Ok, since net_dev_init() is called from genhd.c:53, I figured it'd work 
just fine if I removed it from net/core/dev.c in register_netdev(). 
 Since I'm not as well steeped in this code, those who know better are 
invited to correct me.

--- linux-2.4.6.orig/net/core/dev.c     Sat Jul 14 17:48:57 2001
+++ linux-2.4.6/net/core/dev.c  Wed Jun 20 21:00:55 2001
@@ -2405,6 +2401,9 @@
 #ifdef CONFIG_NET_FASTROUTE
        dev->fastpath_lock=RW_LOCK_UNLOCKED;
 #endif
-
-       if (dev_boot_phase)
-               net_dev_init();
 
 #ifdef CONFIG_NET_DIVERT
        ret = alloc_divert_blk(dev);

David

David Ford wrote:

> Ok, the problem is this.  I have TEQL packet scheduling in my config, 
> the kernel runs through this sequence on boot:
>
> net_dev_init()
>    pktsched_init()
>        teql_init()    [starts a lock with rtnl_lock()]
>            register_netdevice()
>                net_dev_init()
>                    pktsched_init()
>                        teql_init() [hangs here...]
>
> Here is the problem.  We enter teql_init() again with a rtnl_lock() 
> already being held.  Do any of the authors of these functions want to 
> jump in here?
>
> David
>
> David Ford wrote:
>
>> [...]
>> I2O LAN OSM (C) 1999 University of Helsinki.
>> early initialization of device teql0 is deferred
>> loop: loaded (max 8 devices)
>> Linux Tulip driver version 0.9.15-pre3 (June 1, 2001)
>> PCI: Found IRQ 5 for device 00:10.0
>>
>> Any comments or suggestions?  2.4.5-ac19 is the last kernel I have 
>> that works.
>