On Fri, Nov 14, 2003 at 12:04:29PM -0500, Nathan J. Williams wrote:
> Jaromir Dolecek <jdolecek@NetBSD.org> writes:
> 
> > I'd perhaps return to the core of argument, that being
> > the proposed change to eliminate the upper cap on number
> > of LWPs on NetBSD/i386. Given that it took some 10 years
> > to actually start tackle the TSS limitation, it should
> > be done in some form or another finally IMHO.
> 
> "It should be done in some form or another? Is that true? Why should
> it be done at all? Sure, it's conceptually bad to be limited to a
> fixed number of LWPs, but is it a practical problem? What we're seeing
> in this discussion is that there are side effects to the change beyond
> simply removing a cap, and so we have to consider whether it's a cap
> that is actually causing trouble.

One of problems in the current GDT management is that the GDT only
grows and never shrinks. There was gdt_compact() (and others) some
time ago, but unfornately, it caused a kernel crash (PR#9368). Worse
is, nobody could fix it for years. So somebody commented it out, fvdl
removed $if 0'ed gdt_compact(), and I removed #if 0'ed gdt_shrink().

By adopting yamt's per-CPU TSS, you don't have to care about memory
leakage caused by massive process creation and termination any longer.

Jun-Young

-- 
Bang Jun-Young <junyoung@NetBSD.org>