include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit...
[safe/jmp/linux-2.6] / drivers / infiniband / core / user_mad.c
index a64d6b4..e7db054 100644 (file)
@@ -1,7 +1,8 @@
 /*
  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
- * Copyright (c) 2005 Voltaire, Inc. All rights reserved. 
+ * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
+ * Copyright (c) 2008 Cisco. All rights reserved.
  *
  * This software is available to you under a choice of one of two
  * licenses.  You may choose to be licensed under the terms of the GNU
@@ -30,8 +31,6 @@
  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
- *
- * $Id: user_mad.c 2814 2005-07-06 19:14:09Z halr $
  */
 
 #include <linux/module.h>
 #include <linux/err.h>
 #include <linux/fs.h>
 #include <linux/cdev.h>
-#include <linux/pci.h>
 #include <linux/dma-mapping.h>
 #include <linux/poll.h>
-#include <linux/rwsem.h>
+#include <linux/mutex.h>
 #include <linux/kref.h>
+#include <linux/compat.h>
+#include <linux/sched.h>
+#include <linux/semaphore.h>
+#include <linux/slab.h>
 
 #include <asm/uaccess.h>
-#include <asm/semaphore.h>
 
 #include <rdma/ib_mad.h>
 #include <rdma/ib_user_mad.h>
@@ -64,18 +65,34 @@ enum {
        IB_UMAD_MINOR_BASE = 0
 };
 
+/*
+ * Our lifetime rules for these structs are the following:
+ * device special file is opened, we take a reference on the
+ * ib_umad_port's struct ib_umad_device. We drop these
+ * references in the corresponding close().
+ *
+ * In addition to references coming from open character devices, there
+ * is one more reference to each ib_umad_device representing the
+ * module's reference taken when allocating the ib_umad_device in
+ * ib_umad_add_one().
+ *
+ * When destroying an ib_umad_device, we drop the module's reference.
+ */
+
 struct ib_umad_port {
-       int                    devnum;
-       struct cdev            dev;
-       struct class_device    class_dev;
+       struct cdev           cdev;
+       struct device         *dev;
 
-       int                    sm_devnum;
-       struct cdev            sm_dev;
-       struct class_device    sm_class_dev;
+       struct cdev           sm_cdev;
+       struct device         *sm_dev;
        struct semaphore       sm_sem;
 
+       struct mutex           file_mutex;
+       struct list_head       file_list;
+
        struct ib_device      *ib_dev;
        struct ib_umad_device *umad_dev;
+       int                    dev_num;
        u8                     port_num;
 };
 
@@ -86,83 +103,104 @@ struct ib_umad_device {
 };
 
 struct ib_umad_file {
-       struct ib_umad_port *port;
-       spinlock_t           recv_lock;
-       struct list_head     recv_list;
-       wait_queue_head_t    recv_wait;
-       struct rw_semaphore  agent_mutex;
-       struct ib_mad_agent *agent[IB_UMAD_MAX_AGENTS];
-       struct ib_mr        *mr[IB_UMAD_MAX_AGENTS];
+       struct mutex            mutex;
+       struct ib_umad_port    *port;
+       struct list_head        recv_list;
+       struct list_head        send_list;
+       struct list_head        port_list;
+       spinlock_t              send_lock;
+       wait_queue_head_t       recv_wait;
+       struct ib_mad_agent    *agent[IB_UMAD_MAX_AGENTS];
+       int                     agents_dead;
+       u8                      use_pkey_index;
+       u8                      already_used;
 };
 
 struct ib_umad_packet {
-       struct ib_ah      *ah;
        struct ib_mad_send_buf *msg;
+       struct ib_mad_recv_wc  *recv_wc;
        struct list_head   list;
        int                length;
-       DECLARE_PCI_UNMAP_ADDR(mapping)
        struct ib_user_mad mad;
 };
 
+static struct class *umad_class;
+
 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
-static spinlock_t map_lock;
-static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS * 2);
+
+static DEFINE_SPINLOCK(port_lock);
+static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS);
 
 static void ib_umad_add_one(struct ib_device *device);
 static void ib_umad_remove_one(struct ib_device *device);
 
+static void ib_umad_release_dev(struct kref *ref)
+{
+       struct ib_umad_device *dev =
+               container_of(ref, struct ib_umad_device, ref);
+
+       kfree(dev);
+}
+
+static int hdr_size(struct ib_umad_file *file)
+{
+       return file->use_pkey_index ? sizeof (struct ib_user_mad_hdr) :
+               sizeof (struct ib_user_mad_hdr_old);
+}
+
+/* caller must hold file->mutex */
+static struct ib_mad_agent *__get_agent(struct ib_umad_file *file, int id)
+{
+       return file->agents_dead ? NULL : file->agent[id];
+}
+
 static int queue_packet(struct ib_umad_file *file,
                        struct ib_mad_agent *agent,
                        struct ib_umad_packet *packet)
 {
        int ret = 1;
 
-       down_read(&file->agent_mutex);
+       mutex_lock(&file->mutex);
+
        for (packet->mad.hdr.id = 0;
             packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
             packet->mad.hdr.id++)
-               if (agent == file->agent[packet->mad.hdr.id]) {
-                       spin_lock_irq(&file->recv_lock);
+               if (agent == __get_agent(file, packet->mad.hdr.id)) {
                        list_add_tail(&packet->list, &file->recv_list);
-                       spin_unlock_irq(&file->recv_lock);
                        wake_up_interruptible(&file->recv_wait);
                        ret = 0;
                        break;
                }
 
-       up_read(&file->agent_mutex);
+       mutex_unlock(&file->mutex);
 
        return ret;
 }
 
+static void dequeue_send(struct ib_umad_file *file,
+                        struct ib_umad_packet *packet)
+{
+       spin_lock_irq(&file->send_lock);
+       list_del(&packet->list);
+       spin_unlock_irq(&file->send_lock);
+}
+
 static void send_handler(struct ib_mad_agent *agent,
                         struct ib_mad_send_wc *send_wc)
 {
        struct ib_umad_file *file = agent->context;
-       struct ib_umad_packet *timeout, *packet =
-               (void *) (unsigned long) send_wc->wr_id;
+       struct ib_umad_packet *packet = send_wc->send_buf->context[0];
 
