
+               if (cma_internal_listen(id_priv)) {
+                       cma_destroy_listen(id_priv);
+                       continue;
+               }
+
+               list_del(&id_priv->list);
+               list_add_tail(&id_priv->list, &remove_list);
+               atomic_inc(&id_priv->refcount);
+               mutex_unlock(&lock);
+
+               ret = cma_remove_id_dev(id_priv);
+               cma_deref_id(id_priv);
+               if (ret)
+                       rdma_destroy_id(&id_priv->id);
+
+               mutex_lock(&lock);
+       }
+       mutex_unlock(&lock);
+
+       cma_deref_dev(cma_dev);
+       wait_for_completion(&cma_dev->comp);
+}
+
+static void cma_remove_one(struct ib_device *device)
+{
+       struct cma_device *cma_dev;
+
+       cma_dev = ib_get_client_data(device, &cma_client);
+       if (!cma_dev)
+               return;
+
+       mutex_lock(&lock);
+       list_del(&cma_dev->list);
+       mutex_unlock(&lock);
+
+       cma_process_remove(cma_dev);
+       kfree(cma_dev);
+}
+
+static int cma_init(void)
+{
+       int ret;
+
+       cma_wq = create_singlethread_workqueue("rdma_cm_wq");
+       if (!cma_wq)
+               return -ENOMEM;
+
+       ret = ib_register_client(&cma_client);
+       if (ret)
+               goto err;
+       return 0;
+
+err:
+       destroy_workqueue(cma_wq);
+       return ret;
+}
+
+static void cma_cleanup(void)
+{
+       ib_unregister_client(&cma_client);
+       destroy_workqueue(cma_wq);
+       idr_destroy(&sdp_ps);
+       idr_destroy(&tcp_ps);
+}
+
+module_init(cma_init);
+module_exit(cma_cleanup);
diff --git a/include/rdma/rdma_cm.h b/include/rdma/rdma_cm.h
new file mode 100644 (file)
index 0000000..402c63d
--- /dev/null
@@ -0,0 +1,256 @@
+/*
+ * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
+ * Copyright (c) 2005 Intel Corporation.  All rights reserved.
+ *
+ * This Software is licensed under one of the following licenses:
+ *
+ * 1) under the terms of the "Common Public License 1.0" a copy of which is
+ *    available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/cpl.php.
+ *
+ * 2) under the terms of the "The BSD License" a copy of which is
+ *    available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/bsd-license.php.
+ *
+ * 3) under the terms of the "GNU General Public License (GPL) Version 2" a
+ *    copy of which is available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/gpl-license.php.
+ *
+ * Licensee has the right to choose one of the above licenses.
+ *
+ * Redistributions of source code must retain the above copyright
+ * notice and one of the license notices.
+ *
+ * Redistributions in binary form must reproduce both the above copyright
+ * notice, one of the license notices in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ */
+
+#if !defined(RDMA_CM_H)
+#define RDMA_CM_H
+
+#include <linux/socket.h>
+#include <linux/in6.h>
+#include <rdma/ib_addr.h>
+#include <rdma/ib_sa.h>
+
+/*
+ * Upon receiving a device removal event, users must destroy the associated
+ * RDMA identifier and release all resources allocated with the device.
+ */
+enum rdma_cm_event_type {
+       RDMA_CM_EVENT_ADDR_RESOLVED,
+       RDMA_CM_EVENT_ADDR_ERROR,
+       RDMA_CM_EVENT_ROUTE_RESOLVED,
+       RDMA_CM_EVENT_ROUTE_ERROR,
+       RDMA_CM_EVENT_CONNECT_REQUEST,
+       RDMA_CM_EVENT_CONNECT_RESPONSE,
+       RDMA_CM_EVENT_CONNECT_ERROR,
+       RDMA_CM_EVENT_UNREACHABLE,
+       RDMA_CM_EVENT_REJECTED,
+       RDMA_CM_EVENT_ESTABLISHED,
+       RDMA_CM_EVENT_DISCONNECTED,
+       RDMA_CM_EVENT_DEVICE_REMOVAL,
+};
+
+enum rdma_port_space {
+       RDMA_PS_SDP  = 0x0001,
+       RDMA_PS_TCP  = 0x0106,
+       RDMA_PS_UDP  = 0x0111,
+       RDMA_PS_SCTP = 0x0183
+};
+
+struct rdma_addr {
+       struct sockaddr src_addr;
+       u8              src_pad[sizeof(struct sockaddr_in6) -
+                               sizeof(struct sockaddr)];
+       struct sockaddr dst_addr;
+       u8              dst_pad[sizeof(struct sockaddr_in6) -
+                               sizeof(struct sockaddr)];
+       struct rdma_dev_addr dev_addr;
+};
+
+struct rdma_route {
+       struct rdma_addr addr;
+       struct ib_sa_path_rec *path_rec;
+       int num_paths;
+};
+
+struct rdma_cm_event {
+       enum rdma_cm_event_type  event;
+       int                      status;
+       void                    *private_data;
+       u8                       private_data_len;
+};
+
+struct rdma_cm_id;
+
+/**
+ * rdma_cm_event_handler - Callback used to report user events.
