Hard time understanding recursive mapping

[mariuszp]

Previously when I had to modify the data of a page (such as a page table, page directory, etc, as they are also one page long each), I just mapped them to a page called an "ISP", and I switched the frame of the ISP whenever I had to work with a different page. This proved extremely inefficient when trying to copy address spaces (e.g. when calling fork()).

I'm trying to do recursive mapping. I don't need to do this for the entire PML4 - each process only gets one PDPT (mapped as the first PML4 entry) and I want to recursive-map that. So the equations for a PD and PT would be as follows:

Code: Select all

addr_pdpt = 0xFFFFFFFFF000;
addr_pd = 0xFFFFFFE00000 + i_pd * 0x1000;
addr_pt = 0xFFFFC0000000 + i_pd * 0x200000 + i_pt * 0x1000;

But if I substitute 511 into i_pd, I get addr_pd = addr_pdpt, so the PD struct for the last directory would be at the same address as the address of the PDPT struct. Then how do I access the last PD so that I can map more tables into it (as tables in the last PD contain pages that map to PTs of other PDs).

Am I missing something here? If I treat the PDPT and the last PD as the same structure, then setting its first entry to some frame would mean that adress 0 would simultaneously contain the page data and all levels of paging structures, which clearly makes no sense.

Could someone show some sample code maybe? (or pseudo-code?)

Please help

P.S. I'm in 64-bit Long Mode.

[eryjus]

You might take a look at this post: http://forum.osdev.org/viewtopic.php?f=1&t=28734

[jbemmel]

Or this one: http://forum.osdev.org/viewtopic.php?f=15&t=25545

[iansjack]

There's an explanation, with diagrams, that might help here: http://pdosnew.csail.mit.edu/6.828/2014 ... /uvpt.html

[thomasloven]

And yet another one: http://thomasloven.com/blog/2012/06/Rec ... Directory/
though not for long mode..

[mariuszp]

So, I take that mapping the PDPT's last entry onto itself would cause the last 1GB it encompasses to be mapped, and therefore each structure will be accessible, and then I can use the equations above to find each structure and therefore not bother about mapping the last PD, PT etc? But then I'm still confused, because only 1 frame gets allocated, and all those structures cannot be stored in 1 frame.

Just to make sure I understand: I initalize the PDPT as follows, correct?

Code: Select all

PDPT *pdpt = getPDPT();
pdpt->entries[511].present = 1;
pdpt->entries[511].pdPhysFrame = pdptPhysFrame;

Then, if I want a pointer to a specific page table entry, I'd do the following, correct?

Code: Select all

PTe *getPTE(uint64_t addr)
{
	uint64_t pageIndex = (addr & 0xFFF);
	addr >>= 12;
	uint64_t tableIndex = (addr & 0x1FF);
	addr >>= 12;
	uint64_t dirIndex = (addr & 0x1FF);

	PDPT *pdpt = (PDPT*) (0xFFFFFFFFF000);
	if (!pdpt->entries[dirIndex].present)
	{
		pdpt->entries[dirIndex].present = 1;
		pdpt->entries[dirIndex].pdPhysFrame = phmAllocFrame();
	};
	
	PD *pd = (PD*) (0xFFFFFFE00000 + dirIndex * 0x1000);
	if (!pd->entries[tableIndex].present)
	{
		pd->entries[tableIndex].present = 1;
		pd->entries[tableIndex].ptPhysFrame = phmAllocFrame();
	};
	
	PT *pt = (PT*) (0xFFFFC0000000 + dirIndex * 0x200000 + tableIndex * 0x1000);
	return &pt->entries[pageIndex];
};

Am I getting this right? The PDPT itself is the first entry of the PML4.

[Candy]

Yes. The trick is that you should see the table as containing itself 4 times -- every time mapped as the top (or bottom) entry. Accessing the next layer is done from the top or bottom as you first enter the location of that entry in the toplevel page table, then the next layer (which is now mappable as there is a page table for it!) and so on.

[mariuszp]

Yes. The trick is that you should see the table as containing itself 4 times -- every time mapped as the top (or bottom) entry. Accessing the next layer is done from the top or bottom as you first enter the location of that entry in the toplevel page table, then the next layer (which is now mappable as there is a page table for it!) and so on.

OK, thank you for the help. It appears to be working