-       ib_destroy_ah(packet->msg->send_wr.wr.ud.ah);
+       dequeue_send(file, packet);
+       ib_destroy_ah(packet->msg->ah);
        ib_free_send_mad(packet->msg);
 
        if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
-               timeout = kmalloc(sizeof *timeout + sizeof (struct ib_mad_hdr),
-                                 GFP_KERNEL);
-               if (!timeout)
-                       goto out;
-
-               memset(timeout, 0, sizeof *timeout + sizeof (struct ib_mad_hdr));
-
-               timeout->length = sizeof (struct ib_mad_hdr);
-               timeout->mad.hdr.id = packet->mad.hdr.id;
-               timeout->mad.hdr.status = ETIMEDOUT;
-               memcpy(timeout->mad.data, packet->mad.data,
-                      sizeof (struct ib_mad_hdr));
-
-               if (!queue_packet(file, agent, timeout))
-                               return;
+               packet->length = IB_MGMT_MAD_HDR;
+               packet->mad.hdr.status = ETIMEDOUT;
+               if (!queue_packet(file, agent, packet))
+                       return;
        }
-out:
        kfree(packet);
 }
 
@@ -171,44 +209,118 @@ static void recv_handler(struct ib_mad_agent *agent,
 {
        struct ib_umad_file *file = agent->context;
        struct ib_umad_packet *packet;
-       int length;
 
        if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
-               goto out;
+               goto err1;
 
-       length = mad_recv_wc->mad_len;
-       packet = kmalloc(sizeof *packet + length, GFP_KERNEL);
+       packet = kzalloc(sizeof *packet, GFP_KERNEL);
        if (!packet)
-               goto out;
-
-       memset(packet, 0, sizeof *packet + length);
-       packet->length = length;
-
-       ib_coalesce_recv_mad(mad_recv_wc, packet->mad.data);
-
-       packet->mad.hdr.status    = 0;
-       packet->mad.hdr.length    = length + sizeof (struct ib_user_mad);
-       packet->mad.hdr.qpn       = cpu_to_be32(mad_recv_wc->wc->src_qp);
-       packet->mad.hdr.lid       = cpu_to_be16(mad_recv_wc->wc->slid);
-       packet->mad.hdr.sl        = mad_recv_wc->wc->sl;
-       packet->mad.hdr.path_bits = mad_recv_wc->wc->dlid_path_bits;
+               goto err1;
+
+       packet->length = mad_recv_wc->mad_len;
+       packet->recv_wc = mad_recv_wc;
+
+       packet->mad.hdr.status     = 0;
+       packet->mad.hdr.length     = hdr_size(file) + mad_recv_wc->mad_len;
+       packet->mad.hdr.qpn        = cpu_to_be32(mad_recv_wc->wc->src_qp);
+       packet->mad.hdr.lid        = cpu_to_be16(mad_recv_wc->wc->slid);
+       packet->mad.hdr.sl         = mad_recv_wc->wc->sl;
+       packet->mad.hdr.path_bits  = mad_recv_wc->wc->dlid_path_bits;
+       packet->mad.hdr.pkey_index = mad_recv_wc->wc->pkey_index;
        packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
        if (packet->mad.hdr.grh_present) {
-               /* XXX parse GRH */
-               packet->mad.hdr.gid_index       = 0;
-               packet->mad.hdr.hop_limit       = 0;
-               packet->mad.hdr.traffic_class   = 0;
-               memset(packet->mad.hdr.gid, 0, 16);
-               packet->mad.hdr.flow_label      = 0;
+               struct ib_ah_attr ah_attr;
+
+               ib_init_ah_from_wc(agent->device, agent->port_num,
+                                  mad_recv_wc->wc, mad_recv_wc->recv_buf.grh,
+                                  &ah_attr);
+
+               packet->mad.hdr.gid_index = ah_attr.grh.sgid_index;
+               packet->mad.hdr.hop_limit = ah_attr.grh.hop_limit;
+               packet->mad.hdr.traffic_class = ah_attr.grh.traffic_class;
+               memcpy(packet->mad.hdr.gid, &ah_attr.grh.dgid, 16);
+               packet->mad.hdr.flow_label = cpu_to_be32(ah_attr.grh.flow_label);
        }
 
        if (queue_packet(file, agent, packet))
-               kfree(packet);
+               goto err2;
+       return;
 
-out:
+err2:
+       kfree(packet);
+err1:
        ib_free_recv_mad(mad_recv_wc);
 }
 
+static ssize_t copy_recv_mad(struct ib_umad_file *file, char __user *buf,
+                            struct ib_umad_packet *packet, size_t count)
+{
+       struct ib_mad_recv_buf *recv_buf;
+       int left, seg_payload, offset, max_seg_payload;
+
+       /* We need enough room to copy the first (or only) MAD segment. */
+       recv_buf = &packet->recv_wc->recv_buf;
+       if ((packet->length <= sizeof (*recv_buf->mad) &&
+            count < hdr_size(file) + packet->length) ||
+           (packet->length > sizeof (*recv_buf->mad) &&
+            count < hdr_size(file) + sizeof (*recv_buf->mad)))
+               return -EINVAL;
+
+       if (copy_to_user(buf, &packet->mad, hdr_size(file)))
+               return -EFAULT;
+
+       buf += hdr_size(file);
+       seg_payload = min_t(int, packet->length, sizeof (*recv_buf->mad));
+       if (copy_to_user(buf, recv_buf->mad, seg_payload))
+               return -EFAULT;
+
+       if (seg_payload < packet->length) {
+               /*
+                * Multipacket RMPP MAD message. Copy remainder of message.
+                * Note that last segment may have a shorter payload.
+                */
+               if (count < hdr_size(file) + packet->length) {
+                       /*
+                        * The buffer is too small, return the first RMPP segment,
+                        * which includes the RMPP message length.
+                        */
+                       return -ENOSPC;
+               }
+               offset = ib_get_mad_data_offset(recv_buf->mad->mad_hdr.mgmt_class);
+               max_seg_payload = sizeof (struct ib_mad) - offset;
+
+               for (left = packet->length - seg_payload, buf += seg_payload;
+                    left; left -= seg_payload, buf += seg_payload) {
+                       recv_buf = container_of(recv_buf->list.next,
+                                               struct ib_mad_recv_buf, list);
+                       seg_payload = min(left, max_seg_payload);
+                       if (copy_to_user(buf, ((void *) recv_buf->mad) + offset,
+                                        seg_payload))
+                               return -EFAULT;
+               }
+       }
+       return hdr_size(file) + packet->length;
+}
+
+static ssize_t copy_send_mad(struct ib_umad_file *file, char __user *buf,
+                            struct ib_umad_packet *packet, size_t count)
+{
+       ssize_t size = hdr_size(file) + packet->length;
+
+       if (count < size)
+               return -EINVAL;
+
+       if (copy_to_user(buf, &packet->mad, hdr_size(file)))
+               return -EFAULT;
+
+       buf += hdr_size(file);
+
+       if (copy_to_user(buf, packet->mad.data, packet->length))
+               return -EFAULT;
+
+       return size;
+}
+
 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
                            size_t count, loff_t *pos)
 {
@@ -216,13 +328,13 @@ static ssize_t ib_umad_read(struct file *filp, char __user *buf,
        struct ib_umad_packet *packet;
        ssize_t ret;
 
