Re: Release Engineering Status Report

Scott Long wrote:
> Agreed.  PAE was merged into -stable in three steps.  Backing out the
> third step and leaving the first two steps removes the instability.
> Unfortunately, it was the third step that also was the most complex.
> In any case, we have 2 weeks to find the resolution before the decision
> must be made on keeping or tossing PAE.  Since PAE is a *highly*
> sought after feature, it would be doing a disservice to our user base
> to remove it without putting in some effort to fix it.

Not to be rude or anything, so please don't take it that way...

Exactly why is PAE "...a *highly* sought after feature..."?  It's
not like it increases the KVA space for the kernel itself, or UVA
space for individual processes.

I could maybe understand if processes were no longer mapped into
KVA on trap/system call entry, since that would cause the UVA and
KVA to both go to the full 4G size, at some additional copying
cost, which is (relatively) low on Intel, for the benefit of a
larger virtual address space that would let you map more kernel
data and/or run larger application working sets for things like
databases.

But as it is, it's basically nothing more than L3 cache, which
has always been seen as being of dubious benefit (otherwise all
machines would come with L3 caches).

-- Terry