RELENG_7 crash

[Robert Watson]

On Tue, 21 Apr 2009, Ruslan Ermilov wrote:

>> Is it possible I am running into some of the interface lock fixes rwatson 
>> has been working on ?  This box has a lot of ng interfaces which come and 
>> go. Perhaps snmp asking about an interface that just went away caused the 
>> panic ? I disabled bsnmp since the reboot and the box has been up for 10hrs 
>> so far.
>>
> It's a documented bug:

Just to follow up on this thread: I've committed a work-around to 7-STABLE as 
well as the 7.2 release engineering branch.  The on-going ifnet 
reference-counting/address list locking/etc work in 8-CURRENT is the long-term 
solution, and I will MFC it after 7.2 is out the door and it's settled some, 
but with any luck the work-around will prevent instant panics in the described 
scenario.

Robert N M Watson
Computer Laboratory
University of Cambridge

>
> : revision 1.281
> : date: 2008/06/26 23:05:28;  author: rwatson;  state: Exp;  lines: +69 -12
> : SVN rev 180042 on 2008-06-26 23:05:28Z by rwatson
> :
> : Introduce locking around use of ifindex_table, whose use was previously
> : unsynchronized.  While races were extremely rare, we've now had a
> : couple of reports of panics in environments involving large numbers of
> : IPSEC tunnels being added very quickly on an active system.
> :
> : - Add accessor functions ifnet_byindex(), ifaddr_byindex(),
> :   ifdev_byindex() to replace existing accessor macros.  These functions
> :   now acquire the ifnet lock before derefencing the table.
> : - Add IFNET_WLOCK_ASSERT().
> : - Add static accessor functions ifnet_setbyindex(), ifdev_setbyindex(),
> :   which set values in the table either asserting of acquiring the ifnet
> :   lock.
> : - Use accessor functions throughout if.c to modify and read
> :   ifindex_table.
> : - Rework ifnet attach/detach to lock around ifindex_table modification.
> :
> : Note that these changes simply close races around use of ifindex_table,
> : and make no attempt to solve the probem of disappearing ifnets.  Further
>      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> : refinement of this work, including with respect to ifindex_table
> : resizing, is still required.
> :
> : In a future change, the ifnet lock should be converted from a mutex to an
> : rwlock in order to reduce contention.
> :
> : Reviewed and tested by:	brooks
>
>
> Cheers,
> -- 
> Ruslan Ermilov
> ru at FreeBSD.org
> FreeBSD committer
>