-       if (count < sizeof (struct ib_user_mad) + sizeof (struct ib_mad))
+       if (count < hdr_size(file))
                return -EINVAL;
 
-       spin_lock_irq(&file->recv_lock);
+       mutex_lock(&file->mutex);
 
        while (list_empty(&file->recv_list)) {
-               spin_unlock_irq(&file->recv_lock);
+               mutex_unlock(&file->mutex);
 
                if (filp->f_flags & O_NONBLOCK)
                        return -EAGAIN;
@@ -231,36 +343,97 @@ static ssize_t ib_umad_read(struct file *filp, char __user *buf,
                                             !list_empty(&file->recv_list)))
                        return -ERESTARTSYS;
 
-               spin_lock_irq(&file->recv_lock);
+               mutex_lock(&file->mutex);
        }
 
        packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
        list_del(&packet->list);
 
-       spin_unlock_irq(&file->recv_lock);
+       mutex_unlock(&file->mutex);
 
-       if (count < packet->length + sizeof (struct ib_user_mad)) {
-               /* Return length needed (and first RMPP segment) if too small */
-               if (copy_to_user(buf, &packet->mad,
-                                sizeof (struct ib_user_mad) + sizeof (struct ib_mad)))
-                       ret = -EFAULT;
-               else
-                       ret = -ENOSPC;
-       } else if (copy_to_user(buf, &packet->mad,
-                             packet->length + sizeof (struct ib_user_mad)))
-               ret = -EFAULT;
+       if (packet->recv_wc)
+               ret = copy_recv_mad(file, buf, packet, count);
        else
-               ret = packet->length + sizeof (struct ib_user_mad);
+               ret = copy_send_mad(file, buf, packet, count);
+
        if (ret < 0) {
                /* Requeue packet */
-               spin_lock_irq(&file->recv_lock);
+               mutex_lock(&file->mutex);
                list_add(&packet->list, &file->recv_list);
-               spin_unlock_irq(&file->recv_lock);
-       } else
+               mutex_unlock(&file->mutex);
+       } else {
+               if (packet->recv_wc)
+                       ib_free_recv_mad(packet->recv_wc);
                kfree(packet);
+       }
        return ret;
 }
 
+static int copy_rmpp_mad(struct ib_mad_send_buf *msg, const char __user *buf)
+{
+       int left, seg;
+
+       /* Copy class specific header */
+       if ((msg->hdr_len > IB_MGMT_RMPP_HDR) &&
+           copy_from_user(msg->mad + IB_MGMT_RMPP_HDR, buf + IB_MGMT_RMPP_HDR,
+                          msg->hdr_len - IB_MGMT_RMPP_HDR))
+               return -EFAULT;
+
+       /* All headers are in place.  Copy data segments. */
+       for (seg = 1, left = msg->data_len, buf += msg->hdr_len; left > 0;
+            seg++, left -= msg->seg_size, buf += msg->seg_size) {
+               if (copy_from_user(ib_get_rmpp_segment(msg, seg), buf,
+                                  min(left, msg->seg_size)))
+                       return -EFAULT;
+       }
+       return 0;
+}
+
+static int same_destination(struct ib_user_mad_hdr *hdr1,
+                           struct ib_user_mad_hdr *hdr2)
+{
+       if (!hdr1->grh_present && !hdr2->grh_present)
+          return (hdr1->lid == hdr2->lid);
+
+       if (hdr1->grh_present && hdr2->grh_present)
+          return !memcmp(hdr1->gid, hdr2->gid, 16);
+
+       return 0;
+}
+
+static int is_duplicate(struct ib_umad_file *file,
+                       struct ib_umad_packet *packet)
+{
+       struct ib_umad_packet *sent_packet;
+       struct ib_mad_hdr *sent_hdr, *hdr;
+
+       hdr = (struct ib_mad_hdr *) packet->mad.data;
+       list_for_each_entry(sent_packet, &file->send_list, list) {
+               sent_hdr = (struct ib_mad_hdr *) sent_packet->mad.data;
+
+               if ((hdr->tid != sent_hdr->tid) ||
+                   (hdr->mgmt_class != sent_hdr->mgmt_class))
+                       continue;
+
+               /*
+                * No need to be overly clever here.  If two new operations have
+                * the same TID, reject the second as a duplicate.  This is more
+                * restrictive than required by the spec.
+                */
+               if (!ib_response_mad((struct ib_mad *) hdr)) {
+                       if (!ib_response_mad((struct ib_mad *) sent_hdr))
+                               return 1;
+                       continue;
+               } else if (!ib_response_mad((struct ib_mad *) sent_hdr))
+                       continue;
+
+               if (same_destination(&packet->mad.hdr, &sent_packet->mad.hdr))
+                       return 1;
+       }
+
+       return 0;
+}
+
 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
                             size_t count, loff_t *pos)
 {
@@ -268,26 +441,19 @@ static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
        struct ib_umad_packet *packet;
        struct ib_mad_agent *agent;
        struct ib_ah_attr ah_attr;
-       struct ib_send_wr *bad_wr;
+       struct ib_ah *ah;
        struct ib_rmpp_mad *rmpp_mad;
-       u8 method;
        __be64 *tid;
-       int ret, length, hdr_len, data_len, rmpp_hdr_size;
-       int rmpp_active = 0;
+       int ret, data_len, hdr_len, copy_offset, rmpp_active;
 
-       if (count < sizeof (struct ib_user_mad))
+       if (count < hdr_size(file) + IB_MGMT_RMPP_HDR)
                return -EINVAL;
 
-       length = count - sizeof (struct ib_user_mad);
-       packet = kmalloc(sizeof *packet + sizeof(struct ib_mad_hdr) +
-                        sizeof(struct ib_rmpp_hdr), GFP_KERNEL);
+       packet = kzalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
        if (!packet)
                return -ENOMEM;
 
-       if (copy_from_user(&packet->mad, buf,
-                           sizeof (struct ib_user_mad) +
-                           sizeof(struct ib_mad_hdr) +
-                           sizeof(struct ib_rmpp_hdr))) {
+       if (copy_from_user(&packet->mad, buf, hdr_size(file))) {
                ret = -EFAULT;
                goto err;
        }
@@ -298,11 +464,16 @@ static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
                goto err;
        }
 
-       packet->length = length;
+       buf += hdr_size(file);
 
-       down_read(&file->agent_mutex);
+       if (copy_from_user(packet->mad.data, buf, IB_MGMT_RMPP_HDR)) {
+               ret = -EFAULT;
+               goto err;
+       }
 
