Michael> I think I see a solution: replace up_write with
    Michael> downgrade_write.  This way ib_umad_close has a read lock
    Michael> most of the time, and write lock only while it is
    Michael> changing the list.

Yes, excellent idea.  I wasn't familiar with that API, but that's
almost exactly what we need.  It's still a little ugly but I think
this works:

--- infiniband/core/user_mad.c  (revision 4008)
+++ infiniband/core/user_mad.c  (working copy)
@@ -110,12 +110,13 @@ struct ib_umad_device {
 };
 
 struct ib_umad_file {
-       struct ib_umad_port *port;
-       struct list_head     recv_list;
-       struct list_head     port_list;
-       spinlock_t           recv_lock;
-       wait_queue_head_t    recv_wait;
-       struct ib_mad_agent *agent[IB_UMAD_MAX_AGENTS];
+       struct ib_umad_port    *port;
+       struct list_head        recv_list;
+       struct list_head        port_list;
+       spinlock_t              recv_lock;
+       wait_queue_head_t       recv_wait;
+       struct ib_mad_agent    *agent[IB_UMAD_MAX_AGENTS];
+       int                     agents_dead;
 };
 
 struct ib_umad_packet {
@@ -144,6 +145,12 @@ static void ib_umad_release_dev(struct k
        kfree(dev);
 }
 
+/* caller must hold port->mutex at least for reading */
+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)
@@ -151,10 +158,11 @@ static int queue_packet(struct ib_umad_f
        int ret = 1;
 
        down_read(&file->port->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]) {
+               if (agent == __get_agent(file, packet->mad.hdr.id)) {
                        spin_lock_irq(&file->recv_lock);
                        list_add_tail(&packet->list, &file->recv_list);
                        spin_unlock_irq(&file->recv_lock);
@@ -326,7 +334,7 @@ static ssize_t ib_umad_write(struct file
 
        down_read(&file->port->mutex);
 
-       agent = file->agent[packet->mad.hdr.id];
+       agent = __get_agent(file, packet->mad.hdr.id);
        if (!agent) {
                ret = -EINVAL;
                goto err_up;
@@ -480,7 +488,7 @@ static int ib_umad_reg_agent(struct ib_u
        }
 
        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;
@@ -530,7 +538,7 @@ static int ib_umad_unreg_agent(struct ib
 
        down_write(&file->port->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;
        }
@@ -608,21 +616,29 @@ static int ib_umad_close(struct inode *i
        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_unregister_mad_agent(file->agent[i]);
+       down_write(&file->port->mutex);
+
+       already_dead = file->agents_dead;
+       file->agents_dead = 1;
 
        list_for_each_entry_safe(packet, tmp, &file->recv_list, list)
                kfree(packet);
 
-       down_write(&file->port->mutex);
        list_del(&file->port_list);
-       up_write(&file->port->mutex);
 
-       kfree(file);
+       downgrade_write(&file->port->mutex);
+
+       if (!already_dead)
+               for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
+                       if (file->agent[i])
+                               ib_unregister_mad_agent(file->agent[i]);
+
+       up_read(&file->port->mutex);
 
+       kfree(file);
        kref_put(&dev->ref, ib_umad_release_dev);
 
        return 0;
@@ -829,7 +845,6 @@ err_cdev:
 static void ib_umad_kill_port(struct ib_umad_port *port)
 {
        struct ib_umad_file *file;
-       struct ib_mad_agent *agent;
        int id;
 
        class_set_devdata(port->class_dev,    NULL);
@@ -849,16 +864,26 @@ static void ib_umad_kill_port(struct ib_
        port->ib_dev = NULL;
        up_write(&port->mutex);
 
-       list_for_each_entry(file, &port->file_list, port_list)
-               for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id) {
-                       down_write(&port->mutex);
-                       agent = file->agent[id];
-                       file->agent[id] = NULL;
-                       up_write(&port->mutex);
+       down_write(&port->mutex);
 
-                       if (agent)
-                               ib_unregister_mad_agent(agent);
-               }
+       while (!list_empty(&port->file_list)) {
+               file = list_entry(port->file_list.next, struct ib_umad_file,
+                                 port_list);
+
+               file->agents_dead = 1;
+               list_del_init(&file->port_list);
+
+               downgrade_write(&port->mutex);
+
+               for (id = 0; id < IB_UMAD_MAX_AGENTS; ++id)
+                       if (file->agent[id])
+                               ib_unregister_mad_agent(file->agent[id]);
+
+               up_read(&port->mutex);
+               down_write(&port->mutex);
+       }
+
+       up_write(&port->mutex);
 
        clear_bit(port->dev_num, dev_map);
 }
_______________________________________________
openib-general mailing list
[email protected]
http://openib.org/mailman/listinfo/openib-general

To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general

Reply via email to