Michael S (already5chosen.delete@this.yahoo.com) on September 20, 2021 7:03 am wrote:
>
> "Single address space" is also a sane policy, at least as long as
> virtual address space is many times bigger than physical memory.

That has actually never been shown to be true in reality.

It might be true.

It almost certainly isn't.

The thing is, your virtual address space doesn't need to be "many times bigger". It needs to be "many orders of magnitude" bigger. That's big. And it makes a lot of the arguments for SAS systems much weaker, since now SAS implies that "much bigger VA space required".

But equally importantly, every single person who ever claims this also assumes a perfect world with no data leakage of those supposedly secret addresses.

IOW, now those magical SAS systems not only require much bigger VA spaces, they also fundamentally rely on something that has never actually been true in practice. Not in hardware, and certainly not in software with all the alleged advantages being all about passing pointers around.

Don't tell me about protected pointers, if you don't also tell me about how you protect against TLB and cache timing leak attacks etc etc. Or how you deal with all the overhead of your segments or capability hardware or whatever magic you use, and how it is magically cheaper than the TLB with ASID's that you wanted to replace.

I think there are a lot of people who base their "single address space is viable" opinion on a very shaky basis.

Of course, if you don't care about protection, SAS is fine and has been proven over and over, and the Amiga people will still talk your ear off about it for as long as you let them.

That "if you don't care about protection and security" a big "if", though, and I think it entirely negates your "sane policy" argument in practice.

Because a thousand academic papers over multiple decades don't really make up for the lack of any actual physical proof.

Linus