-       agent = file->agent[packet->mad.hdr.id];
+       mutex_lock(&file->mutex);
+
+       agent = __get_agent(file, packet->mad.hdr.id);
        if (!agent) {
                ret = -EINVAL;
                goto err_up;
@@ -316,120 +487,97 @@ static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
        if (packet->mad.hdr.grh_present) {
                ah_attr.ah_flags = IB_AH_GRH;
                memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
-               ah_attr.grh.flow_label     = be32_to_cpu(packet->mad.hdr.flow_label);
-               ah_attr.grh.hop_limit      = packet->mad.hdr.hop_limit;
+               ah_attr.grh.sgid_index     = packet->mad.hdr.gid_index;
+               ah_attr.grh.flow_label     = be32_to_cpu(packet->mad.hdr.flow_label);
+               ah_attr.grh.hop_limit      = packet->mad.hdr.hop_limit;
                ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
        }
 
-       packet->ah = ib_create_ah(agent->qp->pd, &ah_attr);
-       if (IS_ERR(packet->ah)) {
-               ret = PTR_ERR(packet->ah);
+       ah = ib_create_ah(agent->qp->pd, &ah_attr);
+       if (IS_ERR(ah)) {
+               ret = PTR_ERR(ah);
                goto err_up;
        }
 
        rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
-       if (ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) & IB_MGMT_RMPP_FLAG_ACTIVE) {
-               /* RMPP active */
-               if (!agent->rmpp_version) {
-                       ret = -EINVAL;
-                       goto err_ah;
-               }
-
-               /* Validate that the management class can support RMPP */
-               if (rmpp_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_ADM) {
-                       hdr_len = offsetof(struct ib_sa_mad, data);
-                       data_len = length - hdr_len;
-               } else if ((rmpp_mad->mad_hdr.mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
-                           (rmpp_mad->mad_hdr.mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)) {
-                               hdr_len = offsetof(struct ib_vendor_mad, data);
-                               data_len = length - hdr_len;
-               } else {
-                       ret = -EINVAL;
-                       goto err_ah;
-               }
-               rmpp_active = 1;
+       hdr_len = ib_get_mad_data_offset(rmpp_mad->mad_hdr.mgmt_class);
+       if (!ib_is_mad_class_rmpp(rmpp_mad->mad_hdr.mgmt_class)) {
+               copy_offset = IB_MGMT_MAD_HDR;
+               rmpp_active = 0;
        } else {
-               if (length > sizeof(struct ib_mad)) {
-                       ret = -EINVAL;
-                       goto err_ah;
-               }
-               hdr_len = offsetof(struct ib_mad, data);
-               data_len = length - hdr_len;
+               copy_offset = IB_MGMT_RMPP_HDR;
+               rmpp_active = ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
+                             IB_MGMT_RMPP_FLAG_ACTIVE;
        }
 
+       data_len = count - hdr_size(file) - hdr_len;
        packet->msg = ib_create_send_mad(agent,
                                         be32_to_cpu(packet->mad.hdr.qpn),
-                                        0, packet->ah, rmpp_active,
-                                        hdr_len, data_len,
-                                        GFP_KERNEL);
+                                        packet->mad.hdr.pkey_index, rmpp_active,
+                                        hdr_len, data_len, GFP_KERNEL);
        if (IS_ERR(packet->msg)) {
                ret = PTR_ERR(packet->msg);
                goto err_ah;
        }
 
-       packet->msg->send_wr.wr.ud.timeout_ms  = packet->mad.hdr.timeout_ms;
-       packet->msg->send_wr.wr.ud.retries = packet->mad.hdr.retries;
+       packet->msg->ah         = ah;
+       packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
+       packet->msg->retries    = packet->mad.hdr.retries;
+       packet->msg->context[0] = packet;
 
-       /* Override send WR WRID initialized in ib_create_send_mad */
-       packet->msg->send_wr.wr_id = (unsigned long) packet;
+       /* Copy MAD header.  Any RMPP header is already in place. */
+       memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
 
        if (!rmpp_active) {
-               /* Copy message from user into send buffer */
-               if (copy_from_user(packet->msg->mad,
-                                  buf + sizeof(struct ib_user_mad), length)) {
+               if (copy_from_user(packet->msg->mad + copy_offset,
+                                  buf + copy_offset,
+                                  hdr_len + data_len - copy_offset)) {
                        ret = -EFAULT;
                        goto err_msg;
                }
        } else {
-               rmpp_hdr_size = sizeof(struct ib_mad_hdr) +
-                               sizeof(struct ib_rmpp_hdr);
-
-               /* Only copy MAD headers (RMPP header in place) */
-               memcpy(packet->msg->mad, packet->mad.data,
-                      sizeof(struct ib_mad_hdr));
-
-               /* Now, copy rest of message from user into send buffer */
-               if (copy_from_user(((struct ib_rmpp_mad *) packet->msg->mad)->data,
-                                  buf + sizeof (struct ib_user_mad) + rmpp_hdr_size,
-                                  length - rmpp_hdr_size)) {
-                       ret = -EFAULT;
+               ret = copy_rmpp_mad(packet->msg, buf);
+               if (ret)
                        goto err_msg;
-               }
        }
 
        /*
-        * If userspace is generating a request that will generate a
-        * response, we need to make sure the high-order part of the
-        * transaction ID matches the agent being used to send the
-        * MAD.
+        * Set the high-order part of the transaction ID to make MADs from
+        * different agents unique, and allow routing responses back to the
+        * original requestor.
         */
-       method = packet->msg->mad->mad_hdr.method;
-
-       if (!(method & IB_MGMT_METHOD_RESP)       &&
-           method != IB_MGMT_METHOD_TRAP_REPRESS &&
-           method != IB_MGMT_METHOD_SEND) {
-               tid = &packet->msg->mad->mad_hdr.tid;
+       if (!ib_response_mad(packet->msg->mad)) {
+               tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
                *tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
                                   (be64_to_cpup(tid) & 0xffffffff));
+               rmpp_mad->mad_hdr.tid = *tid;
        }
 
-       ret = ib_post_send_mad(agent, &packet->msg->send_wr, &bad_wr);
-       if (ret)
+       spin_lock_irq(&file->send_lock);
+       ret = is_duplicate(file, packet);
+       if (!ret)
+               list_add_tail(&packet->list, &file->send_list);
+       spin_unlock_irq(&file->send_lock);
+       if (ret) {
+               ret = -EINVAL;
                goto err_msg;
+       }
 
-       up_read(&file->agent_mutex);
+       ret = ib_post_send_mad(packet->msg, NULL);
+       if (ret)
+               goto err_send;
 
-       return sizeof (struct ib_user_mad_hdr) + packet->length;
+       mutex_unlock(&file->mutex);
+       return count;
 
