Re: mbuf count negative

On Sun, 5 Dec 2004, Stephan Uphoff wrote:

> I am wondering if critical sections are overkill for this application
> since the interrupt blocking capability is not needed. 
> 
> Perhaps a mutex_lock_local, mutex_unlock_local that work on per CPU
> mutexes and require the current thread to be pinned to a CPU would be
> more appropriate. The functions would have the same functionality as
> mutex_lock/mutex_unlock but without using atomic operations.  The
> "local" mutex would be a clone of the UP sleep mutex and the SMP sleep
> mutex and the local mutex could even be used together in any order.
> (With working priority inheritance)  I think I will try this in my p4
> mutex branch in the next days. 
> 
> A "do not preempt" section (Interrupt enabled) would be another
> possibility. 

Silby suggested a UP mutex implementation on SMP to use with pinned
threads a month or two ago, and that would work just fine here; I was
having similar thoughts also -- while the critical sections are brief,
there's really no reason not to allow preemption there.  However, the
critical section optimizations that I was referring to from John basically
do what you suggest second: they implement critical sections via a
software "Don't preempt".  I don't really mind what the mechanism is, as
long as they neither invoke cli/sti or use atomic operations, and it
should be possible to easily substitute in the UMA code once my patch is
in place.

Robert N M Watson             FreeBSD Core Team, TrustedBSD Projects
robert_at_fledge.watson.org      Principal Research Scientist, McAfee Research