cvs commit: src/sys/contrib/ipfilter/netinet ip_auth.c

John Baldwin jhb at
Tue Dec 28 13:50:44 PST 2004

On Tuesday 28 December 2004 12:18 am, Darren Reed wrote:
> On Mon, Dec 27, 2004 at 08:39:44PM -0700, Scott Long wrote:
> > The locking APIs have existed in FreeBSD 5.x for 4 years.  They are
> > documented in manual pages, web pages, and publically available USENIX
> > papers.  Just because you don't agree that sx locks should be sleepable
> > doesn't mean that your opinion is valid.
> Hmmm, maybe I'm not communicating my view properly.
> I'm not fussed about whether or not sx locks are sleepable.
> What's wrong, IMHO, is the witness code and the implication that
> acquiring a sleepable lock, while you have a mutex, is wrong.
> Or vice versa.  So long as the order is always the same, it
> should not present a problem if the programmer is careful.
> Even "debug" kernels on Solaris don't have restrictions like that
> between mutex/rw-locks, although Solaris never allows you to do
> a recursive mutex, like FreeBSD.
> FWIW, John Baldwin's USENIX paper that talks about witness from BSDi
> only talks about it in context of locking order:
> The problem it is complaining about w.r.t ipfilter is not what the
> paper talks about here.
> As it is, if I were to refer to:
> 050 - i've actually changed the code for other reasons so it should
>       not happen at some point in the future
> 051 - why is UDP holding on to the lock for so long ?
>       this doesn't look like an ipfilter problem.
> 052 - i'm dumbfounded.  IFNET_RLOCK uses mtx, not sx.
> 052 is a classic case of why this LOR report is not useful.
> The witness code is complaining because a mutex is being acquired
> while a sleepable lock is held, for the purpose of walking a linked
> list.  I mean WTF?
> Just to reiterate, it's the witness code and the semantics it is trying
> to enforce (which are somewhat different, now?) that I have an issue
> with, not the lock implementations.

sx locks are not identical to Solaris rwlocks.  Solaris rwlocks do priority 
propagation and can't be held across a condition variable type sleep.  sx 
locks on the other hand are more like lockmgr() style locks in that they do 
not do priority propagation and may be held across a sleep (and are even 
implemented using mutexes and condition variables for that matter).  I do 
plan to implement Solaris-style rwlocks that do priority propagation, etc. 
and are not sleepable but have not done so yet.  Until that time, you will 
have to use mutexes as those are the only locks FreeBSD supports that have 
the semantics you want.

John Baldwin <jhb at>  <><
"Power Users Use the Power to Serve"  =

More information about the cvs-src mailing list