How did paging_init ever work with PAE?

How did paging_init ever work with PAE?

[Martin J. Bligh]

We're crashing in paging_init under certain circumstances, but on
closer inspection, I'm unsure how this ever could have worked.
Obviously it does most of the time, so I'm missing something ....

With PAE on, the code path looks like this:

        pagetable_init();
        load_cr3(swapper_pg_dir);
        if (cpu_has_pae)
                set_in_cr4(X86_CR4_PAE);

Hmmm .... pagetable_init sets up a PGD for PAE use, then we load
cr3 with this table .... then we turn on PAE mode.

How are we surviving in this limbo state between the point when
we reload cr3 and when we turn on PAE? If we take a page fault
(which we will, since reloading cr3 flushes the tlb) is the PGD 
somehow dual purpose and works for non-PAE systems as well?
Are we relying on the global bit on entries on the TLB cache which
we're just praying aren't going to fall out?

Seems very strange ..... any clarification much appreciated.

Thanks,

M.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Re: How did paging_init ever work with PAE?

[Hugh Dickins]

On Tue, 30 Jul 2002, Martin J. Bligh wrote:
> We're crashing in paging_init under certain circumstances, but on
> closer inspection, I'm unsure how this ever could have worked.
> Obviously it does most of the time, so I'm missing something ....
> 
> With PAE on, the code path looks like this:
> 
>         pagetable_init();
>         load_cr3(swapper_pg_dir);
>         if (cpu_has_pae)
>                 set_in_cr4(X86_CR4_PAE);
> 
> Hmmm .... pagetable_init sets up a PGD for PAE use, then we load
> cr3 with this table .... then we turn on PAE mode.
> 
> How are we surviving in this limbo state between the point when
> we reload cr3 and when we turn on PAE? If we take a page fault
> (which we will, since reloading cr3 flushes the tlb) is the PGD 
> somehow dual purpose and works for non-PAE systems as well?

Yes.

> Are we relying on the global bit on entries on the TLB cache which
> we're just praying aren't going to fall out?

No.

See swapper_pg_dir in arch/i386/kernel/head.S: at swapper_pg_dir+0xc00
there's a couple of 32-bit non-PAE entries to map 0xc0000000-0xc07fffff
where the kernel is; whereas pagetable_init sets up four 64-bit PAE
entries at swapper_pg_dir (overwriting the non-PAE identity mappings,
yes, but they won't be needed during changeover from non-PAE to PAE).

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/