+err_send:
+       dequeue_send(file, packet);
 err_msg:
        ib_free_send_mad(packet->msg);
-
 err_ah:
-       ib_destroy_ah(packet->ah);
-
+       ib_destroy_ah(ah);
 err_up:
-       up_read(&file->agent_mutex);
-
+       mutex_unlock(&file->mutex);
 err:
        kfree(packet);
        return ret;
@@ -450,17 +598,24 @@ static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wa
        return mask;
 }
 
-static int ib_umad_reg_agent(struct ib_umad_file *file, unsigned long arg)
+static int ib_umad_reg_agent(struct ib_umad_file *file, void __user *arg,
+                            int compat_method_mask)
 {
        struct ib_user_mad_reg_req ureq;
        struct ib_mad_reg_req req;
-       struct ib_mad_agent *agent;
+       struct ib_mad_agent *agent = NULL;
        int agent_id;
        int ret;
 
-       down_write(&file->agent_mutex);
+       mutex_lock(&file->port->file_mutex);
+       mutex_lock(&file->mutex);
 
-       if (copy_from_user(&ureq, (void __user *) arg, sizeof ureq)) {
+       if (!file->port->ib_dev) {
+               ret = -EPIPE;
+               goto out;
+       }
+
+       if (copy_from_user(&ureq, arg, sizeof ureq)) {
                ret = -EFAULT;
                goto out;
        }
@@ -471,7 +626,7 @@ static int ib_umad_reg_agent(struct ib_umad_file *file, unsigned long arg)
        }
 
        for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
-               if (!file->agent[agent_id])
+               if (!__get_agent(file, agent_id))
                        goto found;
 
        ret = -ENOMEM;
@@ -481,8 +636,18 @@ found:
        if (ureq.mgmt_class) {
                req.mgmt_class         = ureq.mgmt_class;
                req.mgmt_class_version = ureq.mgmt_class_version;
-               memcpy(req.method_mask, ureq.method_mask, sizeof req.method_mask);
-               memcpy(req.oui,         ureq.oui,         sizeof req.oui);
+               memcpy(req.oui, ureq.oui, sizeof req.oui);
+
+               if (compat_method_mask) {
+                       u32 *umm = (u32 *) ureq.method_mask;
+                       int i;
+
+                       for (i = 0; i < BITS_TO_LONGS(IB_MGMT_MAX_METHODS); ++i)
+                               req.method_mask[i] =
+                                       umm[i * 2] | ((u64) umm[i * 2 + 1] << 32);
+               } else
+                       memcpy(req.method_mask, ureq.method_mask,
+                              sizeof req.method_mask);
        }
 
        agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
@@ -492,61 +657,82 @@ found:
                                      send_handler, recv_handler, file);
        if (IS_ERR(agent)) {
                ret = PTR_ERR(agent);
+               agent = NULL;
                goto out;
        }
 
-       file->agent[agent_id] = agent;
-
-       file->mr[agent_id] = ib_get_dma_mr(agent->qp->pd, IB_ACCESS_LOCAL_WRITE);
-       if (IS_ERR(file->mr[agent_id])) {
-               ret = -ENOMEM;
-               goto err;
-       }
-
        if (put_user(agent_id,
                     (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
                ret = -EFAULT;
-               goto err_mr;
+               goto out;
        }
 
+       if (!file->already_used) {
+               file->already_used = 1;
+               if (!file->use_pkey_index) {
+                       printk(KERN_WARNING "user_mad: process %s did not enable "
+                              "P_Key index support.\n", current->comm);
+                       printk(KERN_WARNING "user_mad:   Documentation/infiniband/user_mad.txt "
+                              "has info on the new ABI.\n");
+               }
+       }
+
+       file->agent[agent_id] = agent;
        ret = 0;
-       goto out;
 
-err_mr:
-       ib_dereg_mr(file->mr[agent_id]);
+out:
+       mutex_unlock(&file->mutex);
 
-err:
-       file->agent[agent_id] = NULL;
-       ib_unregister_mad_agent(agent);
+       if (ret && agent)
+               ib_unregister_mad_agent(agent);
+
+       mutex_unlock(&file->port->file_mutex);
 
-out:
-       up_write(&file->agent_mutex);
        return ret;
 }
 
-static int ib_umad_unreg_agent(struct ib_umad_file *file, unsigned long arg)
+static int ib_umad_unreg_agent(struct ib_umad_file *file, u32 __user *arg)
 {
+       struct ib_mad_agent *agent = NULL;
        u32 id;
        int ret = 0;
 
-       down_write(&file->agent_mutex);
+       if (get_user(id, arg))
+               return -EFAULT;
 
-       if (get_user(id, (u32 __user *) arg)) {
-               ret = -EFAULT;
-               goto out;
-       }
+       mutex_lock(&file->port->file_mutex);
+       mutex_lock(&file->mutex);
 
-       if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !file->agent[id]) {
+       if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !__get_agent(file, id)) {
                ret = -EINVAL;
                goto out;
        }
 
-       ib_dereg_mr(file->mr[id]);
-       ib_unregister_mad_agent(file->agent[id]);
+       agent = file->agent[id];
        file->agent[id] = NULL;
 
 out:
-       up_write(&file->agent_mutex);
+       mutex_unlock(&file->mutex);
+
+       if (agent)
+               ib_unregister_mad_agent(agent);
+
+       mutex_unlock(&file->port->file_mutex);
+
+       return ret;
+}
+
+static long ib_umad_enable_pkey(struct ib_umad_file *file)
+{
+       int ret = 0;
+
+       mutex_lock(&file->mutex);
+       if (file->already_used)
+               ret = -EINVAL;
+       else
+               file->use_pkey_index = 1;
+       mutex_unlock(&file->mutex);
+
        return ret;
 }
 
@@ -555,94 +741,173 @@ static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
 {
        switch (cmd) {
        case IB_USER_MAD_REGISTER_AGENT:
-               return ib_umad_reg_agent(filp->private_data, arg);
+               return ib_umad_reg_agent(filp->private_data, (void __user *) arg, 0);
        case IB_USER_MAD_UNREGISTER_AGENT:
-               return ib_umad_unreg_agent(filp->private_data, arg);
+               return ib_umad_unreg_agent(filp->private_data, (__u32 __user *) arg);
+       case IB_USER_MAD_ENABLE_PKEY:
+               return ib_umad_enable_pkey(filp->private_data);
        default:
                return -ENOIOCTLCMD;
        }
 }
 
