cvs commit: src/sys/kern subr_turnstile.c

[John Baldwin]

On Tuesday 09 September 2008 06:09:34 am Robert Watson wrote:
> On Mon, 8 Sep 2008, John Baldwin wrote:
> 
> >  SVN rev 182879 on 2008-09-08 21:40:15Z by jhb
> >
> >  - Reduce scope of #ifdef's in uma_zcreate() call in init_turnstile0().
> >  - Set UMA_ZONE_NOFREE so that the per-turnstile spin locks are type 
stable
> >    to avoid a race where one thread might dereference a lock in a free'd
> >    turnstile that was previously used by another thread.
> 
> Is this a feature or a workaround for a bug?  Normally in the above scenario 
> we would consider use-after-free a bug or symptom of a larger architectural 
> problem rather than a feature.  At least, that's what I consider similar use 
> of UMA_ZONE_NOFREE where it persists in the network stack :-).

Well, it's a workaround for the fact that the way thread_lock works is it 
tries to acquire what it thinks is the current lock for a given thread.  Once 
it has that lock, then it checks to see if the thread has switched to a 
different lock.  If so, it drops the lock it has and tries to get the "new" 
lock.  Anytime you lose this race, you can end up holding a lock that isn't 
necessarily associated with the thread anymore.  For that reason, locks used 
as thread locks should generally be type-stable.  Most of the locks used as 
thread locks are in static data structures (runqueues, sleepq hash table 
buckets, the global "blocked lock", etc.) so they are already type-stable.  
The turnstile locks are the one case where locks used as thread locks are 
dynamically allocated IIRC.

-- 
John Baldwin