Linus Torvalds (torvalds.delete@this.linux-foundation.org) on September 21, 2021 1:55 pm wrote:

> And yes, the x86 segments were bad. They could have done more interesting things, and having segments be
> truly independent address spaces (ie give them separate page tables) could have made them much more powerful
> and useful. They had other issues too. But the fact that nobody else has ever done the concept well - whether
> you call it capability pointers or segments or something else - should give people pause.
>
> Linus


Why would segments need different page tables? What would be point of doing so instead of different processes?


8086 segments were bad but those had nothing to do with MMU and privilege levels. 80286's MMU and specially its extended version on 80386 instead had some serious brainpower behind them. Remember they were designed in beginning of 80's and did have hardware limitation of what they could implement. I believe that I read that 386 was on reticle limit on it's first process.

And yes, I don't mean that somebody should replace different address spaces and paging with segmentation, neither did Intel. No matter if there's one or many address spaces memory coherence needs system to combine all memory mappings together for coherence checks and doing so for segmentation model isn't wildly different to multiple independent address spaces.

But I think that there's nothing wrong with segmentation. Heck, what's the point of object orientation if there's no hardware to protect object from each others? Stop gap solution is to try make software-based protection but why if there's a way to do it with hardware?

And again, what's the main problem with 386's segmentation? 4 privilege levels - different segments for code, stack and data and ability to map four data segments simultaneously. Remember segmentation main target isn't protect from other processes but instead protect your code and data from your own (buggy)code. It also gives a simple way to do hardware-based locking between threads.


Yes programming sure is easier if there's no need to bother with segmentation - but as today software is pumped up with hundreds of different software-based complex implementations against side-channels and other data leaks so I think they should really take a new look about segmentation and it's design targets.