+#ifdef CONFIG_COMPAT
+static long ib_umad_compat_ioctl(struct file *filp, unsigned int cmd,
+                                unsigned long arg)
+{
+       switch (cmd) {
+       case IB_USER_MAD_REGISTER_AGENT:
+               return ib_umad_reg_agent(filp->private_data, compat_ptr(arg), 1);
+       case IB_USER_MAD_UNREGISTER_AGENT:
+               return ib_umad_unreg_agent(filp->private_data, compat_ptr(arg));
+       case IB_USER_MAD_ENABLE_PKEY:
+               return ib_umad_enable_pkey(filp->private_data);
+       default:
+               return -ENOIOCTLCMD;
+       }
+}
+#endif
+
+/*
+ * ib_umad_open() does not need the BKL:
+ *
+ *  - the ib_umad_port structures are properly reference counted, and
+ *    everything else is purely local to the file being created, so
+ *    races against other open calls are not a problem;
+ *  - the ioctl method does not affect any global state outside of the
+ *    file structure being operated on;
+ */
 static int ib_umad_open(struct inode *inode, struct file *filp)
 {
-       struct ib_umad_port *port =
-               container_of(inode->i_cdev, struct ib_umad_port, dev);
+       struct ib_umad_port *port;
        struct ib_umad_file *file;
+       int ret = 0;
 
-       file = kmalloc(sizeof *file, GFP_KERNEL);
-       if (!file)
-               return -ENOMEM;
+       port = container_of(inode->i_cdev, struct ib_umad_port, cdev);
+       if (port)
+               kref_get(&port->umad_dev->ref);
+       else
+               return -ENXIO;
 
-       memset(file, 0, sizeof *file);
+       mutex_lock(&port->file_mutex);
 
-       spin_lock_init(&file->recv_lock);
-       init_rwsem(&file->agent_mutex);
+       if (!port->ib_dev) {
+               ret = -ENXIO;
+               goto out;
+       }
+
+       file = kzalloc(sizeof *file, GFP_KERNEL);
+       if (!file) {
+               kref_put(&port->umad_dev->ref, ib_umad_release_dev);
+               ret = -ENOMEM;
+               goto out;
+       }
+
+       mutex_init(&file->mutex);
+       spin_lock_init(&file->send_lock);
        INIT_LIST_HEAD(&file->recv_list);
+       INIT_LIST_HEAD(&file->send_list);
        init_waitqueue_head(&file->recv_wait);
 
        file->port = port;
        filp->private_data = file;
 
-       return 0;
+       list_add_tail(&file->port_list, &port->file_list);
+
+out:
+       mutex_unlock(&port->file_mutex);
+       return ret;
 }
 
 static int ib_umad_close(struct inode *inode, struct file *filp)
 {
        struct ib_umad_file *file = filp->private_data;
+       struct ib_umad_device *dev = file->port->umad_dev;
        struct ib_umad_packet *packet, *tmp;
+       int already_dead;
        int i;
 
-       for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
-               if (file->agent[i]) {
-                       ib_dereg_mr(file->mr[i]);
-                       ib_unregister_mad_agent(file->agent[i]);
-               }
+       mutex_lock(&file->port->file_mutex);
+       mutex_lock(&file->mutex);
 
-       list_for_each_entry_safe(packet, tmp, &file->recv_list, list)
+       already_dead = file->agents_dead;
+       file->agents_dead = 1;
+
+       list_for_each_entry_safe(packet, tmp, &file->recv_list, list) {
+               if (packet->recv_wc)
+                       ib_free_recv_mad(packet->recv_wc);
                kfree(packet);
+       }
+
+       list_del(&file->port_list);
+
+       mutex_unlock(&file->mutex);
+
+       if (!already_dead)
+               for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
+                       if (file->agent[i])
+                               ib_unregister_mad_agent(file->agent[i]);
+
+       mutex_unlock(&file->port->file_mutex);
 
        kfree(file);
+       kref_put(&dev->ref, ib_umad_release_dev);
 
        return 0;
 }
 
-static struct file_operations umad_fops = {
-       .owner          = THIS_MODULE,
-       .read           = ib_umad_read,
-       .write          = ib_umad_write,
-       .poll           = ib_umad_poll,
+static const struct file_operations umad_fops = {
+       .owner          = THIS_MODULE,
+       .read           = ib_umad_read,
+       .write          = ib_umad_write,
+       .poll           = ib_umad_poll,
        .unlocked_ioctl = ib_umad_ioctl,
-       .compat_ioctl   = ib_umad_ioctl,
-       .open           = ib_umad_open,
-       .release        = ib_umad_close
+#ifdef CONFIG_COMPAT
+       .compat_ioctl   = ib_umad_compat_ioctl,
+#endif
+       .open           = ib_umad_open,
+       .release        = ib_umad_close
 };
 
 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
 {
-       struct ib_umad_port *port =
-               container_of(inode->i_cdev, struct ib_umad_port, sm_dev);
+       struct ib_umad_port *port;
        struct ib_port_modify props = {
                .set_port_cap_mask = IB_PORT_SM
        };
        int ret;
 
+       port = container_of(inode->i_cdev, struct ib_umad_port, sm_cdev);
+       if (port)
+               kref_get(&port->umad_dev->ref);
+       else
+               return -ENXIO;
+
        if (filp->f_flags & O_NONBLOCK) {
-               if (down_trylock(&port->sm_sem))
-                       return -EAGAIN;
+               if (down_trylock(&port->sm_sem)) {
+                       ret = -EAGAIN;
+                       goto fail;
+               }
        } else {
-               if (down_interruptible(&port->sm_sem))
-                       return -ERESTARTSYS;
+               if (down_interruptible(&port->sm_sem)) {
+                       ret = -ERESTARTSYS;
+                       goto fail;
+               }
        }
 
        ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
        if (ret) {
                up(&port->sm_sem);
-               return ret;
+               goto fail;
        }
 
        filp->private_data = port;
 
        return 0;
+
+fail:
+       kref_put(&port->umad_dev->ref, ib_umad_release_dev);
+       return ret;
 }
 
 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
@@ -651,17 +916,23 @@ static int ib_umad_sm_close(struct inode *inode, struct file *filp)
        struct ib_port_modify props = {
                .clr_port_cap_mask = IB_PORT_SM
        };
-       int ret;
+       int ret = 0;
+
+       mutex_lock(&port->file_mutex);
+       if (port->ib_dev)
+               ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
+       mutex_unlock(&port->file_mutex);
 
-       ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
        up(&port->sm_sem);
 
+       kref_put(&port->umad_dev->ref, ib_umad_release_dev);
+
        return ret;
 }
 
-static struct file_operations umad_sm_fops = {
-       .owner   = THIS_MODULE,
-       .open    = ib_umad_sm_open,
+static const struct file_operations umad_sm_fops = {
+       .owner   = THIS_MODULE,
+       .open    = ib_umad_sm_open,
        .release = ib_umad_sm_close
 };
 
@@ -671,176 +942,201 @@ static struct ib_client umad_client = {
        .remove = ib_umad_remove_one
 };
 
