On Sun, 25 Mar 2001, Linus Torvalds wrote:

> So my suggestion for PAE is:
>
>  - populate in gdb_alloc() (ie just do the three "alloc_page()" calls to
>    allocate the PMD's immediately)
>
>    NOTE: This makes the race go away, and will actually speed things up as
>    we will pretty much in practice always populate the PGD _anyway_, the
>    way the VM areas are laid out.
>
>  - make "pgd_present()" always return 1.
>
>    NOTE: This will speed up the page table walkers anyway. It will also
>    avoid the problem above.
>
>  - make "free_pmd()" a no-op.
>
> All of the above will (a) simplify things (b) remove special cases and
> (c) remove actual and existing bugs.

yep, this truly makes things so much easier and cleaner! There was only
one thing missing to make it work:

   - make "pgd_clear()" a no-op.

[the reason for the slightly more complex pgd_alloc_slow() code is to
support non-default virtual memory splits as well, where the number of
user pgds is not necesserily 3.]

plus i took to opportunity to reduce the allocation size of PAE-pgds.
Their size is only 32 bytes, and we allocated a full page. Now the code
kmalloc()s a 32 byte cacheline for the pgd. (there is a hardware
constraint on alignment: this cacheline must be at least 16-byte aligned,
which is true for the current kmalloc() code.) So the per-process cost is
reduced by almost 4 KB.

and i got rid of pgalloc-[2|3]level.h - with the pmds merged into the pgd
logic the algorithmic difference between 2-level and this pseudo-3-level
PAE paging is not that big anymore. The pgtable-[2|3]level.h files are
still separate.

the attached pae-2.4.3-B2 patch (against 2.4.3-pre7) compiles & boots just
fine both in PAE and non-PAE mode. The patch removes 217 lines, and adds
only 78 lines.

	Ingo