On Tue, Jan 30, 2007 at 11:37:25PM -0600, Donald Lee wrote:
> One, at the end of the file, where the "${si}" gets put back in the
> pool, the operation is protected by "splvm()/splx()" rather than
> the msr ops that are now sprinkled in the code.  I thought the
> reason for the msrs was because the IPLs were unreliable.  Why
> are they OK here?

The pool operations just need splvm protection because other pool
operations can not happen at higher IPL than IPL_VM.  Note the
comments above the functions.  The enqueue and dequeue operations
are what need to be interrupt-safe because they can be called at
pretty much any IPL and you don't want to get an interrupt in the
middle of the operation that might lead to a re-entry into that code.

> Two, it seems that this newer code does not provide for
> multiple CPUs in the code.  I can see where if multiple CPUs got into
> the code, we'd have a mess.  Is this covered elsewhere?
> (BTW - the MP issue does not matter to me.  I only have one CPU)

Right now, I think the biglock model protects against that.  In a
fully-reentrant MP system, we would need to hold a mutex on the
softintr structure for the enqueue/dequeue, and a different one
for access to the pool (probably encapsulated in the pool code).

-allen

-- 
Allen Briggs  |  http://www.ninthwonder.com/~briggs/  |  briggs@ninthwonder.com