-static ssize_t show_dev(struct class_device *class_dev, char *buf)
+static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
+                         char *buf)
 {
-       struct ib_umad_port *port = class_get_devdata(class_dev);
-
-       if (class_dev == &port->class_dev)
-               return print_dev_t(buf, port->dev.dev);
-       else
-               return print_dev_t(buf, port->sm_dev.dev);
-}
-static CLASS_DEVICE_ATTR(dev, S_IRUGO, show_dev, NULL);
+       struct ib_umad_port *port = dev_get_drvdata(dev);
 
-static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
-{
-       struct ib_umad_port *port = class_get_devdata(class_dev);
+       if (!port)
+               return -ENODEV;
 
        return sprintf(buf, "%s\n", port->ib_dev->name);
 }
-static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
+static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 
-static ssize_t show_port(struct class_device *class_dev, char *buf)
+static ssize_t show_port(struct device *dev, struct device_attribute *attr,
+                        char *buf)
 {
-       struct ib_umad_port *port = class_get_devdata(class_dev);
+       struct ib_umad_port *port = dev_get_drvdata(dev);
+
+       if (!port)
+               return -ENODEV;
 
        return sprintf(buf, "%d\n", port->port_num);
 }
-static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
+static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
 
-static void ib_umad_release_dev(struct kref *ref)
-{
-       struct ib_umad_device *dev =
-               container_of(ref, struct ib_umad_device, ref);
+static CLASS_ATTR_STRING(abi_version, S_IRUGO,
+                        __stringify(IB_USER_MAD_ABI_VERSION));
 
-       kfree(dev);
-}
-
-static void ib_umad_release_port(struct class_device *class_dev)
+static dev_t overflow_maj;
+static DECLARE_BITMAP(overflow_map, IB_UMAD_MAX_PORTS);
+static int find_overflow_devnum(void)
 {
-       struct ib_umad_port *port = class_get_devdata(class_dev);
+       int ret;
 
-       if (class_dev == &port->class_dev) {
-               cdev_del(&port->dev);
-               clear_bit(port->devnum, dev_map);
-       } else {
-               cdev_del(&port->sm_dev);
-               clear_bit(port->sm_devnum, dev_map);
+       if (!overflow_maj) {
+               ret = alloc_chrdev_region(&overflow_maj, 0, IB_UMAD_MAX_PORTS * 2,
+                                         "infiniband_mad");
+               if (ret) {
+                       printk(KERN_ERR "user_mad: couldn't register dynamic device number\n");
+                       return ret;
+               }
        }
 
-       kref_put(&port->umad_dev->ref, ib_umad_release_dev);
-}
-
-static struct class umad_class = {
-       .name    = "infiniband_mad",
-       .release = ib_umad_release_port
-};
+       ret = find_first_zero_bit(overflow_map, IB_UMAD_MAX_PORTS);
+       if (ret >= IB_UMAD_MAX_PORTS)
+               return -1;
 
-static ssize_t show_abi_version(struct class *class, char *buf)
-{
-       return sprintf(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
+       return ret;
 }
-static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 
 static int ib_umad_init_port(struct ib_device *device, int port_num,
                             struct ib_umad_port *port)
 {
-       spin_lock(&map_lock);
-       port->devnum = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
-       if (port->devnum >= IB_UMAD_MAX_PORTS) {
-               spin_unlock(&map_lock);
-               return -1;
-       }
-       port->sm_devnum = find_next_zero_bit(dev_map, IB_UMAD_MAX_PORTS * 2, IB_UMAD_MAX_PORTS);
-       if (port->sm_devnum >= IB_UMAD_MAX_PORTS * 2) {
-               spin_unlock(&map_lock);
-               return -1;
+       int devnum;
+       dev_t base;
+
+       spin_lock(&port_lock);
+       devnum = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
+       if (devnum >= IB_UMAD_MAX_PORTS) {
+               spin_unlock(&port_lock);
+               devnum = find_overflow_devnum();
+               if (devnum < 0)
+                       return -1;
+
+               spin_lock(&port_lock);
+               port->dev_num = devnum + IB_UMAD_MAX_PORTS;
+               base = devnum + overflow_maj;
+               set_bit(devnum, overflow_map);
+       } else {
+               port->dev_num = devnum;
+               base = devnum + base_dev;
+               set_bit(devnum, dev_map);
        }
-       set_bit(port->devnum, dev_map);
-       set_bit(port->sm_devnum, dev_map);
-       spin_unlock(&map_lock);
+       spin_unlock(&port_lock);
 
        port->ib_dev   = device;
        port->port_num = port_num;
        init_MUTEX(&port->sm_sem);
+       mutex_init(&port->file_mutex);
+       INIT_LIST_HEAD(&port->file_list);
 
-       cdev_init(&port->dev, &umad_fops);
-       port->dev.owner = THIS_MODULE;
-       kobject_set_name(&port->dev.kobj, "umad%d", port->devnum);
-       if (cdev_add(&port->dev, base_dev + port->devnum, 1))
-               return -1;
-
-       port->class_dev.class = &umad_class;
-       port->class_dev.dev   = device->dma_device;
-
-       snprintf(port->class_dev.class_id, BUS_ID_SIZE, "umad%d", port->devnum);
-
-       if (class_device_register(&port->class_dev))
+       cdev_init(&port->cdev, &umad_fops);
+       port->cdev.owner = THIS_MODULE;
+       kobject_set_name(&port->cdev.kobj, "umad%d", port->dev_num);
+       if (cdev_add(&port->cdev, base, 1))
                goto err_cdev;
 
-       class_set_devdata(&port->class_dev, port);
-       kref_get(&port->umad_dev->ref);
-
-       if (class_device_create_file(&port->class_dev, &class_device_attr_dev))
-               goto err_class;
-       if (class_device_create_file(&port->class_dev, &class_device_attr_ibdev))
-               goto err_class;
-       if (class_device_create_file(&port->class_dev, &class_device_attr_port))
-               goto err_class;
-
-       cdev_init(&port->sm_dev, &umad_sm_fops);
-       port->sm_dev.owner = THIS_MODULE;
-       kobject_set_name(&port->dev.kobj, "issm%d", port->sm_devnum - IB_UMAD_MAX_PORTS);
-       if (cdev_add(&port->sm_dev, base_dev + port->sm_devnum, 1))
-               return -1;
-
-       port->sm_class_dev.class = &umad_class;
-       port->sm_class_dev.dev   = device->dma_device;
+       port->dev = device_create(umad_class, device->dma_device,
+                                 port->cdev.dev, port,
+                                 "umad%d", port->dev_num);
+       if (IS_ERR(port->dev))
+               goto err_cdev;
 
-       snprintf(port->sm_class_dev.class_id, BUS_ID_SIZE, "issm%d", port->sm_devnum - IB_UMAD_MAX_PORTS);
+       if (device_create_file(port->dev, &dev_attr_ibdev))
+               goto err_dev;
+       if (device_create_file(port->dev, &dev_attr_port))
+               goto err_dev;
 