+ *
+ * Notes: Users may not call rdma_destroy_id from this callback to destroy
+ *   the passed in id, or a corresponding listen id.  Returning a
+ *   non-zero value from the callback will destroy the passed in id.
+ */
+typedef int (*rdma_cm_event_handler)(struct rdma_cm_id *id,
+                                    struct rdma_cm_event *event);
+
+struct rdma_cm_id {
+       struct ib_device        *device;
+       void                    *context;
+       struct ib_qp            *qp;
+       rdma_cm_event_handler    event_handler;
+       struct rdma_route        route;
+       enum rdma_port_space     ps;
+       u8                       port_num;
+};
+
+/**
+ * rdma_create_id - Create an RDMA identifier.
+ *
+ * @event_handler: User callback invoked to report events associated with the
+ *   returned rdma_id.
+ * @context: User specified context associated with the id.
+ * @ps: RDMA port space.
+ */
+struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
+                                 void *context, enum rdma_port_space ps);
+
+void rdma_destroy_id(struct rdma_cm_id *id);
+
+/**
+ * rdma_bind_addr - Bind an RDMA identifier to a source address and
+ *   associated RDMA device, if needed.
+ *
+ * @id: RDMA identifier.
+ * @addr: Local address information.  Wildcard values are permitted.
+ *
+ * This associates a source address with the RDMA identifier before calling
+ * rdma_listen.  If a specific local address is given, the RDMA identifier will
+ * be bound to a local RDMA device.
+ */
+int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr);
+
+/**
+ * rdma_resolve_addr - Resolve destination and optional source addresses
+ *   from IP addresses to an RDMA address.  If successful, the specified
+ *   rdma_cm_id will be bound to a local device.
+ *
+ * @id: RDMA identifier.
+ * @src_addr: Source address information.  This parameter may be NULL.
+ * @dst_addr: Destination address information.
+ * @timeout_ms: Time to wait for resolution to complete.
+ */
+int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
+                     struct sockaddr *dst_addr, int timeout_ms);
+
+/**
+ * rdma_resolve_route - Resolve the RDMA address bound to the RDMA identifier
+ *   into route information needed to establish a connection.
+ *
+ * This is called on the client side of a connection.
+ * Users must have first called rdma_resolve_addr to resolve a dst_addr
+ * into an RDMA address before calling this routine.
+ */
+int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms);
+
+/**
+ * rdma_create_qp - Allocate a QP and associate it with the specified RDMA
+ * identifier.
+ *
+ * QPs allocated to an rdma_cm_id will automatically be transitioned by the CMA
+ * through their states.
+ */
+int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
+                  struct ib_qp_init_attr *qp_init_attr);
+
+/**
+ * rdma_destroy_qp - Deallocate the QP associated with the specified RDMA
+ * identifier.
+ *
+ * Users must destroy any QP associated with an RDMA identifier before
+ * destroying the RDMA ID.
+ */
+void rdma_destroy_qp(struct rdma_cm_id *id);
+
+/**
+ * rdma_init_qp_attr - Initializes the QP attributes for use in transitioning
+ *   to a specified QP state.
+ * @id: Communication identifier associated with the QP attributes to
+ *   initialize.
+ * @qp_attr: On input, specifies the desired QP state.  On output, the
+ *   mandatory and desired optional attributes will be set in order to
+ *   modify the QP to the specified state.
