Support PAE on ia32

[Jakub Jermář]

Add support for Physical Address Extension to our ia32 port so that more than 4G of physical memory can be addressed.

Details

On all currently supported 32-bit platforms (arm32, ia32, mips32, ppc32), HelenOS assumes 32-bit physical addresses. This allows the system to make use of 4G of physical memory in total. Some of these architectures, however, provide extensions (e.g. PAE on ia32, LPAE on arm32) that make it possible to address more physical memory by using wider physical addresses (e.g. 36-bit or 40-bit).

There are actually two goals for this project. The first is to modify HelenOS to use a dedicated type for representing physical addresses instead of the current uintptr_t or void *, because the assumption that both virtual and physical addresses have the same amount of bits will no longer hold. The second goal is to implement the actual support for PAE in the form of PAE page table format and PAE feature detection and initialization.

What Gains and Benefits will this bring?

By having the PAE support on ia32, HelenOS will be able to utilize more of the installed memory. HelenOS will also become ready to support similar features on other architectures (think LPAE on arm32).

Difficulty

High

Required skills

Kernel programming skills are needed and the applicant should be strong in C and should have the ability to understand the HelenOS memory management subsystem quickly.

Documentation

• Developer documentation for IA-32 architecture
• ​HelenOS abstraction for virtual address translation
• ​Memory management implementation for IA-32 (​headers)

Possible mentors

HelenOS Core Team, Jakub Jermar

[Jiří Zárevúcky]

Is this still something we are interested in? ia32 is very nearly extinct by now, arm's 64b variant is slowly entering mainstream as well, and I have personally never seen an arm32 board with more than 4GB RAM.

[Martin Decky]

Despite the arguable obscurity I would still prefer to keep the ticket. It is an interesting opportunity to come up with a generic support for the case where the physical address space is larger than the virtual address space (or, even more generally, with a well-designed type-safe support for physical and virtual pointers). The priority is minimal, of course.

[Jiri Svoboda]

Aren't there 32-bit Intel Atoms? While desktop/server processors are moving to 64-bit, 32-bit processors seem to be rather popular choice for embedded use, in general (AVR32, Leon is a 32-bit SPARC, etc.). ia32 seems like a good particular case to implement the general feature, which can be useful in other cases (e.g. >64 bit physical space).

[Jiří Zárevúcky]

Atom CPUs have had 64b support since 2008 (though according to wiki, motherboard support is another story). Besides, 32b embedded boards with >32b physical address space sounds like something utterly cost-inefficient.

But I do agree that it's a good test case for improving kernel interfaces.