-       if (class_device_register(&port->sm_class_dev))
+       base += IB_UMAD_MAX_PORTS;
+       cdev_init(&port->sm_cdev, &umad_sm_fops);
+       port->sm_cdev.owner = THIS_MODULE;
+       kobject_set_name(&port->sm_cdev.kobj, "issm%d", port->dev_num);
+       if (cdev_add(&port->sm_cdev, base, 1))
                goto err_sm_cdev;
 
-       class_set_devdata(&port->sm_class_dev, port);
-       kref_get(&port->umad_dev->ref);
+       port->sm_dev = device_create(umad_class, device->dma_device,
+                                    port->sm_cdev.dev, port,
+                                    "issm%d", port->dev_num);
+       if (IS_ERR(port->sm_dev))
+               goto err_sm_cdev;
 
-       if (class_device_create_file(&port->sm_class_dev, &class_device_attr_dev))
-               goto err_sm_class;
-       if (class_device_create_file(&port->sm_class_dev, &class_device_attr_ibdev))
-               goto err_sm_class;
-       if (class_device_create_file(&port->sm_class_dev, &class_device_attr_port))
-               goto err_sm_class;
+       if (device_create_file(port->sm_dev, &dev_attr_ibdev))
+               goto err_sm_dev;
+       if (device_create_file(port->sm_dev, &dev_attr_port))
+               goto err_sm_dev;
 
        return 0;
 
-err_sm_class:
-       class_device_unregister(&port->sm_class_dev);
+err_sm_dev:
+       device_destroy(umad_class, port->sm_cdev.dev);
 
 err_sm_cdev:
-       cdev_del(&port->sm_dev);
+       cdev_del(&port->sm_cdev);
 
-err_class:
-       class_device_unregister(&port->class_dev);
+err_dev:
+       device_destroy(umad_class, port->cdev.dev);
 
 err_cdev:
-       cdev_del(&port->dev);
-       clear_bit(port->devnum, dev_map);
+       cdev_del(&port->cdev);
+       if (port->dev_num < IB_UMAD_MAX_PORTS)
+               clear_bit(devnum, dev_map);
+       else
+               clear_bit(devnum, overflow_map);
 
        return -1;
 }
 
+static void ib_umad_kill_port(struct ib_umad_port *port)
+{
+       struct ib_umad_file *file;
+       int already_dead;
+       int id;
+
+       dev_set_drvdata(port->dev,    NULL);
+       dev_set_drvdata(port->sm_dev, NULL);
+
+       device_destroy(umad_class, port->cdev.dev);
+       device_destroy(umad_class, port->sm_cdev.dev);
+
+       cdev_del(&port->cdev);
+       cdev_del(&port->sm_cdev);
+
+       mutex_lock(&port->file_mutex);
+
+       port->ib_dev = NULL;
+
+       list_for_each_entry(file, &port->file_list, port_list) {
+               mutex_lock(&file->mutex);
+               already_dead = file->agents_dead;
+               file->agents_dead = 1;
+               mutex_unlock(&file->mutex);
+
+               for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
+                       if (file->agent[id])
+                               ib_unregister_mad_agent(file->agent[id]);
+       }
+
+       mutex_unlock(&port->file_mutex);
+
+       if (port->dev_num < IB_UMAD_MAX_PORTS)
+               clear_bit(port->dev_num, dev_map);
+       else
+               clear_bit(port->dev_num - IB_UMAD_MAX_PORTS, overflow_map);
+}
+
 static void ib_umad_add_one(struct ib_device *device)
 {
        struct ib_umad_device *umad_dev;
        int s, e, i;
 
-       if (device->node_type == IB_NODE_SWITCH)
+       if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
+               return;
+
+       if (device->node_type == RDMA_NODE_IB_SWITCH)
                s = e = 0;
        else {
                s = 1;
                e = device->phys_port_cnt;
        }
 
-       umad_dev = kmalloc(sizeof *umad_dev +
+       umad_dev = kzalloc(sizeof *umad_dev +
                           (e - s + 1) * sizeof (struct ib_umad_port),
                           GFP_KERNEL);
        if (!umad_dev)
                return;
 
-       memset(umad_dev, 0, sizeof *umad_dev +
-              (e - s + 1) * sizeof (struct ib_umad_port));
-
        kref_init(&umad_dev->ref);
 
        umad_dev->start_port = s;
@@ -858,10 +1154,8 @@ static void ib_umad_add_one(struct ib_device *device)
        return;
 
 err:
-       while (--i >= s) {
-               class_device_unregister(&umad_dev->port[i - s].class_dev);
-               class_device_unregister(&umad_dev->port[i - s].sm_class_dev);
-       }
+       while (--i >= s)
+               ib_umad_kill_port(&umad_dev->port[i - s]);
 
        kref_put(&umad_dev->ref, ib_umad_release_dev);
 }
@@ -874,10 +1168,8 @@ static void ib_umad_remove_one(struct ib_device *device)
        if (!umad_dev)
                return;
 
-       for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i) {
-               class_device_unregister(&umad_dev->port[i].class_dev);
-               class_device_unregister(&umad_dev->port[i].sm_class_dev);
-       }
+       for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i)
+               ib_umad_kill_port(&umad_dev->port[i]);
 
        kref_put(&umad_dev->ref, ib_umad_release_dev);
 }
@@ -886,8 +1178,6 @@ static int __init ib_umad_init(void)
 {
        int ret;
 
-       spin_lock_init(&map_lock);
-
        ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
                                     "infiniband_mad");
        if (ret) {
@@ -895,13 +1185,14 @@ static int __init ib_umad_init(void)
                goto out;
        }
 
-       ret = class_register(&umad_class);
-       if (ret) {
+       umad_class = class_create(THIS_MODULE, "infiniband_mad");
+       if (IS_ERR(umad_class)) {
+               ret = PTR_ERR(umad_class);
                printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
                goto out_chrdev;
        }
 
-       ret = class_create_file(&umad_class, &class_attr_abi_version);
+       ret = class_create_file(umad_class, &class_attr_abi_version.attr);
        if (ret) {
                printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
                goto out_class;
@@ -916,7 +1207,7 @@ static int __init ib_umad_init(void)
        return 0;
 
 out_class:
-       class_unregister(&umad_class);
+       class_destroy(umad_class);
 
 out_chrdev:
        unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
@@ -928,8 +1219,10 @@ out:
 static void __exit ib_umad_cleanup(void)
 {
        ib_unregister_client(&umad_client);
-       class_unregister(&umad_class);
+       class_destroy(umad_class);
        unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
+       if (overflow_maj)
+               unregister_chrdev_region(overflow_maj, IB_UMAD_MAX_PORTS * 2);
 }
 
 module_init(ib_umad_init);