+ * @qp_attr_mask: The QP attribute mask that may be used to transition the
+ *   QP to the specified state.
+ *
+ * Users must set the @qp_attr->qp_state to the desired QP state.  This call
+ * will set all required attributes for the given transition, along with
+ * known optional attributes.  Users may override the attributes returned from
+ * this call before calling ib_modify_qp.
+ *
+ * Users that wish to have their QP automatically transitioned through its
+ * states can associate a QP with the rdma_cm_id by calling rdma_create_qp().
+ */
+int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
+                      int *qp_attr_mask);
+
+struct rdma_conn_param {
+       const void *private_data;
+       u8 private_data_len;
+       u8 responder_resources;
+       u8 initiator_depth;
+       u8 flow_control;
+       u8 retry_count;         /* ignored when accepting */
+       u8 rnr_retry_count;
+       /* Fields below ignored if a QP is created on the rdma_cm_id. */
+       u8 srq;
+       u32 qp_num;
+       enum ib_qp_type qp_type;
+};
+
+/**
+ * rdma_connect - Initiate an active connection request.
+ *
+ * Users must have resolved a route for the rdma_cm_id to connect with
+ * by having called rdma_resolve_route before calling this routine.
+ */
+int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param);
+
+/**
+ * rdma_listen - This function is called by the passive side to
+ *   listen for incoming connection requests.
+ *
+ * Users must have bound the rdma_cm_id to a local address by calling
+ * rdma_bind_addr before calling this routine.
+ */
+int rdma_listen(struct rdma_cm_id *id, int backlog);
+
+/**
+ * rdma_accept - Called to accept a connection request or response.
+ * @id: Connection identifier associated with the request.
+ * @conn_param: Information needed to establish the connection.  This must be
+ *   provided if accepting a connection request.  If accepting a connection
+ *   response, this parameter must be NULL.
+ *
+ * Typically, this routine is only called by the listener to accept a connection
+ * request.  It must also be called on the active side of a connection if the
+ * user is performing their own QP transitions.
+ */
+int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param);
+
+/**
+ * rdma_reject - Called to reject a connection request or response.
+ */
+int rdma_reject(struct rdma_cm_id *id, const void *private_data,
+               u8 private_data_len);
+
+/**
+ * rdma_disconnect - This function disconnects the associated QP and
+ *   transitions it into the error state.
+ */
+int rdma_disconnect(struct rdma_cm_id *id);
+
+#endif /* RDMA_CM_H */
+
diff --git a/include/rdma/rdma_cm_ib.h b/include/rdma/rdma_cm_ib.h
new file mode 100644 (file)
index 0000000..e8c3af1
--- /dev/null
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2006 Intel Corporation.  All rights reserved.
+ *
+ * This Software is licensed under one of the following licenses:
+ *
+ * 1) under the terms of the "Common Public License 1.0" a copy of which is
+ *    available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/cpl.php.
+ *
+ * 2) under the terms of the "The BSD License" a copy of which is
+ *    available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/bsd-license.php.
+ *
+ * 3) under the terms of the "GNU General Public License (GPL) Version 2" a
+ *    copy of which is available from the Open Source Initiative, see
+ *    http://www.opensource.org/licenses/gpl-license.php.
+ *
+ * Licensee has the right to choose one of the above licenses.
+ *
+ * Redistributions of source code must retain the above copyright
+ * notice and one of the license notices.
+ *
+ * Redistributions in binary form must reproduce both the above copyright
+ * notice, one of the license notices in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ */
+
+#if !defined(RDMA_CM_IB_H)
+#define RDMA_CM_IB_H
+
+#include <rdma/rdma_cm.h>
+
+/**
+ * rdma_set_ib_paths - Manually sets the path records used to establish a
+ *   connection.
+ * @id: Connection identifier associated with the request.
+ * @path_rec: Reference to the path record
+ *
+ * This call permits a user to specify routing information for rdma_cm_id's
+ * bound to Infiniband devices.  It is called on the client side of a
+ * connection and replaces the call to rdma_resolve_route.
+ */
+int rdma_set_ib_paths(struct rdma_cm_id *id,
+                     struct ib_sa_path_rec *path_rec, int num_paths);
+
+#endif /* RDMA_CM_IB_H */