Hello Ralf,

Running kernel 3.0.3 with rose patches.
Looking at the following code it seems that first spin_unlock is misplaced
just before memcpy().

I tried to move spin_unlock just after the call to memcpy without change.
The DEADLOCK is still there.


static void rose_get_neigh_callsign(ax25_address *rose_call,
         struct rose_neigh *neigh)
{
         spin_lock(&rose_callsign_lock);
         if (ax25cmp(&rose_callsign, &null_ax25_address) == 0) {
                 spin_unlock(&rose_callsign_lock);
                 memcpy(rose_call, neigh->dev->dev_addr, sizeof(rose_call));

                 return;
         }

When examining the inconsistent lock state report I may understand the 
reason.
The sequence is :

rose_route_frame()
rose_link_rx_start()
rose_send_frame()
rose_get_neigh_callsign()

However, rose_route_frame() is locking rose_route_list and 
rose_route_neigh_list with spin_lock_bh and if lci==0 there is a call in 
the function to rose_link_rx_restart before unlocking the lists.
If (neigh->restarted) rose_send_frame() is called and in turn will call 
rose_get_neigh_callsign() that tries to spin_lock rose_callsign_lock and 
here comes the conflicting situation.

Thus I modified rose_route_frame() in order to spin_unlock both lists 
before calling rose_link_rx_start().

Here are my patches.

However, another inconsistent lock state remains as you will see in the 
following attached file.


73 de Bernard


Le 08/08/2011 17:33, f6bvp a écrit :
> Hello Ralf,
>
> Booting 3.0.1 kernel with ROSE patches in SMP mode gives systematically
> the following Inconsistent Lock State.
> See attached file.
>
> Bernard
>
>