You claimed that portability is a moot point.
Which is obviously wrong, because plenty of people are out there demonstrating that portability is very far from moot.
What Linus knows is how to achieve portability. As happens with many well-designed solutions, what you have to do makes the problem so transparent, that it is easy to miss that anything was done.
See [link|http://kt.zork.net/kernel-traffic/kt20000501_65.html#5|this] for a longer explanation of how you achieve portability. Then re-read the page that you quoted from. That is what Linus is doing.
Just in case someone missed it, here is how it works. What you do is define a simplified idealized model. Program to that model. For each architecture, supply compatibility macros so that the ugly details of that architecture look like that model. Except in a general outline, the architectures need not work the same way. This approach allows you to hide that fact in a clean way, with the ugly details hidden away nicely in an unobtrusive fashion.
However if you attempted to do the same thing using a different design, then you would very quickly find out that the differences are not minor, and portability is very, very far from being a moot point.
Cheers,
Ben